# From Parisi full support to deterministic QAOA circuits

## A long reasoning record

3 November 2026. Mathematical sources: OpenAI's full-support
manuscript of 27 September 2026 and QAOA manuscript of 25 September 2026,
at the [pinned public revision](https://github.com/openai/math/tree/adc7f1241b42e322a6451854ab7e4b4c146bf78a).

**AI assistance.** Will Cook directed the work. AI agents based on large
language models produced and revised this manuscript.

<!-- passage:overview -->
OpenAI's full-support manuscript proves that an admissible integrable
minimizer of the zero-temperature Parisi functional for the pure zero-field
Sherrington–Kirkpatrick model has Stieltjes support equal to every overlap
scale in [0,1).

The proof turns a hypothetical support gap into a constant interval for the
order parameter. Variational contact at the ends of that interval then
conflicts with a strict curvature inequality inside it. Along the auxiliary
diffusion, the spatial gradient of the value function is a bounded
martingale. Full support makes its second moment equal to time, and its
terminal correlation with Brownian motion gives the ground-state value.

The manuscript's further argument gives a smooth Stieltjes density, strictly
positive on $(0,1)$, and no atom at zero. Sections 9.1–9.2 derive this stronger
ordinary conclusion separately from the selected formal support statement.

We reconstruct these arguments and follow the finite Gaussian approximation
into the separate QAOA manuscript, which claims that finite deterministic
circuits approach the SK optimum when system size tends to infinity before
circuit depth increases. Its circuit construction needs four further
arguments: approximation, synthesis, seed removal and transfer. The two
formal statements concern full support and the scalar value consequences,
and neither states QAOA circuit convergence. Sections 12–13 describe what
they cover. See [OpenAI (2026a), §§1–2][fs-main] and
[OpenAI (2026b), Theorem 1.1][q-intro].
<!-- /passage:overview -->

The support and scalar arguments assume real analysis, stochastic calculus
and variational minimization. Section 8 also uses Hilbert-space tensor
products, symmetric Fock space, unbounded operators and quantum control. Its
weighted-domain and control estimates are further analytic inputs from the
QAOA manuscript, and the text gives each one's precise use and proof
destination.

The companion short paper concentrates on full support and the strictly
positive interior density. The exposition introduces the support argument
through examples. Throughout, OpenAI (2026a) and OpenAI (2026b) name the two
manuscripts.

Sections 1–8 give the continuous proof: full support, the terminal value,
finite Gaussian correlations and, for each accuracy, one ordinary QAOA word.
Section 9 derives the density and finite-disorder consequences. Section 10
gives an independent scalar comparison and a heat-family lower bound.
Section 11 treats alternative circuit constructions and further limits.
Section 12 gives the exact formal statements and compares them with the
model, and Section 13 lists the sources and the finite calculation.

<a id="model-and-three-targets"></a>

## 1. The SK model and the Parisi variational problem

<!-- passage:model -->
For independent standard Gaussian couplings $J_{ij}$, put

$$
H_n(\sigma)=\frac1{\sqrt n}\sum_{i<j}J_{ij}\sigma_i\sigma_j,\qquad
\sigma\in\{-1,1\}^n,\qquad
P_*=\lim_{n\to\infty}\frac1n\mathbb E\max_\sigma H_n(\sigma).
$$

For overlap $R_{12}=n^{-1}\sum_i\sigma_i^1\sigma_i^2$, the covariance is
$nR_{12}^2/2-1/2$. Thus the convention is $\xi(t)=t^2/2$, with
$\xi''=1$. A centered Gaussian of variance $1/2$, common to all
configurations, removes the final constant and leaves the expected maximum
unchanged. Every normalization below uses this convention.

Let $\mathcal U$ consist of nonnegative, nondecreasing, right-continuous
functions $\gamma:[0,1)\to[0,\infty)$ with
$\int_0^1\gamma(t)\,dt<\infty$. Their Stieltjes measures satisfy
$\mu_\gamma([0,t])=\gamma(t)$, including
$\mu_\gamma(\{0\})=\gamma(0)$. They are locally finite. They need not be
probability measures, and their total mass may be infinite. Support is taken
relative to [0,1).

Let $B$ be standard Brownian motion. The control value and Parisi
functional are

$$
\Phi_\gamma(t,x)=\sup_{|\alpha|\le1}\mathbb E\left[
 \left|x+B_1-B_t+\int_t^1\gamma(s)\alpha_s\,ds\right|
 -\frac12\int_t^1\gamma(s)\alpha_s^2\,ds\right],
$$
$$
\mathcal P(\gamma)=\Phi_\gamma(0,0)-\frac12\int_0^1t\gamma(t)\,dt.
$$

Controls are progressively measurable for Brownian increments after $t$.
The control formula fixes the solution, and the PDE is
$\Phi_t=-\tfrac12(\Phi_{xx}+\gamma\Phi_x^2)$, with
$\Phi(1,x)=|x|$. For a general admissible coefficient, the solution and its
spatial derivatives are locally absolutely continuous in time on compact
strips, and the time equation holds almost everywhere.

Auffinger and Chen proved the zero-temperature Parisi formula and its
attainment, which identify $P_*$ with $\min_{\gamma\in\mathcal U}\mathcal P(\gamma)$.
The support theorem itself takes a minimizer as given and proves a property
of it. See [OpenAI (2026a), §§1–2][fs-main]
and [Auffinger and Chen (2017), Theorem 1 and proof of Lemma 3][ac2017].
<!-- /passage:model -->

<!-- passage:targets -->
Write $u=\Phi_x$, $a=\Phi_{xx}$, and
$dX_t=\gamma(t)u(t,X_t)\,dt+dB_t$, $X_0=0$.
Write $q(t)=\mathbb E[u(t,X_t)^2]$ for the second moment of this
gradient along $X$.

The first result assumes an admissible minimizer. It constructs its
Stieltjes measure and an associated diffusion, proves
$\operatorname{supp}\mu_\gamma=[0,1)$ and strict increase of $\gamma$,
and obtains $\mathbb E u(t,X_t)^2=t$ and
$\mathbb E a(t,X_t)^2=1$ for every associated diffusion. The selected
formal statement has exactly this conditional scope and asserts no smooth
density.

Its [Lean implementation at the fixed source revision][support-main] implements the
[separately stated formal challenge][support-type], and the
[Comparator configuration][support-config] identifies the intended
comparison. The evidence kept with this draft does not establish a
successful research-target replay.

The second result concerns the scalar value. A bounded martingale ends
at a spin-valued random variable whose Brownian correlation represents
$P_*$. Finite Gaussian coefficients approximate that correlation at each
fixed $T<1$, and then $T\uparrow1$. Its formal statement is separate
and keeps the minimizer and diffusion hypotheses where they are needed.

The QAOA theorem goes further. For each requested accuracy it selects one
finite deterministic circuit, independent of size and disorder, whose
limiting expected energy approaches $P_*$. Neither formal statement asserts
that circuit conclusion.
The scalar clauses are in a [separate Lean implementation][value-main], with their
own [formal challenge][value-type] and [Comparator configuration][value-config].
Those links identify the selected statements and give no completed replay
receipt for either research target.
<!-- /passage:targets -->

<a id="argument-map"></a>

## 2. Analytic preliminaries and prior work

<!-- passage:dependencies -->

### 2.1 Outline of the proof

We begin with regularity on strips bounded away from terminal time, which
makes the diffusion $X$ and the gradient moment $q$ available, so that a
first variation can show where the minimizer is in contact. Then, on a constant
interval, the shape inequalities and the rational certificate give the
strict curvature needed to contradict that contact. The zero-interval case
uses a separate heat argument, and the possible terminal plateau needs its
own exclusion. Together these steps give full support and the two moment
equalities.

The value argument then uses bounded-martingale convergence and the
attained Parisi formula to express $P_*$ as a terminal Brownian correlation
and as an integral of expected curvature. Section 6 defines the terminal
martingale variable and proves these equalities. The fixed-$T$ Gaussian
approximation recovers this integral through Euler coupling, orthogonal
normalization and mean-preserving rounding.

The QAOA theorem needs four more steps: rooted energy identification,
classical synthesis, removal of temporary seeds and transfer to fixed
angles. The main route is §§3–5, then §§6–7, then §§8.1–8.6, and §12
compares the two formal interfaces. The original analytic details are in
[OpenAI (2026a), §§2–7][fs-main],
and the circuit construction is in [OpenAI (2026b), §§2–6][q-intro].

The density theorem in §§9.1–9.2 follows from support and the PDE, and
neither the scalar argument nor the circuit construction uses it. The
positive-temperature route in §11.1 independently supplies the same bounded
classical input used for synthesis. The probability consequence in §11.2
requires a separate fixed-word second-moment theorem. These branches show
why the two selected formal statements fall short of a formal proof of the
complete circuit theorem.

The run-in paragraphs below separate the premises of the support proof from
its consequences and locate the detailed arguments behind the short paper's
support and density theorem.

**The diffusion and moment identity.** For any admissible $\gamma$, the
compact-strip estimates justify defining $X,q$ and computing
$q'=\mathbb E a^2$; see [§2.4](#strip-estimates) and
[§3.1](#gradient-moment).

**Variational contact.** Minimality, admissible mass addition and compact
mass removal give $G\ge0$, its vanishing on support and the contact slopes.
The steps are in [§3.2](#feasible-variation), [§3.3](#measure-contact)
and [§3.4](#contact-slopes).

**Strict curvature.** This step uses only admissibility and a positive
constant $c$ on an interval. The shape comparisons, the algebra and the
limiting argument are in [§4.3](#shape-comparisons),
[§4.4](#curvature-algebra) and [§4.5](#plateau-limit).

**Exclusion of gaps.** We combine contact with the positive, zero or
terminal argument, as the case requires. Each case has its own destination:
[terminal](#terminal-gap), [bounded](#bounded-gaps) and
[initial](#initial-gaps) gaps.

**Positive density.** Full support comes first. Differentiated moments
then give [smoothness in §9.1](#smooth-density-consequence), and a local
comparison gives [strict positivity in §9.2](#positive-density).

**Value and circuits.** We keep the same minimizer and diffusion through
[the value identity in §6](#scalar-value-identities) and
[the finite Gaussian construction in §7](#finite-gaussian-approximation).
[The circuit argument in §8](#qaoa-construction) requires additional theorems.

### 2.2 Compact-strip regularity and the terminal bound

**Proposition 2.1 (compact-strip regularity).** For $\gamma\in\mathcal U$
and $0\le T<1$, the value has bounded, jointly continuous spatial derivatives
of every positive finite order on $[0,T]\times\mathbb R$. If admissible
coefficients converge in $L^1$ and their smooth, even, convex, one-Lipschitz
terminal values converge uniformly to $|x|$, these spatial derivatives
converge uniformly on each such strip. The bounds depend on the derivative
order, the fixed strip and a common coefficient mass bound. Time derivatives need only exist
almost everywhere across coefficient jumps. On an open constant interval,
the heat formula supplies the additional time smoothness used in §4.
The terminal estimate is
$$
|x|\le\Phi_\gamma(t,x)
\le |x|+\sqrt{2(1-t)/\pi}+\frac12\int_t^1\gamma(s)\,ds.
$$
It controls the approach to $|x|$ without differentiating the cusp. The
strip estimates justify the interior calculations and finite meshes, while
the terminal bound and integrability justify the later endpoint limits. We
keep the two roles separate, since each depends on the region and the limit.

On a fixed $[0,T]$ with $T<1$, bounded spatial derivatives and locally
absolutely continuous time dependence justify the Itô identities and Euler
coupling. Across a jump of $\gamma$, the spatial derivative moments remain
continuous and the time equations hold almost everywhere. This is enough
for the contact slopes and the endpoint chord comparison. Inside a constant
interval, classical time differentiation permits the strict-curvature
calculation.

As $T\uparrow1$ the controls change. The displayed value bound, integrable
coefficient tails and martingale convergence justify the first variation and
the terminal value, so the argument uses no differentiated estimate at the
terminal cusp.

These are [OpenAI (2026a), Lemma 2.1 and its diffusion consequences][fs-reg],
and [§2.4](#strip-estimates) reconstructs their uniformity.

The hypotheses also have distinct uses. Zero field supplies evenness,
$u(0,0)=0$ and reflection in the shape argument. The convention
$\xi''=1$ fixes the curvature coefficients. Monotonicity gives bounded
coefficients before time one and nonnegative coefficient jumps, and it
turns a support gap into a constant interval. Integrability controls
terminal tails and Fubini interchanges, while local finiteness of the
Stieltjes measure is enough for compact mass removal.

Minimality first enters through the variational inequality in §3. The
constant-interval curvature estimate applies to any admissible coefficient,
minimiser or not. [OpenAI (2026a), §§1–2][fs-main],
[OpenAI (2026a), §2][fs-reg] and [OpenAI (2026a), §6][fs-var].

<a id="prior-work-comparison"></a>

### 2.3 Prior results and their roles

The ground-state formula of Auffinger and Chen supplies two facts with different uses. Their Theorem 1 identifies the ground-state limit with an infimum over $\mathcal U$. To choose a coefficient attaining that infimum, we also need their construction in the proof of Lemma 3. In the cited arXiv v2, the theorem is on p. 3, the compactness construction on pp. 7–9 and the lower-bound proof on pp. 11–12.

The support proof is conditional on a minimizer. A reader who grants that
hypothesis can continue directly with
[the compact-strip estimates in §2.4](#strip-estimates). The construction
below shows how attainment supplies the minimizer later used in the value
and circuit arguments.

<a id="minimizer-attainment"></a>

Here is the part of that construction which supplies the witness. Let $\alpha_\beta$ be a positive-temperature Parisi distribution function. The source bounds $\beta\alpha_\beta(t)$ uniformly on each interval ending before one. Helly selection and a diagonal subsequence produce a nonnegative, nondecreasing right-continuous limit $\gamma_0$. In the SK normalization, the source's weighted integral bound also controls $\int_0^1\beta\alpha_\beta(t)\,dt$. For every nonnegative nondecreasing $h$, splitting at $1/2$ gives $\int_0^1 h(t)\,dt\le4\int_0^1t h(t)\,dt$: the first half is bounded by $h(1/2)/2$, and on the second half $t\ge1/2$. Fatou's lemma then gives $\gamma_0\in\mathcal U$.

Mass can concentrate at time one during this limit, so compact convergence alone does not identify the limiting control problem. Write the limiting measure of the time coefficients as
$$
\nu_0(dt)=\gamma_0(t)\,dt+m\delta_1(dt),\qquad m\geq0.
$$
For an arbitrary admissible control, let $S$ denote its terminal state with coefficient $\gamma_0$. To see why the terminal choice is admissible, let the original control be
$\alpha$ and put $S_s=B_s+\int_0^s\gamma_0(r)\alpha_r\,dr$.
For a cutoff $\tau<1$ use $\alpha_s$ before $\tau$ and
$g_k(S_s)$ afterwards, where $g_k$ is a continuous function bounded by
one and converging pointwise to the sign with $\operatorname{sgn}(0)=1$.
This control uses only information available at time $s$. At fixed
$\tau,k$, we first take the temperature limit and then let $\tau\uparrow1$
and $k\to\infty$. The continuous terminal evaluation has thereby
approached $\operatorname{sgn}(S)$ without anticipating a future state. The endpoint atom then contributes
$$
|S+m\operatorname{sgn}(S)|-|S|=m
$$
to the reward and
$$
\tfrac12m\bigl(\operatorname{sgn}(S)^2+1\bigr)=m
$$
to the combined control and Parisi penalties. These contributions cancel. Taking the supremum over the remaining controls proves $P_*\geq\mathcal P(\gamma_0)$. The upper bound proved earlier in that paper gives $P_*\leq\inf_{\gamma\in\mathcal U}\mathcal P(\gamma)$. Therefore
$$
P_*\leq\inf_{\gamma\in\mathcal U}\mathcal P(\gamma)
\leq\mathcal P(\gamma_0)\leq P_*.
$$
Thus $\gamma_0$ attains the infimum. The cancellation is a step in the imported existence proof and adds no assumption about the Stieltjes measure $d\gamma_0$. In particular, $\nu_0$ here is the limit of coefficient-times-Lebesgue measures, distinct from the order parameter's Stieltjes measure, and the two must not be identified. [Auffinger and Chen, arXiv:1606.05335v2, pp. 7–12][ac17].

The cited theorem is stated for the almost-sure normalized maximum. Its expected-value formulation used here follows from Gaussian concentration. For
$L_n=n^{-1}\max_\sigma H_n(\sigma)$, the Lipschitz constant in the Gaussian couplings is
$\sqrt{(n-1)/(2n^2)}$. Hence $L_n-\mathbb E L_n\to0$ in probability. Since the almost-sure limit is the deterministic Parisi value, the deterministic sequence $\mathbb E L_n$ has the same limit. The full-support theorem below remains conditional on a minimizer, and neither of the selected formal interfaces constructs one. [OpenAI full-support manuscript, Introduction, ground-state formula and its footnote][fs-main].

Auffinger, Chen and Zeng proved infinite support, which still allows an
interior gap. Their terminal-plateau observation anticipates one endpoint
obstruction, and the argument below supplies the quantitative
boundary-layer proof used here.

The full-support manuscript develops a rational coercivity route and
credits Chen's 2026 theorem for the same full-support conclusion and
smooth absolute continuity. Chen's Proposition 4.2 has the conditional
crossing form $q''(t)=0\Rightarrow q'''(t)>0$. The cited rational
certificate gives $q'''(t)>0$ throughout a positive constant interval.
These are different sufficient mechanisms, and this record makes no
priority claim for full support.
[Auffinger, Chen and Zeng (2020), Theorem 1 and Remark 7][acz];
[Chen (2026), Theorem 1.3 and Proposition 4.2][chen2026];
[OpenAI (2026a), §§1–2][fs-main].

The transformed density comparisons also have antecedents in Lopatto's
positive-temperature analysis and Chen's zero-temperature analysis. The
rational coercive identity is the step that distinguishes the full-support
proof.

Zhou's positive-temperature interval-support result applies near
the critical temperature, whereas Lopatto's cited theorem covers the
low-temperature zero-field phase. Neither transfers full Stieltjes support
to zero temperature on its own. For the minimizers, $\beta\mu_\beta$
converges vaguely on $[0,1)$ to the zero-temperature Stieltjes measure,
while the unscaled Parisi probability measures converge weakly on $[0,1]$
to $\delta_1$. These are different measures and topologies, and
convergence without an interval-specific lower bound can lose support
information. The direct zero-temperature gap proof below is therefore
essential. [Chen (2026), Remarks 1.4–1.5][chen2026];
[OpenAI (2026a), §§1–2][fs-main].

<a id="finite-temperature-dictionary"></a>

The coefficient classes in this comparison can be translated explicitly.
In the positive-temperature convention, write
$$
U_t=-\frac{\beta^2}{2}(U_{yy}+\alpha_\nu U_y^2),
\qquad U(1,y)=\log\cosh y.
$$
Set $\Phi_\beta(t,x)=\beta^{-1}U(t,\beta x)$ and
$\gamma_\beta=\beta\alpha_\nu$. Direct differentiation gives
$$
(\Phi_\beta)_t=-\tfrac12\bigl((\Phi_\beta)_{xx}
 +\gamma_\beta(\Phi_\beta)_x^2\bigr),\qquad
\Phi_\beta(1,x)=\beta^{-1}\log\cosh(\beta x).
$$
Dividing the associated diffusion by $\beta$ likewise gives unit Brownian
noise and drift $\gamma_\beta(\Phi_\beta)_x$. Conversely, every bounded
nonnegative, nondecreasing, right-continuous coefficient on $[0,1)$ has
this form: choose $\beta>0$ at least its supremum, take distribution
function $\alpha_\nu=\gamma/\beta$ below one, and put the remaining
probability mass at one. This identifies coefficient classes, not
minimizers at different temperatures. The terminal values differ from
$|x|$ by at most $(\log2)/\beta$, but passing a strict inequality to the
cusp and to coefficients unbounded near one still needs uniform estimates.
A qualitative crossing statement alone supplies no positive limiting
fourth-moment margin. [Lopatto, arXiv:2607.11756v4, equation (1.2),
§2 and Proposition 5.1](https://arxiv.org/abs/2607.11756v4).

Two related works sharpen the comparison without entering the pinned
September proof. [Lopatto, arXiv:2607.11756v4, Proposition 5.1 and
(5.8)–Corollary 5.6](https://arxiv.org/abs/2607.11756v4), proves a
conditional curvature crossing on a positive constant interval. Its
integrating factor makes a transformed curvature derivative strictly
increasing, so the associated contact derivative decreases and then
increases. It does not assert an unconditional positive third derivative.

[Lopatto, arXiv:2610.00127v1, Theorem 2.3 and §8](https://arxiv.org/abs/2610.00127v1),
works at positive temperature and positive external field. Its comparison
requires one-sided bounds on the coefficient before and after the chosen
time, a moment inequality and a stationarity equality. Several of these conditions are automatic in its density application:
monotonicity supplies the one-sided bounds at a continuity point, contact
on a full-support interval makes the scalar inequality an equality, and
the quotient formula for the coefficient supplies stationarity. Its
moment bootstrap then gives smooth, strictly positive density on a
nontrivial support interval, and a singleton support is a separate case.

Appendix B, Theorem B.1 and equation (B.1), also prove strict positivity
in the zero-field positive-temperature setting, where the extra
noncentered-law condition is unnecessary.

These results are therefore close analytic antecedents. The distinction
used below is the
unconditional quantitative fourth-moment margin for the comparison
coefficient and its controlled passage to the zero-temperature cusp.

Here is an explicit example of the information lost in such a limit.
It concerns admissible measures, not Parisi minimizers. Put
$g=\mathbf1_{[0,1/3]\cup[2/3,1)}$ and $q_n=1-n^{-2}$, and for $n\ge2$ define
$$
\mu_n(dt)=\frac{g(t)+n^{-1}}n\mathbf1_{[0,q_n]}(t)\,dt
+\left(1-\frac1n\int_0^{q_n}(g(t)+n^{-1})\,dt\right)\delta_{q_n}(dt).
$$
The continuous part has mass at most $(2/3+1/n)/n<1$, so these are
probability measures with a positive endpoint atom. Their density is
positive throughout $[0,q_n]$, giving $\operatorname{supp}\mu_n=[0,q_n]$.
Their continuous mass tends to zero and $q_n\to1$, hence
$\mu_n\Rightarrow\delta_1$ weakly on $[0,1]$.
For a compactly supported continuous test function on $[0,1)$, the endpoint
atom eventually lies outside its support. Consequently
$n\mu_n\to g(t)\,dt$ vaguely there, although the limit vanishes on
$(1/3,2/3)$. Even the cumulative coefficients converge in the topology
used for the PDE: with $\gamma_n(t)=n\mu_n([0,t])$ and
$\gamma(t)=\int_0^t g(s)\,ds$, their difference is $t/n$ below $q_n$,
and $\gamma_n=n$ on the terminal strip of length $n^{-2}$. Thus
$$
\|\gamma_n-\gamma\|_{L^1(0,1)}\le\frac1{2n}+\frac1n\longrightarrow0.
$$
Every $\gamma_n$ is admissible, but $d\gamma$ has an interior gap.
Minimality and the gap argument supply information that these convergences
alone do not. An atom added at time one is likewise invisible to the
coefficient on $[0,1)$ and hence to its PDE and variational functional,
so the zero-temperature problem selects no endpoint mass.

Control and gradient-martingale methods have antecedents in
[Auffinger–Chen](https://doi.org/10.1007/s00220-014-2254-z) and
[Jagannath–Tobasco](https://arxiv.org/abs/1502.04398v4). A finite-temperature
theorem with smooth terminal value $\log\cosh$ does not alone establish
the result with terminal cusp $|x|$. The required compact-strip and
terminal estimates are supplied in [OpenAI (2026a), §2][fs-reg];
[Chen–Handschy–Lerman](https://arxiv.org/abs/1609.04368v4) supplies earlier
zero-temperature regularity, control and uniqueness results (Theorem 4
for uniqueness). This is distinct from the earlier finite-temperature
uniqueness result of Auffinger and Chen.
[Montanari](https://arxiv.org/abs/1812.10897v2) and
[El Alaoui–Montanari–Sellke](https://arxiv.org/abs/2001.00904v1) inform
normalised discretisation and optimisation. The particular shifted Gaussian
coefficient identity used here is proved in [OpenAI (2026a), §7][fs-cons].
<!-- /passage:dependencies -->

<a id="strip-estimates"></a>

### 2.4 Recovering the compact-strip estimates

<!-- passage:strip-recovery -->
The curvature argument needs estimates uniform over the approximating
coefficients. We derive them from the control formula, with every bound
taken on a fixed strip ending before time one.

For finite-step coefficients and a smooth, even, convex, one-Lipschitz
terminal function $\psi$, the evolution through a step of height $c$ is
$$
\Phi(t,x)=
\begin{cases}
c^{-1}\log P_{b-t}(e^{c\Phi(b,\cdot)})(x),&c>0,\\
P_{b-t}\Phi(b,\cdot)(x),&c=0,
\end{cases}
$$
where $P_h$ denotes convolution with a centred Gaussian of variance $h$.
The heat formula preserves evenness, convexity and the spatial Lipschitz
bound. Along a controlled path $dY_s=\gamma(s)\alpha_s\,ds+dB_s$, the drift
of the value minus the accumulated control cost is
$-\gamma(s)(\alpha_s-\Phi_x(s,Y_s))^2/2$. It is nonpositive, and is zero
for the feedback control. This proves the stated control formula for the
smooth step problem, including attainability by its feedback diffusion.

Using the same control in two such problems and then taking suprema gives
$$
\|\Phi_{\gamma,\psi}-\Phi_{\widetilde\gamma,\widetilde\psi}\|_\infty
\le \|\psi-\widetilde\psi\|_\infty
       +\frac32\|\gamma-\widetilde\gamma\|_{L^1(0,1)}.
$$
The drift part costs at most the coefficient difference in $L^1$, by the
terminal Lipschitz bound, and the quadratic cost contributes another half of
that difference. The supremum norm is of the difference of the values,
which is bounded even though each value grows linearly in $x$.
With $\psi_M(x)=M^{-1}\log(2\cosh Mx)$,
$0\le\psi_M-|x|\le(\log2)/M$. Choose simultaneously $\gamma_m\to\gamma$ in $L^1$ and $M_m\to\infty$.
These are two distinct requirements: $M_m\ge\|\gamma_m\|_\infty$ alone
does not force terminal convergence. This compatible approximation
determines a unique uniform limiting value. Its terminal bound,
given in §2.2, also follows directly: choose the zero control for the lower
bound, and use $|\alpha|-\alpha^2/2\le1/2$ for the upper bound.

To obtain spatial derivatives, choose $T<b<1$. Monotonicity gives the
uniform coefficient estimate
$$
\gamma(t)\le\frac{\|\gamma\|_1}{1-b}\quad(0\le t\le b).
$$
An $L^1$-convergent family thus has a common coefficient bound $\Gamma$ on
this larger strip. For its smooth step approximants, put $U=\Phi_x$.
The differentiated mild equation is
$$
U(t)=P_{b-t}U(b)+\int_t^b\gamma(s)P_{s-t}[U(s)U_x(s)]\,ds.
$$
Since $\|U\|_\infty\le1$ and
$\|\partial_xP_hF\|_\infty\le Ch^{-1/2}\|F\|_\infty$,
$$
\|U_x(t)\|_\infty\le\frac C{\sqrt{b-t}}
 +C\Gamma\int_t^b\frac{\|U_x(s)\|_\infty}{\sqrt{s-t}}\,ds.
$$
Substituting this inequality once into its own integral uses
$\int_t^s((r-t)(s-r))^{-1/2}\,dr=\pi$. It reduces the estimate to an
ordinary backward Gronwall inequality with integrable forcing. This bounds
$U_x$ uniformly on $[0,T]$ without any terminal derivative bound. Repeating
on successively smaller strips bounds each higher spatial derivative: put
one derivative on the heat kernel and the rest on $UU_x$, and only the term
$U\partial_x^mU$ contains the new highest derivative. The remaining products
have already been bounded.

Uniform convergence of the values then yields convergence of every
positive spatial derivative. For a bounded twice differentiable difference
$F$ with bounded second derivative, Taylor's formula gives
$$
\|F'\|_\infty\le \frac{2\|F\|_\infty}{h}
                         +\frac h2\|F''\|_\infty\qquad(h>0).
$$
First choose $h$ small using the common second-derivative bound, then make
$\|F\|_\infty$ small. Iterate on derivative differences. Passing to the
integral form of the differentiated PDE proves local absolute continuity
in time and its almost-everywhere equations. No classical time derivative
at a coefficient jump, and no derivative of $|x|$ at its cusp, is being
assumed.

Finally, couple all feedback diffusions with the same Brownian motion.
On $[0,T]$, let $L$ be a common spatial Lipschitz bound for their drifts.
The difference of the two integral equations gives, for example,
$$
\sup_{s\le T}|X_s^n-X_s|
\le e^{LT}\left(\|\gamma_n-\gamma\|_1
 +\|\gamma_n\|_1\|u_n-u\|_{\infty,[0,T]\times\mathbb R}\right)
 \longrightarrow0.
$$
This is a deterministic bound under the coupling. The derivative bounds
one order higher ensure uniform continuity when derivatives are evaluated
at these moving points. Every fixed polynomial in the bounded derivatives
therefore converges in expectation. This is precisely the convergence
needed for the integrated curvature expression in §4, and none of the forward
density scores or their quotients need converge.
[OpenAI (2026a), Lemmas 2.1–2.3][fs-reg].
<!-- /passage:strip-recovery -->

<!-- passage:approximation-choices -->

### 2.5 The order of the approximation limits

Each limit below fixes its background objects first, so that one symbol for
a small error never hides a change of object. A single admissible minimiser
and its associated diffusion stay fixed throughout the scalar construction.

**Compact-strip regularity.** Fix a derivative order $k\ge1$, a cutoff
$T<1$ and a common coefficient $L^1$ bound. Coefficients converge in $L^1$
within that bound, and spatial derivatives converge uniformly on the strip.

**Shape inequalities on a plateau.** Fix the height $c>0$ and a closed
interior neighbourhood. Choose step coefficients equal to $c$ there and
smooth terminal values with $M_m\ge\|\gamma_m\|_\infty$ and
$M_m\to\infty$. The coefficient and terminal approximations may be
chosen as one compatible sequence.

**First variation.** Fix $\gamma$, a feasible $\eta\in L^1$ and initially
$\epsilon>0$. Remove the terminal cutoff before letting
$\epsilon\downarrow0$. The terminal remainder divided by $\epsilon$ is
bounded by an $L^1$ tail of $\eta$.

**Terminal value.** Keep the chosen minimiser and diffusion fixed. Bounded
martingale convergence and integrability of $\gamma$ justify
$t\uparrow1$ without any terminal derivative bound.

**Finite Gaussian coefficients.** Fix $\gamma,X$ and $T<1$. Refine $N$
using the fixed-strip bounds, then control the remaining value tail by $1-T$.

For the plateau transfer, tail estimates for a fixed approximant may depend
on that approximant. The passage to the limit uses only the integrated
polynomial in the scaled second, third and fourth spatial derivatives of
the value function, together with the common strip bounds. The auxiliary
density scores need no uniform bounds.

Sections 3 and 4 give the variation and plateau-transfer arguments, and §7
gives a complete scalar error budget. The further choices used to produce a
circuit appear in §8.6.
[OpenAI (2026a), Lemma 2.1, Lemma 6.1, Proposition 5.1 and
Proposition 7.2][fs-main].
<!-- /passage:approximation-choices -->

<a id="variational-signs"></a>

## 3. From minimality to contact

<!-- passage:contact -->
**Proposition 3.1 (variational contact).** For an admissible minimizer,
the potential $G(s)=\int_s^1(q(t)-t)\,dt$ is nonnegative on $[0,1)$
and vanishes on $S=\operatorname{supp}\mu_\gamma$. At every $s\in S$,
$q(s)=s$ and $q'(s)\le1$, with a right derivative at zero.

The minimisation takes place over nonnegative Stieltjes measures with an
integrable cumulative function, and their total mass is free. We first
derive the variation of the functional, then use changes of measure which
preserve this admissible class. The result is contact on the *existing*
support. Gaps are excluded afterwards, in §5.

<a id="gradient-moment"></a>

### 3.1 The moment function before minimisation

Fix any $\gamma\in\mathcal U$, without yet assuming that it minimises
$\mathcal P$. Compact-strip regularity and Itô's formula give
$$
u(t,X_t)=\int_0^t a(s,X_s)\,dB_s,\qquad
q(t):=\mathbb E u(t,X_t)^2
     =\int_0^t\mathbb E a(s,X_s)^2\,ds.
$$
The initial term vanishes because zero field makes $\Phi$ even and
$u(0,0)=0$. Thus $q(0)=0$, $0\le q(t)\le1$, and
$$
q'(t)=\mathbb E a(t,X_t)^2\qquad(0\le t<1),
$$
with a right derivative at zero. The integrand is continuous on each
compact time strip, so $q\in C^1[0,1)$. These facts hold for every
admissible coefficient, and the identity $q(t)=t$ is still to be proved.
The source proves them by smooth approximation on $[0,T]$, $T<1$, and
never differentiates the terminal cusp $|x|$ at time one.
[OpenAI (2026a), Lemmas 2.1–2.2][fs-reg].

<a id="feasible-variation"></a>

### 3.2 The feasible right derivative

Let $\eta\in L^1(0,1)$ and suppose that
$\gamma+\varepsilon\eta\in\mathcal U$ for
$0\le\varepsilon\le\varepsilon_0$, where $\varepsilon_0>0$.
This is a feasibility hypothesis: an arbitrary signed bump in $\gamma$
need not preserve monotonicity. Since $\mathcal U$ is convex, it is enough
to check that the endpoint $\gamma+\varepsilon_0\eta$ is admissible.
The formula below is the zero-field, $\xi''=1$ specialization of
[Chen, Handschy and Lerman, arXiv:1609.04368v4, Proposition 4](https://arxiv.org/abs/1609.04368v4). Their chord
$\gamma_\theta=(1-\theta)\gamma+\theta\gamma_1$ gives this direction
by taking $\gamma_1=\gamma+\varepsilon_0\eta$ and
$\theta=\varepsilon/\varepsilon_0$. Their class $\mathcal U$ permits
unnormalized cumulative measures with finite cumulative integral, so
this change of parameter also covers the scaling direction $-\gamma$.
We reconstruct the detailed PDE proof from OpenAI's Lemma 6.1.

For $0<\varepsilon\le\varepsilon_0$, put
$\beta=\gamma+\varepsilon\eta$,
$\widetilde u=\Phi_{\beta,x}$ and $D=\Phi_\beta-\Phi_\gamma$.
Subtracting the two PDEs gives
$$
D_t=-\frac12D_{xx}
    -\frac12\beta(\widetilde u+u)D_x
    -\frac{\varepsilon}{2}\eta u^2.
$$
The choice of diffusion
$$
dY_t^\varepsilon
 =\frac12\beta(t)(\widetilde u+u)(t,Y_t^\varepsilon)\,dt+dB_t,
\qquad Y_0^\varepsilon=0,
$$
cancels the first-order spatial term when Itô's formula is applied to
$D(t,Y_t^\varepsilon)$. Its drift has absolute value at most $\beta(t)$,
which is integrable. On $[0,T]$ the resulting identity is
$$
D(0,0)=\mathbb E D(T,Y_T^\varepsilon)
 +\frac{\varepsilon}{2}\int_0^T
       \eta(t)\mathbb E[u(t,Y_t^\varepsilon)^2]\,dt.
$$
At coefficient jumps the time derivatives need only exist almost
everywhere. The source obtains this identity by approximating both
coefficients by monotone step functions with the same smooth terminal
datum, and then passing to the limit using the compact-strip estimates.

The control representation, applied with the same control on each side,
gives the time-local difference bound
$$
\sup_x|D(T,x)|\le\frac{3\varepsilon}{2}
                         \int_T^1|\eta(t)|\,dt.
$$
Hence $T\uparrow1$, at fixed $\varepsilon$, removes the terminal term
even after division by $\varepsilon$. We obtain the exact finite-change
identity
$$
\frac{\Phi_{\gamma+\varepsilon\eta}(0,0)-\Phi_\gamma(0,0)}
     {\varepsilon}
 =\frac12\int_0^1\eta(t)
                       \mathbb E[u(t,Y_t^\varepsilon)^2]\,dt.
$$
As $\varepsilon\downarrow0$, the coefficients converge in $L^1$ and have
uniformly bounded $L^1$ norms. On each fixed strip, their gradients converge
uniformly and have common spatial Lipschitz bounds. Coupling with the same
Brownian motion and applying Gronwall's inequality therefore gives
$Y^\varepsilon\to X$ uniformly on that strip. The integrand converges to
$\eta(t)q(t)$ for each $t<1$, and its absolute value is at most
$|\eta(t)|$. Dominated convergence, followed by subtraction of the linear
Parisi penalty, now yields
$$
D_+\mathcal P(\gamma)[\eta]
 :=\lim_{\varepsilon\downarrow0}
     \frac{\mathcal P(\gamma+\varepsilon\eta)-\mathcal P(\gamma)}
          {\varepsilon}
 =\frac12\int_0^1\eta(t)(q(t)-t)\,dt.
$$
If $\gamma$ is a minimiser, this derivative is nonnegative for every
feasible direction. It need not vanish: the minimisation has one-sided
constraints. [OpenAI (2026a), Lemma 6.1, equations (6.1)–(6.2)][fs-var].

<a id="measure-contact"></a>

### 3.3 Admissible changes of measure

Write $\mu=\mu_\gamma$ and $S=\operatorname{supp}\mu$, with support
relative to $[0,1)$. Tonelli's theorem gives the useful identity
$$
\int_0^1\gamma(t)\,dt
 =\int_{[0,1)}\!\int_s^1dt\,\mu(ds)
 =\int_{[0,1)}(1-s)\,\mu(ds)<\infty.
$$
This is a weighted finiteness condition, and $\mu([0,1))$ may be infinite.
In particular, mass added at one point need not be removed elsewhere, since
no normalisation constrains the total.

Define the tail potential
$$
G(s)=\int_s^1(q(t)-t)\,dt\qquad(0\le s<1).
$$
It is continuous, satisfies $|G(s)|\le1-s$, and has
$G'(s)=s-q(s)$ and $G''(s)=1-q'(s)$ on compact strips, with one-sided
interpretations at zero. The following two competitors suffice.

| Change to the Stieltjes measure | Direction in the cumulative function | Allowed parameter | Consequence at a minimiser |
| --- | --- | --- | --- |
| Add $\varepsilon\delta_s$, $0\le s<1$ | $\eta(t)=\mathbf1_{[s,1)}(t)$ | $\varepsilon\ge0$ | $G(s)\ge0$ |
| Remove $\varepsilon\nu$, where $\nu=\mu|_K$ and $K\subset[0,1)$ is compact | $\eta_K(t)=-\nu([0,t])$ | $0\le\varepsilon\le1$ | $\int_KG(s)\,\mu(ds)\le0$ |

For atom addition, the new cumulative function is nonnegative,
nondecreasing and right-continuous. Its integral increases by
$\varepsilon(1-s)$, so it remains in $\mathcal U$. This includes $s=0$:
an atom there adds a constant to $\gamma$ on all of $[0,1)$. The point
$1$ is outside the measure's domain and is not an additional competitor.
The derivative formula gives
$D_+\mathcal P(\gamma)[\mathbf1_{[s,1)}]=G(s)/2$, proving
$G\ge0$ everywhere below one.

For removal, choose $T<1$ with $K\subset[0,T]$. Then
$\nu(K)\le\gamma(T)<\infty$. The measure
$\mu-\varepsilon\nu$ is nonnegative for $0\le\varepsilon\le1$,
so its cumulative function is nonnegative, nondecreasing and
right-continuous. It lies between zero and $\gamma$, hence is integrable,
and $|\eta_K|\le\nu(K)$. Fubini is justified quantitatively by
$$
\int_K\!\int_s^1|q(t)-t|\,dt\,\mu(ds)
 \le\int_K(1-s)\,\mu(ds)\le\nu(K)<\infty.
$$
It follows that
$$
0\le D_+\mathcal P(\gamma)[\eta_K]
 =-\frac12\int_0^1\nu([0,t])(q(t)-t)\,dt
 =-\frac12\int_KG(s)\,\mu(ds).
$$
Since $G\ge0$, its integral on every compact $K$ is zero. The compact
intervals $[0,T]$ exhaust $[0,1)$, so $G=0$ $\mu$-almost everywhere.
If a support point $s$ had $G(s)>0$, continuity would give a relative
neighbourhood on which $G$ stays positive. After restricting to a compact
interval containing that neighbourhood, the defining property of support
would give it positive $\mu$-mass, contradicting the zero integral.
Therefore
$$
G(s)=0\qquad(s\in S).
$$
This step applies to atoms, diffuse mass and singular continuous mass.
Removing individual atoms alone would not establish it.
[OpenAI (2026a), Lemma 6.2][fs-var].

<a id="contact-slopes"></a>

### 3.4 Contact values and endpoint slopes

At $s\in S\cap(0,1)$, the function $G$ has an interior minimum of
value zero. Thus $G'(s)=0$ and $G''(s)\ge0$, giving
$$
q(s)=s,\qquad q'(s)\le1.
$$
If $0\in S$, the argument needs the zero-field initial condition in
addition to the one-sided minimum. We have $G(0)=0$ and $q(0)=0$, so
$$
G(h)=G(0)+\int_0^h(t-q(t))\,dt
     =\frac12(1-q'(0))h^2+o(h^2).
$$
Nonnegativity for $h>0$ implies $q'(0)\le1$. This works whether or not
zero is an atom. The equality $q(0)=0$ comes from symmetry, rather than
from differentiating a one-sided minimum. If $0\notin S$, no condition
$G(0)=0$ has been obtained or used.
[OpenAI (2026a), Lemma 6.2][fs-var].

We have now proved, without any no-gap assumption,
$$
G\ge0\text{ on }[0,1),\qquad G=0\text{ on }S,\qquad
q(s)=s\text{ and }q'(s)\le1\text{ for }s\in S.
$$
These are precisely the contact data used in §5. They make no assertion
that every zero of $G$ belongs to $S$, or that $q(t)=t$ at a point outside
$S$. Terminal, interior and initial gaps remain to be excluded.

### 3.5 The reach of the first-order criterion

<a id="global-scaling-contact"></a>

The gap proof needs only the contact data above, so a reader may continue
to [the curvature argument in §4](#curvature-and-gap-certificate). We give
here an alternative contact proof and the converse optimality criterion.

There is also a global alternative to compact mass removal. Atom addition
still gives $G\ge0$. Now replace $\mu$ by $(1-\varepsilon)\mu$,
or equivalently $\gamma$ by $(1-\varepsilon)\gamma$, for
$0\le\varepsilon\le1$. This preserves admissibility, and its direction
$-\gamma$ lies in $L^1$, although it may be unbounded. The first
variation in §3.2 therefore applies. The signed interchange is justified by
$$
\int_{[0,1)}\!\int_s^1|q(t)-t|\,dt\,\mu(ds)
 =\int_0^1\gamma(t)|q(t)-t|\,dt
 \le\int_0^1\gamma(t)\,dt<\infty.
$$
Consequently
$$
0\le D_+\mathcal P(\gamma)[-\gamma]
 =-\frac12\int_{[0,1)}G(s)\,\mu(ds).
$$
Together with $G\ge0$, this gives $\int G\,d\mu=0$. Continuity
then gives $G=0$ on relative support, including zero, as in §3.3.
Finite weighted mass is sufficient throughout, with no assumption of
finite total mass or of a density for $\mu$. This ordinary argument uses the
same integrable-direction first variation as the local removal proof.

For the actual moment function associated with $\gamma$, the two
potential conditions $G\ge0$ and $G=0$ on $S$ are also sufficient for
minimality. This ordinary deduction explains exactly how much variational
information has been extracted. For another admissible coefficient
$\beta$, let $\nu=\mu_\beta$. The weighted identity above and
$|G(s)|\le1-s$ make both measure integrals absolutely convergent, even
when either measure has infinite total mass. Fubini gives
$$
2D_+\mathcal P(\gamma)[\beta-\gamma]
 =\int_{[0,1)}G(s)\,\nu(ds)
  -\int_{[0,1)}G(s)\,\mu(ds)\ge0.
$$
Here the second integral is zero because $\mu$ is concentrated on its
relative support. The control formula also shows that $\mathcal P$ is
convex in its coefficient: for a fixed control, the terminal absolute
value is convex in the coefficient and both penalties are linear, and taking
the supremum over the same control class preserves convexity. Therefore
$$
\mathcal P(\beta)-\mathcal P(\gamma)
 \ge D_+\mathcal P(\gamma)[\beta-\gamma]\ge0.
$$
This proves the converse. It agrees with the chord optimality criterion
of [Chen, Handschy and Lerman, arXiv:1609.04368v4, Proposition 4,
printed p. 14](https://arxiv.org/abs/1609.04368v4).
Chen's zero-temperature treatment obtains the same contact integral by
testing $0$ and $2\gamma$ and then using weighted Fubini
[Chen (2026), Proposition 3.2, equations (3.11)–(3.12)][chen2026].
The compact-removal proof above keeps the local mass operation explicit.

The converse concerns $G$ built from the PDE and diffusion of the same
$\gamma$. It does not say that an arbitrary curve with these contact
properties comes from a Parisi solution. Nor does it prove full support:
the missing input is the strict curvature on an interval where the
coefficient is constant, established next. The formal chord statement and
the ordinary derivative bridge are distinguished in §12.1.
<!-- /passage:contact -->

<a id="curvature-and-gap-certificate"></a>

## 4. Strict curvature on a constant interval

<!-- passage:curvature -->
The gap argument needs a property of the derivative $q'$, which is distinct
from convexity of the spatial function $\Phi$. We establish the following
analytic input before using minimality. Let $\gamma\in\mathcal U$ be
constant with value $c$ on an open interval $I\subset(0,1)$, and let $X$
be its diffusion started at zero. Then

$$
\begin{cases}
q'''(t)\ge \dfrac1{1000c^2}
  \mathbb E\bigl[(c\Phi_{xx}(t,X_t))^4\bigr]>0,&c>0,\\[4pt]
q''(t)=\mathbb E[\Phi_{xxx}(t,X_t)^2]>0,&c=0,
\end{cases}\qquad t\in I.
$$

Thus $q'$ is strictly convex for $c>0$ and strictly increasing for
$c=0$. The coefficient may jump outside $I$ and may be unbounded near
one, provided it is nonnegative, nondecreasing, right-continuous and
integrable. The statement holds for every such coefficient, whatever its
support and whether or not it minimises. The positive-time restriction
matters when a diffusion density is used. Interval endpoints are handled
later by continuity.
[OpenAI (2026a), Proposition 5.1][fs-curv].

The proof has two stages. In §§4.1–4.4 the coefficient is a positive
monotone finite-step function with smooth terminal data. Shape comparisons
and boundary estimates justify the [polynomial correction](#curvature-algebra)
and its zero expectation. In [§4.5](#plateau-limit), approximants
$\gamma_n,\Phi_n,X^n$ approach an arbitrary admissible coefficient.
Only the integrated expression in spatial derivatives passes to the limit,
and the auxiliary forward scores need not converge. Bounds depend on a
fixed compact strip and a common coefficient mass bound. Finally, we
justify the strictly positive right side for the limiting solution itself.

<a id="curvature-moments"></a>

### 4.1 Why the uncorrected expression has no pointwise sign

On a positive constant interval put
$r=c\Phi_x$, $v=r_x$, $w=r_{xx}$ and $z=r_{xxx}$.
In forward time,
$r_t=-\tfrac12r_{xx}-rr_x$ and $dX_t=r(t,X_t)\,dt+dB_t$.
Differentiating the PDE and applying Itô's formula gives

$$
dr(t,X_t)=v(t,X_t)\,dB_t,
$$
$$
dv(t,X_t)=-v(t,X_t)^2\,dt+w(t,X_t)\,dB_t,
\qquad
dw(t,X_t)=-3v(t,X_t)w(t,X_t)\,dt+z(t,X_t)\,dB_t.
$$

Since $q'=c^{-2}\mathbb E v^2$, the drift of $v^2$ yields
$c^2q''=\mathbb E(w^2-2v^3)$. The drifts of $w^2$ and $v^3$ are,
respectively, $z^2-6vw^2$ and $3vw^2-3v^4$. Therefore

$$
c^2q'''=\mathbb E\bigl[z^2-12vw^2+6v^4\bigr].
$$

These are interior-time identities. On a smaller constant interval the
heat formula gives classical time smoothness and bounded spatial
derivatives of sufficiently high order. Hence the local martingale terms
have zero mean, and the expectation differentiations are justified. The
identities leave the sign of $q''$ undetermined.
[OpenAI (2026a), Lemma 2.3][fs-reg].

A concrete example shows that the last display cannot be read
pointwise. For the smooth finite-step problem $c=M=1$ on $[0,1)$, with terminal
value $\log(2\cosh x)$, the exact solution is

$$
\Phi(t,x)=\log(2\cosh x)+\frac{1-t}{2},\qquad r=\tanh x.
$$

Indeed, $\Phi_{xx}+\Phi_x^2=1$. Writing $R=r^2$ gives
$v=1-R$, $w=-2rv$ and $z=-2v^2+4Rv$. At $R=1/2$,
$v=1/2$, $w^2=1/2$ and $z=1/2$, so

$$
z^2-12vw^2+6v^4=-\frac{19}{8}<0.
$$

This example lies in the smooth analytic class used below. Its diffusion
density, obtained from the heat transition formula, is
$f_t(x)=e^{-t/2}\cosh(x)p_t(x)$, and it integrates to one because
$\mathbb E\cosh(B_t)=e^{t/2}$.

The example shows that the integration argument is necessary. It uses a
smooth terminal value and makes no claim to be the zero-temperature
minimiser. The negative value above is an exact elementary calculation,
and the heat formula itself is [OpenAI (2026a), Lemma 2.1][fs-reg].

<a id="pointwise-curvature-obstruction"></a>

The same pointwise obstruction occurs with the actual terminal datum
$|x|$. For the admissible constant coefficient $\gamma_c\equiv c>0$,
write $\tau=1-t>0$ and use the same heat formula:
$$
\Phi_c(t,x)=c^{-1}\log\mathbb E e^{c|x+\sqrt\tau Z|}.
$$
At $c=0$ its continuous extension is the heat value, whose derivatives are
$$
a_0=\sqrt{\frac2{\pi\tau}}e^{-x^2/(2\tau)},\qquad
b_0=-\frac{x}{\tau}a_0,\qquad
d_0=\left(\frac{x^2}{\tau^2}-\frac1\tau\right)a_0.
$$
At $x=\sqrt\tau$, $d_0=0$ and $b_0^2=a_0^2/\tau>0$.
For fixed $t,x$, Gaussian convolution may be differentiated in $x$
through its kernel and expanded in $c$ near zero, since exponential moments
dominate every resulting polynomial factor. The apparent singularity
of $c^{-1}\log\mathbb E e^{c|x+\sqrt\tau Z|}$ is removable, and its
spatial derivatives through order four are smooth in $c$. Hence
$$
d_c^2-12c a_cb_c^2+6c^2a_c^4
 =-\frac{12a_0^3}{\tau}c+O(c^2)<0
$$
for all sufficiently small positive $c$, at that fixed point.
The sign follows from the constant-coefficient formula, in which all four
derivatives come from one solution. The example asserts nothing about the
minimality of $\gamma_c$ and is consistent with the positive expectation
proved below.

It shows that, even with the absolute-value terminal value, that
expectation cannot be replaced by a pointwise inequality.

### 4.2 The smooth finite-step problem and its boundary control

Take $0=t_0<\cdots<t_N=1$,
$0<c_1\le\cdots\le c_N\le M$, and
$\gamma(t)=c_i$ on $[t_{i-1},t_i)$. The terminal value is
$M^{-1}\log(2\cosh Mx)$. On a step of height $c$, set $l=c\Phi$ and
$h=e^l$. Then $h_t=-\tfrac12h_{xx}$. For $a<t$ in that step the
diffusion transition density is

$$
p_{t-a}(x-y)\frac{h(t,x)}{h(a,y)},\qquad
p_t(x)=(2\pi t)^{-1/2}e^{-x^2/(2t)}.
$$

Its mass is one by $h(a,\cdot)=P_{t-a}h(t,\cdot)$. Factoring the
positive-time density as $f=e^l g$ gives $g_t=\tfrac12g_{xx}$ on each
step. Put $s=-\partial_x\log g$. On the first step,
$g(t,x)=p_t(x)e^{-l(0,0)}$ and $s=x/t$. At a forward jump
$\delta=c_+-c_-\ge0$, continuity of $f$ gives
$g_+=g_-e^{-\delta\Phi}$. The functions $r$ and $s$ are odd.
[OpenAI (2026a), §3, before Lemma 3.1][fs-signs].

The comparison arguments take place on unbounded spatial domains, so
the boundary estimates must precede the signs. For these fixed smooth
data, Gaussian convolution and the finite jump rules give, as
$x\to+\infty$,

$$
l(t,x)=cx+C(t)+O(e^{-\kappa x}),
$$
$$
g(t,x)=\exp\!\left(-\frac{x^2}{2t}+\beta(t)x+C_0(t)\right)
       (1+O(e^{-\kappa x})).
$$

The errors may be differentiated any fixed finite number of times, and
the estimates are uniform on each closed step, with positive time
bounded away from zero for the forward estimate. Their constants may
depend on the step data.

For example, after completing the square in a
Gaussian convolution, the error is an expectation at
$Y_x=\lambda x+b+\sqrt VZ$ with $\lambda$ bounded below by a positive
constant. Split at $Y_x=\mathbb EY_x/2$. The first part inherits the
exponential tail. On the second, the Gaussian probability dominates the
permitted exponential growth of the error and all its specified
derivatives. This proves the differentiated estimates by finite induction
through heat evolution and jumps. In particular,
$s=x/t-\beta(t)+O(e^{-\kappa x})$, $s_x=1/t+O(e^{-\kappa x})$ and
$s_{xx}=O(e^{-\kappa x})$.
[OpenAI (2026a), Lemma 3.1][fs-signs].

The comparison principle used below is also explicit. If
$U_\tau-\tfrac12U_{xx}-bU_x-dU\ge0$ on a compact cylinder,
$d\le C$, the initial values are nonnegative and the lateral values are
at least $-\varepsilon$, then
$U\ge-\varepsilon e^{(\max(C,0)+1)(\tau-\tau_0)}$.
Multiply by the corresponding decreasing exponential and apply the
minimum principle. Exhaustion is legitimate when the upper bound for
$d$ is independent of the cylinder and the lateral errors tend to zero.
The first-order coefficient need only be bounded on each cylinder, which
permits the $1/x$ coefficients that appear below.
[OpenAI (2026a), comparison preceding Lemma 3.2][fs-signs].

<a id="shape-comparisons"></a>

### 4.3 Backward and forward comparisons, including jumps

We work in backward time $\tau$ on a single step, where the equation is
$r_\tau=\tfrac12r_{xx}+rr_x$. The logarithmic heat formula preserves
strict positivity of $v$: its second spatial derivative is the tilted
mean of the old second derivative plus the tilted variance of the old
first derivative. Initially $v=cM\operatorname{sech}^2(Mx)>0$, and
positive rescaling at a backward jump preserves this property.
Consequently $r(x)>0$ for $x>0$.

Twice differentiating gives
$w_\tau=\tfrac12w_{xx}+rw_x+3vw$. At the terminal endpoint
$w=-2cM^2\operatorname{sech}^2(Mx)\tanh(Mx)\le0$ for $x>0$.
The boundary value at zero is zero, and the differentiated tails make
the far boundary tend uniformly to zero. Comparison applied to $-w$
gives $w\le0$ on the positive half-line.

Define

$$
H=\frac{r_\tau}{r}=\frac{w}{2r}+v.
$$

Oddness and $r_x(0)=v(0)>0$ make $w/r$, hence $H$, smooth and even
at zero. Direct differentiation gives

$$
H_\tau=\tfrac12H_{xx}+(v/r+r)H_x+vH,
$$
$$
(H_x)_\tau=\tfrac12(H_x)_{xx}+(v/r+r)(H_x)_x
             +((v/r)_x+2v)H_x+wH.
$$

The terminal signs are
$H=(cM-M^2)\operatorname{sech}^2(Mx)\le0$ and $H_x\ge0$ on
$x>0$, using $M\ge c$. For the first comparison write
$v/r+r=1/x+\eta(x)$, where $\eta$ is smooth and odd near zero.
The radial extension $H(|y|)$ to $\mathbb R^3$ turns
$\tfrac12H_{xx}+x^{-1}H_x$ into $\tfrac12\Delta H$.
It is smooth at the origin, as is the extra drift, and $H$ tends
uniformly to zero at infinity. Comparison on expanding balls proves
$H\le0$.

For the second comparison, $wH\ge0$ and
$(v/r)_x=w/r-v^2/r^2<0$, so the zeroth-order coefficient has upper
bound $2\sup v$. On $[\varepsilon,A]$, the two lateral errors are
$H_x(\varepsilon)=O(\varepsilon)$ and $H_x(A)=o_A(1)$, uniformly
on the closed step. Taking $\varepsilon\downarrow0$ and
$A\to\infty$ gives $H_x\ge0$. At a backward jump
$r_{\rm new}=\alpha r_{\rm old}$ with $0<\alpha\le1$,

$$
H_{\rm new}=H_{\rm old}-(1-\alpha)v_{\rm old},\qquad
(H_x)_{\rm new}=(H_x)_{\rm old}-(1-\alpha)w_{\rm old}.
$$

Thus both signs survive every jump.
[OpenAI (2026a), Lemma 3.2][fs-signs].

For the forward score use $a_s=s_x$, temporarily distinguishing it
from the Hessian $a=\Phi_{xx}$. The forward heat equation gives

$$
s_t=\tfrac12s_{xx}-ss_x,\qquad
(a_s)_t=\tfrac12(a_s)_{xx}-s(a_s)_x-a_s^2,
$$
$$
(s_{xx})_t=\tfrac12(s_{xx})_{xx}-s(s_{xx})_x-3a_s s_{xx}.
$$

The first step has $s=x/t$. Thereafter $a_s-1/t$ satisfies the
linear equation with drift $-s$ and zeroth-order coefficient
$-(a_s+1/t)$. For fixed smooth data this coefficient is bounded above
on a closed step, independently of spatial truncation. This uses the
tail bound and continuity, before its sign is known. Comparison gives
$a_s\ge1/t$. The last displayed equation then gives $s_{xx}\le0$
on $x>0$. At a forward jump,

$$
s_+=s_-+\delta\Phi_x,\quad
(a_s)_+=(a_s)_-+\delta\Phi_{xx},\quad
(s_{xx})_+=(s_{xx})_-+\delta\Phi_{xxx}.
$$

The backward signs and $\delta\ge0$ preserve the forward signs.

Concavity alone is insufficient here, so we need one further comparison.
On $x>0$ set
$d_s=s_x/s$, $d_r=r_x/r$ and $\rho=d_s-d_r$.
The heat equation gives $s_t=\tfrac12s_{xx}-ss_x$, whereas $r_t=-rH$.
Taking logarithmic derivatives yields
$$
(d_s)_t=\tfrac12(d_s)_{xx}+d_s(d_s)_x-s_{xx},\qquad
(d_r)_t=-H_x.
$$
Since $\tfrac12(d_r)_{xx}+d_r(d_r)_x=\tfrac12(w/r)_x=H_x-w$,
substitution of $d_s=d_r+\rho$ gives its equation

$$
\rho_t=\tfrac12\rho_{xx}+d_s\rho_x+(d_r)_x\rho
                      +2H_x-w-s_{xx}.
$$

All three terms in the source are nonnegative by the comparisons
already proved, and $(d_r)_x<0$. On the first step, concavity gives
$r(x)=\int_0^xv(y)\,dy\ge xv(x)$, so
$\rho=1/x-v/r\ge0$. Odd Taylor expansions give
$d_s=1/x+O(x)$, $d_r=1/x+O(x)$ and $\rho=O(x)$ near zero, and
the tail estimates give $\rho\to0$ at infinity. Comparison on
$[\varepsilon,A]$ therefore proves $\rho\ge0$ on each subsequent
step provided its initial sign is preserved. Define

$$
j=s\rho=s_x-(s/r)v.
$$

At a forward jump, with
$\alpha=c_+/c_-$ and $\lambda=(c_+-c_-)/c_-$,
$r_+=\alpha r_-$, $v_+=\alpha v_-$,
$s_+=s_-+\lambda r_-$ and $(s_x)_+=(s_x)_-+\lambda v_-$.
Substitution cancels the added terms and gives $j_+=j_-$.
Since $s_+>0$ on $x>0$, the sign of $\rho$ is preserved.
The quotient $s/r$ extends smoothly and evenly across zero, with
$j(0)=0$. Reflection then gives $j\ge0$ everywhere.
[OpenAI (2026a), Lemma 3.3][fs-signs].

This ratio comparison also follows from the earlier lower-history
inequalities in [Zhou (2026), Appendix E.3, (E.37) and (E.41)][zhou-gardner].
To compare the finite-step problems, choose $M>\|\gamma_{\rm step}\|_\infty$
and put $T=M^2t$, $y=Mx$, $U(T,y)=M\Phi(t,x)$ and $m=c/M$.
Then $U$ has terminal data $\log(2\cosh y)$, whose additive constant
has no effect on the derivatives, and $0<m<1$. In Zhou's notation,
$B=U_y$, $C=U_{yy}$ and $V_y=y/T+W_y$. The correspondence is
$r=MmB$, $s=MV_y$, so
$$
j=M^2\left(V_{yy}-\frac{C V_y}{B}\right)
 =M^2\left(\frac{B-yC}{TB}
              +B\,\partial_y\frac{W_y}{B}\right)\ge0.
$$
Here $B\ge yC$ follows from decreasing $C$ on the positive half-line
(Zhou, p. 130), while (E.37) implies (E.41),
$\partial_y(W_y/B)\ge0$. At a mass jump $m\mapsto m+d$, Zhou's
update $W\mapsto W+dU$ in (E.28) adds $dB$ to $V_y$ and $dC$ to
$V_{yy}$, leaving the displayed expression for $j$ unchanged.
This identifies an earlier source for the ratio sign and its jump
invariance in the smoothed finite-step problem; the approximation and
coercivity arguments below are still needed for the full-support result.

<a id="curvature-algebra"></a>

### 4.4 The polynomial correction with zero expectation

Here the coefficient is a positive constant $c$ on the step containing
the chosen time. The variables are the rescaled spatial derivatives
$$
r=c\Phi_x,\qquad v=c\Phi_{xx},\qquad
w=c\Phi_{xxx},\qquad z=c\Phi_{xxxx}.
$$
Write $h=e^{c\Phi}$ and factor the forward diffusion density as $f=hg$,
with auxiliary score $s=-\partial_x\log g$. These are the backward
and forward quantities compared in §4.3. On the positive half-line,
where $r,s>0$, set
$$
H=\frac{w}{2r}+v,\qquad
j=s_x-\frac{s}{r}v,\qquad
\rho=\frac{s_x}{s}-\frac vr,
\qquad j=s\rho.
$$
In this certificate only,
$b=-w/r$ is a quotient. In the unscaled curvature formula, $b$ denotes
instead the derivative $\Phi_{xxx}$. At a fixed positive time
inside the step define

$$
R=r^2,\quad K=rs,\quad b=-w/r,\quad B=b-2v,
\quad L=-rs_{xx},\quad S_1=z+bv-2Rb.
$$

The subscript on $S_1$ distinguishes it from the support set. The
previous comparisons give

$$
R,K,v,B,j,L,S_1\ge0,\qquad b=B+2v\ge0,\qquad Rb^2=w^2,
$$

because $B=-2H$ and $S_1=2rH_x$. Here $s=-\partial_x\log g$
belongs to the forward heat factor, while $\partial_x\log f=r-s$
is the logarithmic derivative of the diffusion density. The signs have
separate sources: $B\ge0$ comes from $H\le0$, $S_1\ge0$ from
$H_x\ge0$ on the positive half-line, $j=s\rho\ge0$ from the
relative-slope comparison, and $L=-rs_{xx}\ge0$ from forward concavity.

Reflection extends these signs. The quotients were defined for $x>0$,
where $r>0$ and $v>0$ by the tilted mean-and-variance argument in
[§4.3](#shape-comparisons). At zero, $r$ and $w$ are odd,
$r_x(0)=v(0)>0$, and division by $r$ is therefore division by a function
with a simple zero. This gives the smooth even extension
$b(0)=-z(0)/v(0)$, and the other comparison identities give $j(0)=0$ and
$S_1(0)=0$. Thus the integration has no boundary at zero.

For $\mathcal D=r\partial_x$, direct differentiation gives

$$
\mathcal D(R,K,v,b,j)
 =(2Rv,\ 2vK+Rj,\ -Rb,\ -z-bv,\ -jv+Kb-L).
$$

Also $f_x/f=r-s$. Hence
$\mathcal W(F)=\mathcal DF+(v+R-K)F$ satisfies

$$
f\mathcal W(F)=\partial_x(rfF).
$$

By itself the identity leaves the boundary terms at infinity to be controlled.
The tail estimates give $r=O(1)$, exponential decay of $v,w,z,b$ and their
derivatives, $s=x/t+O(1)$, $j=1/t+O((1+x)e^{-\kappa x})$ and $K=O(1+x)$.
Every polynomial $F$ in $R,K,v,b,j$ and its differentiated expression
therefore has at most polynomial growth. Meanwhile
$f(t,x)\le C\exp(-x^2/(2t)+C|x|)$. Thus every term used below is
absolutely integrable and $rfF\to0$ at both infinities. Smooth
extension across zero prevents an interior boundary term. Integrating
on $[-A,A]$ and taking $A\to\infty$ now gives
$\mathbb E_f\mathcal W(F)=0$.
[OpenAI (2026a), Lemma 3.4][fs-signs].

The curvature estimate rests on four separate steps. The comparisons
supply the signs above, and the tail estimates make the correction
integrate to zero. The rational identity below gives a positive remainder,
and finite arithmetic verifies this third step. Finally,
[§4.5](#plateau-limit) passes the resulting expectation inequality to the
original coefficient.

Set

$$
(A,C,k,T_0,g_0,h_0)=
\left(\frac1{13},\frac{53}{24},\frac{19}{16},\frac{73}{48},
      \frac{2331}{10000},\frac{87}{2500}\right),
$$
$$
(x_0,y,m,n_0)=
\left(\frac{4663}{5000},\frac{6853}{10000},
      -\frac{1051}{400},-\frac{1073}{500}\right).
$$

Define

$$
Q=Avb+Cv^2+kKv-2kRb+T_0Rv,\qquad T_1=v(g_0R+h_0j),
$$
$$
\begin{aligned}
F_2={}&(Av-2kR)b^2
 +b\bigl(2Cv^2+2kKv+(g_0+2T_0)Rv+h_0jv\bigr)\\
 &+x_0v^2j+yv^3+mKv^2+n_0Rv^2.
\end{aligned}
$$

The terms containing $z$ motivate this choice. Indeed,
$$
\begin{aligned}
\partial_bF_2
 &=2(Av-2kR)b+2Cv^2+2kKv+(g_0+2T_0)Rv+h_0jv\\
 &=2Q+T_1,\\
{}[z]\mathcal W(F_2)&=-\partial_bF_2=-(2Q+T_1).
\end{aligned}
$$
Here $[z]$ means the coefficient of $z$, and only
$\mathcal Db=-z-bv$ contributes it. The square $(z+Q)^2$ and
$S_1T_1$ contribute $2Q+T_1$, so these terms cancel exactly.
Therefore subtracting
$(z+Q)^2+S_1T_1+\mathcal W(F_2)$ from the curvature expression
cancels every term involving $z$. The other favourable derivative is
$\partial_jF_2=x_0v^2+h_0bv\ge0$, so the $-L$ in $\mathcal Dj$
leaves a nonnegative correction.

Put

$$
\begin{aligned}
P_0&=Rj(b-7v)^2,&
P_1&=RB(b-\tfrac73v)^2,\\
P_2&=Kv(b-\tfrac92v)^2,&
P_3&=vB(K+\tfrac14R-\tfrac17v)^2,\\
P_4&=(vb-\tfrac{14}{5}v^2-\tfrac{13}{4}Kv+\tfrac43Rv)^2,
\end{aligned}
$$
$$
(p_0,p_1,p_2,p_3,p_4)=
\left(\frac{337}{10000},\frac{183}{2500},\frac{411}{10000},
      \frac{11461}{5000},\frac{647}{10000}\right).
$$

Exact expansion, with $b=B+2v$, gives

$$
z^2-12Rvb^2+6v^4
 =(z+Q)^2+S_1T_1+\sum_{i=0}^4p_iP_i
   +\mathcal W(F_2)+L(x_0v^2+h_0bv)+P_{\rm rem}.
$$

The table gives the entire remainder, and every unlisted coefficient is
zero.

| Monomial | Coefficient | Monomial | Coefficient |
| --- | ---: | --- | ---: |
| $R^2v^2$ | $223/160000$ | $R^2vB$ | $4007/240000$ |
| $R^2B^2$ | $1097/120000$ | $RKv^2$ | $83/80000$ |
| $RKvB$ | $3343/24000$ | $Rv^3$ | $16259/23400000$ |
| $Rv^2B$ | $30179/5460000$ | $Rv^2j$ | $26/625$ |
| $RvB^2$ | $107/15000$ | $RvBj$ | $3/2500$ |
| $RB^3$ | $121/32500$ | $RB^2j$ | $11/10000$ |
| $K^2v^2$ | $579/20000$ | $K^2vB$ | $207/2500$ |
| $Kv^3$ | $911/975000$ | $Kv^2B$ | $349/420000$ |
| $Kv^2j$ | $5011/5000$ | $KvB^2$ | $133/130000$ |
| $KvBj$ | $87/2500$ | $v^4$ | $1661507/1521000000$ |
| $v^3B$ | $633799/621075000$ | $v^2B^2$ | $10657/1690000$ |

All square prefactors $Rj,RB,Kv,vB,1$ are nonnegative, as are
$S_1T_1$, $L(x_0v^2+h_0bv)$ and every remainder monomial. Integration
removes only $\mathcal W(F_2)$, giving

$$
\mathbb E_f[z^2-12vw^2+6v^4]
 \ge\frac{1661507}{1521000000}\mathbb E_fv^4
 \ge\frac1{1000}\mathbb E_fv^4.
$$

The surplus of the first coefficient over $1/1000$ is exactly
$140507/1521000000>0$. The rational identity and the positivity of the
table reduce to finite arithmetic. Their application to this diffusion
also uses the signs, quotient behaviour, integrability and vanishing
flux proved above. Dividing the integrated inequality by $c^2$ returns
$$
\mathbb E[\Phi_{xxxx}^2-12c\Phi_{xx}\Phi_{xxx}^2
                  +6c^2\Phi_{xx}^4]
\ge\frac{c^2}{1000}\mathbb E\Phi_{xx}^4.
$$
This is the unscaled estimate needed by the gap argument. Section 4.5
passes it to the original coefficient and supplies strictness for the
limiting solution. [OpenAI (2026a), Lemma 4.1 and Proposition 4.2][fs-cert].

<!-- /passage:curvature -->

<!-- passage:fixed-time-coercivity -->
<a id="plateau-limit"></a>

### 4.5 Coercivity at a fixed time and on plateaus

<a id="fixed-time-coercivity"></a>

**Proposition 4.1 (coercivity at a fixed time).** Let $\gamma\in\mathcal U$
and let $X$ be its associated diffusion, started at zero. Suppose that
$0<t<1$, $\gamma$ is continuous at $t$, and $c=\gamma(t)>0$. With
$a=\Phi_{xx}$, $b=\Phi_{xxx}$ and $d=\Phi_{xxxx}$, evaluated at
$(t,X_t)$, one has
$$
\mathcal C_\gamma(t):=\mathbb E[d^2-12cab^2+6c^2a^4]
 \ge \frac{c^2}{1000}\mathbb E a^4.
\tag{4.1}
$$
The proposition assumes neither minimality nor a constant neighbourhood
of $t$, and it is the common estimate behind plateau curvature and
positive density.

When $\gamma$ is smooth near $t$, Itô's formula gives
$q''=\mathbb E b^2-2\gamma\mathbb E a^3$ and
$$
(\mathbb E b^2)'=\mathbb E d^2-6\gamma\mathbb E(ab^2),\qquad
(\mathbb E a^3)'=3\mathbb E(ab^2)-3\gamma\mathbb E a^4.
$$
Differentiating the first identity therefore gives
$$
q'''(t)+2\gamma'(t)\mathbb E a^3=\mathcal C_\gamma(t).
\tag{4.2}
$$
On a plateau the coefficient derivative vanishes. At a smooth minimizer,
$q=t$ makes the third derivative of $q$ vanish instead.

**Proof.** Fix $t<T<T'<1$ and a shrinking interval
$[t-\delta_n,t+\delta_n)\subset(0,T)$. Approximate $\gamma$ by
$$
h_n=\max\{\epsilon_n,\min\{K_n,
                  \eta_n\lfloor\gamma/\eta_n\rfloor\}\},
$$
where $0<\epsilon_n<c<K_n$, $\epsilon_n,\eta_n\downarrow0$ and
$K_n\uparrow\infty$. Use $\min(c,h_n)$ to the left of the interval,
$c$ on it, and $\max(c,h_n)$ to its right. Monotonicity makes these
positive, nondecreasing, right-continuous finite-step coefficients
$\gamma_n$. Clipping toward $c$ outside the interval cannot increase
the error, and
$$
\|\gamma_n-\gamma\|_1
\le\epsilon_n+\eta_n+\int_0^1(\gamma-K_n)_+\,ds
 +2\delta_n\sup_{|s-t|\le\delta_n}|\gamma(s)-c|
 \longrightarrow0.
\tag{4.3}
$$
Integrability controls the capped tail and continuity at $t$ the last
term. Choose smooth terminal parameters $M_n\ge\max(n,K_n)$.
The terminal values then converge uniformly to $|x|$, and the coefficient
masses have a common bound. The finite-step inequality of §4.4 applies
at $t$, where every $\gamma_n$ equals the same positive $c$.

On a fixed strip containing $t$ and ending before one, compact-strip
regularity gives uniform spatial convergence and bounds through order
four. The next derivative bounds give a common spatial Lipschitz
constant for those derivatives. Coupling with the same Brownian motion
and applying Gronwall gives a deterministic bound
$\sup_{s\le T}|X_s^n-X_s|\to0$. If $D_k^n=\partial_x^k\Phi_n$
and $D_k=\partial_x^k\Phi$, then, for $1\le k\le4$,

$$
|D_k^n(t,X_t^n)-D_k(t,X_t)|
\le\|D_k^n-D_k\|_{\infty,[0,T]\times\mathbb R}
   +C_{k,T}|X_t^n-X_t|\longrightarrow0.
$$

After the nonnegative terms are discarded, the two surviving integrands
$z_n^2-12v_nw_n^2+6v_n^4$ and $v_n^4$ are uniformly bounded,
with $v_n=cD_2^n$, $w_n=cD_3^n$ and $z_n=cD_4^n$. Dominated convergence therefore passes their expectations
to the limit.

To make this quantitative, let $V,W,Z$ bound both the limiting and approximating
$v,w,z$, and let $e_v,e_w,e_z$ bound their coupled evaluation errors.
Uniform spatial convergence and the path bound above give these errors
with $e_v,e_w,e_z\to0$. Difference-of-powers identities give
$$
\begin{aligned}
|\mathbb E\mathcal C(v_n,w_n,z_n)-\mathbb E\mathcal C(v,w,z)|
 &\le(12W^2+24V^3)e_v+24VW e_w+2Ze_z,\\
|\mathbb E v_n^4-\mathbb E v^4|&\le4V^3e_v,
\end{aligned}
$$
where $\mathcal C(v,w,z)=z^2-12vw^2+6v^4$. These constants depend only
on the common strip and the coefficient mass bound, and in particular are
independent of the plateau width. After division by $c^2$, this is (4.1).

Only spatial moments pass to the limit. The auxiliary quantities $s_n$ and
$j_n$ have already disappeared after integration, so the transfer needs
neither their convergence nor a uniform bound on their tail constants.
[OpenAI (2026a), Lemmas 2.1–2.3][fs-reg];
[Proposition 5.1][fs-curv].

<a id="fixed-time-jump-extension"></a>

**The spatial inequality at a coefficient jump.** Proposition 4.1 and the
sampled formal interface assume continuity at $t$, and the spatial-moment
argument permits a little more. Fix $c=\gamma(t)>0$ and replace $\gamma$ by $c$ on
$[t-\delta,t+\delta)$, leaving it unchanged elsewhere. Monotonicity
preserves both joins and right continuity. For fixed
$0<\delta_0<\min(t,1-t)$ and $0<\delta\le\delta_0$,
$$
\|\gamma_\delta-\gamma\|_1
\le2\delta\,\gamma(t+\delta_0)\longrightarrow0.
$$
Thus local boundedness suffices for the same strip stability and coupled
expectation limit. Each replacement has a positive plateau at $t$, so
(4.1) holds at a jump too, with the stated right-continuous value $c$.
This extension concerns spatial moments only. It supplies no time
derivative $q'''(t)$ or $\gamma'(t)$ at a jump, and the formal declaration
keeps every one of its hypotheses. Neither application below uses the
extension.

**Corollary 4.2 (positive plateaus).** Suppose now that $\gamma=c>0$
on an open interval $I\subset(0,1)$ and $t\in I$. Then (4.2) gives
$q'''(t)=\mathcal C_\gamma(t)$. To establish strictness for this
limiting solution, choose $b\in I$ later than $t$ and put
$\psi=\Phi_\gamma(b,\cdot)$. This is smooth, convex, even and a
bounded distance from $|x|$, so $\psi'$ cannot be constant. The
Cole–Hopf formula gives

$$
\Phi_{xx}(t,x)
=\mathbb E^{\rm tilt}_{t,x}\psi''
  +c\operatorname{Var}^{\rm tilt}_{t,x}(\psi'),
$$
$$
\nu_{t,x}(dy)=
\frac{e^{c\psi(y)}p_{b-t}(y-x)}
 {\int_{\mathbb R}e^{c\psi(z)}p_{b-t}(z-x)\,dz}\,dy.
$$

The tilted density is positive everywhere. Hence the tilted variance of
the nonconstant continuous function $\psi'$ is strictly positive, while the
tilted mean of $\psi''$ is nonnegative. Thus $v(t,x)>0$ for every $x$, and
$\mathbb E v(t,X_t)^4>0$. Combining the limiting inequality with
the differentiated identity proves
$q'''(t)\ge(1000c^2)^{-1}\mathbb E v(t,X_t)^4>0$.
In particular, $q'$ is strictly convex on $I$.
[OpenAI (2026a), Lemma 2.3][fs-reg]; [Proposition 5.1][fs-curv].

<a id="zero-plateau"></a>

### 4.6 The zero interval as a separate case

When $c=0$, the heat equation gives
$q''(t)=\mathbb E\Phi_{xxx}(t,X_t)^2$ without division by $c$.
Start the drift-free evolution at an earlier point of $I$. Its
Gaussian transition kernel gives the law of $X_t$ a positive
density everywhere. If the expectation were zero, continuity would
force $\Phi_{xxx}(t,\cdot)\equiv0$. Then $\Phi(t,\cdot)$ would
be an even quadratic polynomial. Its one-Lipschitz bound forces that
polynomial to be constant, contradicting its bounded distance from
$|x|$. Therefore $q''(t)>0$, giving strict increase of $q'$.
[OpenAI (2026a), Lemma 2.3][fs-reg].

The analytic certificate is thus a proved input to the gap argument in
this source, under the stated model and regularity assumptions. An
abstract polynomial or flux lemma assumes those analytic hypotheses, and
the finite-step comparisons and the transfer argument discharge them.
Section 13 separates inspection of the corresponding Lean sources, which
shows that construction, from compilation, scoped dependency extraction and
Comparator replay, and reports the observed outcome of each.
<!-- /passage:fixed-time-coercivity -->

<a id="gap-elimination"></a>

## 5. Excluding terminal, interior and initial gaps

<!-- passage:gaps -->
**Theorem 5.1 (full support).** In the pure zero-field SK model with
$\xi(t)=t^2/2$, let $\gamma\in\mathcal U$ minimize $\mathcal P$. Then
$\operatorname{supp}\mu_\gamma=[0,1)$ and $\gamma$ is strictly increasing.
For every associated diffusion started at zero,
$\mathbb E u(t,X_t)^2=t$ and $\mathbb E a(t,X_t)^2=1$ for
$0\le t<1$.

We keep the same minimiser, diffusion and moment function $q$ throughout.
Section 3 has established $G\ge0$, $G=0$ on the support
$S=\operatorname{supp}\mu_\gamma$, and the contact conditions
$q(s)=s$, $q'(s)\le1$ at support points. At zero, the value identity
comes from symmetry, and the slope bound requires $0\in S$.
Section 4 supplies curvature on constant intervals for every admissible
coefficient. These premises are available before any support gap is
excluded. We first deal with the endpoint at one, and then use the
curvature estimate on the remaining intervals.

For a terminal interval $(a,1)$, including empty support $[0,1)$,
the cusp estimate gives $1-q(t)\gtrsim\sqrt{1-t}$ and hence $G<0$
near one. On a positive plateau with two contact endpoints, strict
convexity puts $q'$ below its endpoint chord, although its integral
must equal the interval length. On a zero plateau the heat argument
makes $q'$ strictly increasing, below the right contact slope.
At an initial gap we use the one-sided contact slope if $0\in S$.
Otherwise the coefficient is zero, and symmetry supplies $q(0)=0$.

<a id="terminal-gap"></a>

**A terminal plateau violates the variational inequality.** Suppose that $\gamma=c$ on $(a,1)$, where $0\le a<1$ and
$c\ge0$ is finite. The endpoint $1$ lies outside the support domain and
carries no contact condition, so an endpoint-chord argument is
unavailable. Instead we prove
$$
1-q(t)\ge C\sqrt{1-t}
\tag{5.1}
$$
for some $C>0$ and every $t$ sufficiently close to one. The proof
separates two facts: the diffusion puts positive probability density
near zero, and the smoothed terminal slope has a definite deficit from
$\pm1$ in a shrinking window around zero.

Fix $s\in(a,1)$ and let $\rho_s$ be the law of $X_s$. Choose
$R<\infty$ such that $p_0=\rho_s([-R,R])>0$. For $c>0$, set
$$
h(t,y)=\mathbb E e^{c|y+\sqrt{1-t}Z|},\qquad
\Phi(t,y)=c^{-1}\log h(t,y),
$$
where $Z$ is standard normal. The drift on this interval is
$c\Phi_x=h_x/h$, and the transition density is
$$
k_{s,t}(x,y)=p_{t-s}(y-x)\frac{h(t,y)}{h(s,x)},\qquad
p_v(z)=(2\pi v)^{-1/2}e^{-z^2/(2v)}.
\tag{5.2}
$$
The heat-semigroup identity normalises this kernel. Equivalently, for
Brownian motion started at $(s,x)$, $h(t,B_t)/h(s,x)$ is the positive
mean-one change of measure which adds the drift $h_x/h$. For $c=0$,
the same kernel formula applies
with $h=1$, and the value function is the ordinary heat evolution of
$|x|$. [OpenAI (2026a), Lemma 6.3, proof of (6.6)][fs-var].

Put $L=1-s$ and $\delta=L/2$. For
$s+\delta\le t<1$, $|x|\le R$ and $|y|\le1$,
$$
p_{t-s}(y-x)\ge
(2\pi L)^{-1/2}e^{-(R+1)^2/(2\delta)}=:k_*>0,
$$
while $h(t,y)\ge1$ and
$h(s,x)\le e^{cR}\mathbb E e^{c\sqrt L|Z|}=:M<\infty$.
Integrating (5.2) over the part of $\rho_s$ in $[-R,R]$ therefore
gives a density $f_t$ of $X_t$ with
$$
f_t(y)\ge p_0k_*/M=:m>0
\quad(s+\delta\le t<1,\ |y|\le1).
\tag{5.3}
$$
At time $s$ we used only the positive mass of a compact set, with no
density assumption on $\rho_s$.

Now write $\varepsilon=1-t$ and restrict to
$|y|\le\sqrt\varepsilon$. Differentiating the terminal heat formula
gives, with the unweighted interpretation for $c=0$,
$$
u(t,y)=
\frac{\mathbb E[\operatorname{sgn}(y+\sqrt\varepsilon Z)
                     e^{c|y+\sqrt\varepsilon Z|}]}
     {\mathbb E e^{c|y+\sqrt\varepsilon Z|}}.
$$
The denominator is at most
$D=e^c\mathbb E e^{c|Z|}<\infty$. Each sign has unnormalised
mass at least $p_*:=\mathbb P(Z\ge1)>0$, since the threshold
$|y/\sqrt\varepsilon|$ is at most one, and the exponential weight
is at least one. Hence each sign has tilted probability at least
$\theta=p_*/D\in(0,1/2)$. It follows that
$$
|u(t,y)|\le1-2\theta,\qquad
1-u(t,y)^2\ge4\theta(1-\theta)=:\kappa>0.
$$
Multiplying this uniform deficit by the density lower bound (5.3)
and the window width $2\sqrt\varepsilon$ yields
$$
1-q(t)=\int_{\mathbb R}(1-u(t,y)^2)f_t(y)\,dy
\ge 2m\kappa\sqrt{1-t}.
$$
This proves (5.1) with $C=2m\kappa$.
[OpenAI (2026a), Lemma 6.3, density and terminal-window estimates][fs-var].

We can now read off the contradiction. For $s'$ sufficiently
close to one, the estimate holds at every $t\in[s',1)$, so
$$
\begin{aligned}
G(s')&=\int_{s'}^1(q(t)-t)\,dt\\
&\le\frac{(1-s')^2}{2}
       -\frac{2C}{3}(1-s')^{3/2}<0.
\end{aligned}
$$
The final inequality holds once $\sqrt{1-s'}<4C/3$. It contradicts
the nonnegative cost of adding an atom at $s'$. This proof covers
$c=0$ as well as $c>0$. In particular, empty support is impossible:
then $\mu_\gamma=0$ and $\gamma=0$ on all of $[0,1)$.
The terminal-plateau conclusion appears already in
[Auffinger, Chen and Zeng (2020), Remark 7][acz], and the estimates above
supply the quantitative argument needed here.

This excludes the terminal gap. The support proof resumes with
[the remaining finite components](#finite-gap-cases). The next argument
gives a shorter route to the same exclusion, without its quantitative bound.

<a id="qualitative-terminal-gap"></a>

**A shorter qualitative exclusion.** The curvature input of §4 also
excludes a terminal support gap, with no estimate of its boundary layer.
One elementary calculus step suffices. Suppose
$q\in C^1([r,1))$, $q(r)=r$, $q'(r)\le1$, $q\le1$, and $q'$
is strictly convex on $(r,1)$. If $q(t)\ge t$ at some $r<t<1$,
strict integration below the finite endpoint chord gives
$$
t-r\le\int_r^t q'(v)\,dv
 <\frac{t-r}{2}\bigl(q'(r)+q'(t)\bigr),
$$
so $q'(t)>1$. For $t<s<1$, convexity then gives
$$
q'(s)\ge q'(t)+\frac{s-t}{t-r}\bigl(q'(t)-q'(r)\bigr)
\ge q'(t)>1.
$$
Consequently $q(s)\ge t+(s-t)q'(t)>1$ for $s$ sufficiently
close to one, a contradiction. Thus $q(t)<t$ throughout $(r,1)$.
All chord endpoints lie below one, so no derivative at one is used.

<a id="empty-support-chord"></a>

The same inputs close the empty-support case. Empty Stieltjes support
forces $\gamma=0$ everywhere. The
[zero-interval argument](#zero-plateau) makes $q$ strictly convex on
$[0,1)$, whereas the positive-plateau argument above makes $q'$ strictly
convex. Here $q(0)=0$ and $0\le q\le1$. For $0<t<s<1$, convexity gives
$$
q(t)\le\frac ts q(s)\le\frac ts.
$$
Letting $s\uparrow1$ in this bound gives $q(t)\le t$. Now fix
$t<u<1$. Strict convexity on the finite interval $[0,u]$ gives
$$
q(t)<\frac tu q(u)\le t.
$$
Thus the bounded continuous integrand $q(t)-t$ is strictly negative on
$(0,1)$, so $G(0)<0$, contrary to the allowed addition of mass at zero.
The argument assumes no value or derivative at one. The
[explicit zero-coefficient calculation](#zero-coefficient-example)
also gives $D_+\mathcal P(0)[1]=1/4-1/\pi$, and the qualitative
argument needs only its sign.

A nonempty support with a terminal complementary component has a last
point $r<1$. Its plateau
has positive height, since a neighbourhood of $r$ has positive measure.
The contact values and slope at $r$, including the one-sided case $r=0$,
and positive-plateau curvature meet the calculus hypotheses. Since
$0\le q\le1$, integration gives $G(r)<0$, contradicting atom addition.
This follows from the inputs already stated. We keep the quantitative
proof above for its stronger estimate (5.1).

<a id="finite-gap-cases"></a>

**The remaining gaps have a finite right contact point.**

The support is relatively closed in $[0,1)$. Its complement is a union
of relative open intervals, and the Stieltjes measure vanishes on each
such interval. For $x<y$ in one interval,
$$
\gamma(y)-\gamma(x)=\mu_\gamma((x,y])=0,
$$
so $\gamma$ is constant there. The terminal argument excludes a component
ending at one. Every remaining component therefore ends at a point
$b<1$ of $S$.

<a id="bounded-gaps"></a>

First consider a gap $(a,b)$ with both endpoints in $S$, allowing
$a=0$. Compact-strip regularity gives continuity of $q'$ on $[a,b]$,
even if $\gamma$ jumps at an endpoint. Contact gives
$$
q(a)=a,\qquad q(b)=b,\qquad q'(a),q'(b)\le1.
\tag{5.4}
$$
If the constant coefficient is $c>0$, Section 4 makes $q'$ strictly
convex in the gap. Continuity extends the chord comparison to its
endpoints, giving
$$
q'(t)<\frac{b-t}{b-a}q'(a)+\frac{t-a}{b-a}q'(b)\le1,
\qquad a<t<b.
\tag{5.5}
$$
Integrating contradicts $q(b)-q(a)=b-a$. Equivalently, the mean value
theorem applied to (5.4) requires an interior point with $q'=1$,
whereas (5.5) excludes every such point. The endpoint derivatives need
only be at most one.

<a id="endpoint-chord-area"></a>

With $h=q'$ and $C_{a,b}$ its endpoint chord, the contradiction is the exact area identity
$$
(b-a)-\int_a^b h
=(b-a)\left(1-\frac{h(a)+h(b)}2\right)
 +\int_a^b(C_{a,b}-h)>0.
$$
The first term is nonnegative, and strict convexity makes the second
positive.
Only continuity of $q'$ is needed at the ends. For example, a jump
$\Delta\gamma$ at $s$ gives the one-sided identity
$$
q''(s+)-q''(s-)=-2\Delta\gamma\,\mathbb E a(s,X_s)^3.
$$
It follows by applying Itô's formula to $a^2$ and using continuity of
its derivative moments on a compact strip. Taking $\gamma=0$ before
$s=1/2$ and $\gamma=1/2$ afterward makes this jump strictly negative,
because the terminal Cole–Hopf solution has positive Hessian. Thus
continuity of $q'$ gives no classical higher time derivatives across
coefficient jumps. [OpenAI (2026a), proof of Corollary 6.4][fs-density].

If $c=0$, the separate heat argument in Section 4 makes $q'$ strictly
increasing. Then $q'(t)<q'(b)\le1$ for $a<t<b$, and the same
contradiction follows. The left slope bound is unnecessary in this case.
[OpenAI (2026a), proof of Theorem 1.1 after Lemma 6.3][fs-var].

<a id="initial-gaps"></a>

There are two distinct possibilities at the left endpoint. If
$(0,b)$ is a gap and $0\in S$, the preceding proof applies using
the one-sided contact slope bound at zero. Because $(0,b)$ contains
no mass, the support condition at zero forces an atom there and hence
$c=\gamma(0)>0$. At this stage the atom has not yet been excluded.

If instead the complementary component is $[0,b)$ with $0\notin S$,
there is no mass anywhere in $[0,b)$. Thus $\gamma=0$ there.
The identity $q(0)=0$ comes from zero-field symmetry, and the
right endpoint gives $q(b)=b$, $q'(b)\le1$. Here $q''(t)>0$ for
$0<t<b$, and at zero evenness gives $\Phi_{xxx}(0,0)=0$ and hence
$q''(0)=0$. Strict increase on the open interval still forces $q'<1$
throughout $(0,b)$, an impossibility.

**Full support and the moment identities.**

The cases are exhaustive, so $S=[0,1)$. Only now can contact be
applied at every $0<t<1$, giving $q(t)=t$ throughout the domain.
The value at zero was already known. Differentiating and using the
gradient-martingale identity gives
$$
\mathbb E a(t,X_t)^2=q'(t)=1,\qquad 0\le t<1,
$$
with the right derivative at zero.

<a id="strict-increase"></a>

Finally, every nonempty open subinterval $(a,b)$ receives positive mass. Therefore
$$
\gamma(b)-\gamma(a)=\mu_\gamma((a,b])>0
\qquad(0\le a<b<1).
$$
The proof establishes full support, strict increase and the two
moment identities. Smooth absolute continuity, including the later
conclusion $\gamma(0)=0$, uses the source's further argument from
saturated consistency, and neither contact nor the exclusion of an
initial gap depends on it. [OpenAI (2026a), conclusion of the
proof of Theorem 1.1][fs-var].

A weaker analytic condition suffices for the bounded-gap contradiction.
If $h\in C([a,b])\cap C^2((a,b))$ has the displayed endpoint and
integral data, then $h'=0\Rightarrow h''>0$ is already impossible:
either $h\le1$ everywhere, forcing $h\equiv1$ by its integral, or it
attains an interior maximum above one, where $h'=0$ and $h''\le0$.
This is an elementary deduction from these premises. Chen's conditional
crossing theorem, with its own hypotheses, has this form. Section 4
proves the stronger $q'''>0$ throughout a positive plateau, for any
admissible coefficient. See [Chen, arXiv:2607.18032v1, Proposition 4.2 and the proof
of Proposition 4.3][chen26]. Neither statement supplies terminal contact
at the excluded endpoint one.
<!-- /passage:gaps -->

<a id="scalar-value-identities"></a>

<a id="terminal-value"></a>

## 6. Recovering the value from a martingale

<!-- passage:value -->
**Proposition 6.1 (terminal value).** For the minimizer and diffusion of
Theorem 5.1 there is a spin-valued terminal variable $U_1$ such that
$u(t,X_t)=\mathbb E[U_1\mid\mathcal F_t]$ and
$$
P_*=\mathbb E[U_1B_1]=\int_0^1\mathbb E a(t,X_t)\,dt.
\tag{6.1}
$$
Its finite-cutoff tail lies in $[0,1-T]$. Besides the support theorem,
the proposition uses the attained Parisi formula.

The bounded martingale $u(t,X_t)$ converges almost surely and in $L^2$
to $U_1$. Its second moment tends to one because $q(t)=t$.
Since $|U_1|\le1$, $\mathbb E U_1^2=1$ forces $U_1^2=1$ almost surely.
Uniform integrability identifies $u(t,X_t)=\mathbb E[U_1\mid\mathcal F_t]$,
which is the conditional-expectation identity used below.

The drift bound $|\gamma u|\le\gamma\in L^1$ gives convergence of $X_t$
to $X_1$, and uniform terminal convergence of $\Phi$ to $|x|$, together
with convexity, gives $u(t,x_t)\to\operatorname{sgn}(x)$ when
$x_t\to x\ne0$. Thus $U_1X_1=|X_1|$, including when $X_1=0$, so we
never need the probability of that event.

Itô's formula gives
$$
\mathbb E\Phi(t,X_t)=\Phi(0,0)+\frac12\int_0^t\gamma(s)s\,ds.
$$
The terminal limit first gives
$\Phi(0,0)=\mathbb E|X_1|-\tfrac12\int_0^1s\gamma(s)ds$.
Subtracting the second half from the definition of the functional yields
$\mathcal P(\gamma)=\mathbb E|X_1|-\int_0^1s\gamma(s)\,ds$.
The ground-state Parisi formula identifies this with $P_*$.
The diffusion equation, conditional expectation and Fubini give
$$
\mathbb E[U_1X_1]=\mathbb E[U_1B_1]
 +\int_0^1\gamma(s)\mathbb E[u(s,X_s)^2]\,ds.
$$
Bounded $u,U_1$ and integrable $\gamma$ justify these terms.
Combining the identities and using stochastic integral covariance gives
$$
P_*=\mathbb E|X_1|-\int_0^1t\gamma(t)\,dt
 =\mathbb E[U_1B_1]=\int_0^1\mathbb E a(t,X_t)\,dt.
$$
The last integral is absolutely bounded by one because
$\mathbb E|a|\le(\mathbb E a^2)^{1/2}=1$.
Full support supplies $\mathbb E a^2=1$, whereas the value uses
$\int\mathbb E a$, so substituting the second moment into the value
identity would change the assertion. [OpenAI (2026a), §7][fs-cons].

The prediction error and the energy tail have different sizes:
$$
\|U_1-u(T,X_T)\|_2^2=1-T,
$$
$$
P_*-\int_0^T\mathbb E a(t,X_t)\,dt
=\mathbb E[(U_1-u(T,X_T))(B_1-B_T)]\in[0,1-T].
$$
The first identity is orthogonal projection onto $\mathcal F_T$.
Martingale orthogonality removes the past Brownian component in the
second. Cauchy–Schwarz bounds the product by $1-T$, and convexity
makes it nonnegative, so no uniform terminal curvature bound is needed.

Sellke gave a direct antecedent of this stochastic value calculation in
[Sellke, arXiv:2105.03506v4, Definition 1.3 and Lemma 3.11](https://arxiv.org/abs/2105.03506v4).
There, “optimizable” means the contact identity $\mathbb E u(t,X_t)^2=t$
on $[q,1)$, with $q$ the infimum of the support of the coefficient
function. Lemma 3.11 derives the finite-cutoff value identity by applying
Itô's formula to $\Phi$ and $xu$, before taking the terminal limit.
This is the same value-recovery mechanism used here, specialized to
zero field and $\xi''=1$.

A cutoff must retain the full Parisi penalty. Set
$\Psi(T,x)=\Phi(T,x)-xu(T,x)$ and
$I(T)=\int_0^T\mathbb E a(t,X_t)\,dt$. Subtracting the complete penalty
from the finite-time Itô identity gives
$$
\mathcal P(\gamma)-I(T)
 =\mathbb E\Psi(T,X_T)
  +\frac12\int_0^T\gamma(t)(q(t)-t)\,dt
  -\frac12\int_T^1t\gamma(t)\,dt.
$$
Under contact the middle term vanishes. The final term is absent from
the displayed equality in [El Alaoui, Montanari and Sellke,
arXiv:2001.00904v1, §6.2, pp. 32–33][eams20]. Its sign is nonpositive,
so the upper bound by $\mathbb E\Psi(T,X_T)$ used there remains valid,
and its magnitude tends to zero by integrability. This finite-cutoff
correction is consistent with the exact terminal identity above.
<!-- /passage:value -->

<!-- passage:coefficient-budget -->
<a id="coefficient-budget"></a>

**A coefficient budget from the first curvature moment.** Full support
also bounds the minimizing coefficient. Put $m(t)=\mathbb E a(t,X_t)$.
The compact-strip identity
$$
da(t,X_t)=-\gamma(t)a(t,X_t)^2\,dt+b(t,X_t)\,dB_t
$$
holds before any smoothness conclusion for $\gamma$, and the bounded strip
derivatives make its stochastic integral a martingale.
[OpenAI (2026a), proof of Corollary 6.4][fs-density].
At zero, $X_0=0$, $a(0,0)\ge0$ and $\mathbb E a(0,X_0)^2=1$
give $a(0,0)=1$. Taking expectations and using
$\mathbb E a(t,X_t)^2=1$ therefore yields
$$
m(t)=1-\int_0^t\gamma(s)\,ds,\qquad 0\le t<1.
$$
Convexity gives $m(t)\ge0$. Letting $t\uparrow1$ and using
monotonicity of $\gamma$ on every tail gives
$$
\int_0^1\gamma(s)\,ds\le1,\qquad
0\le\gamma(t)\le\frac{m(t)}{1-t}\le\frac1{1-t}.
$$
Indeed, $(s-t)\gamma(t)\le\int_t^s\gamma\le m(t)$ for
$t<s<1$. Integrating the first-moment identity also gives
$$
I(T)=T-\int_0^T(T-s)\gamma(s)\,ds,\qquad
P_*=1-\int_0^1(1-s)\gamma(s)\,ds.
$$
The second equality uses (6.1) and so depends on its ground-state
identification. Both follow directly in the normalization $\xi''=1$.
The budget alone leaves the terminal first moment undetermined.

<a id="coefficient-total-mass"></a>

**Proposition 6.2 (terminal normalization in zero field).** For the
minimizer and diffusion of Theorem 5.1,
$$
\lim_{t\uparrow1}m(t)=0,\qquad
\int_0^1\gamma(s)\,ds=1,\qquad
m(t)=\int_t^1\gamma(s)\,ds,
$$
and consequently
$$
P_*=\int_0^1s\gamma(s)\,ds.
$$
Thus $\gamma(s)\,ds$ has total mass one. This says nothing about the
total mass of the Stieltjes measure $d\gamma$. The proposition is an
analytic consequence outside the selected formal endpoints.

**Proof.** The backward comparison in §4.3 gives
$\Phi_{xxx}(t,x)\le0$ for $x>0$ in every positive finite-step
approximation. Compact-strip convergence of the third spatial derivative
passes this inequality to $\gamma$. Together with evenness and convexity,
it says that $a(t,x)$ is nonnegative, even and nonincreasing in $|x|$.
These inputs are [OpenAI (2026a), Lemma 3.2][fs-signs] and
[Lemma 2.1][fs-reg], and neither requires differentiability of $\gamma$.
Uniform terminal convergence to $|x|$ and convex difference quotients
give $u(t,x)\to1$ at every fixed $x>0$. For each $\delta>0$,
$$
0\le a(t,\delta)
\le\frac2\delta\int_{\delta/2}^{\delta}a(t,x)\,dx
=\frac2\delta\bigl(u(t,\delta)-u(t,\delta/2)\bigr)
\longrightarrow0.
$$
The spatial monotonicity is essential here, since convergence of convex
functions alone would not imply convergence of their second derivatives.

It remains to control the probability of the shrinking region where
curvature can be large. Oddness and monotonicity give $u(t,x)\ge0$
for $x\ge0$, so the drift points away from zero. Tanaka's formula for
$R_t=|X_t|$ gives
$$
R_t=W_t+V_t+L_t,\qquad
W_t=\int_0^t\operatorname{sgn}(X_s)\,dB_s,\qquad
V_t=\int_0^t\gamma(s)u(s,R_s)\,ds.
$$
Here $L$ is the local time of $X$ at zero.
The occupation-time formula and $d\langle X\rangle_s=ds$ imply
$\int_0^t\mathbf1_{\{X_s=0\}}\,ds=0$. Thus $\langle W\rangle_t=t$,
and the continuous martingale $W$ is Brownian motion. The process $V$
is nondecreasing and starts at zero, and $L$ increases only when $R=0$.
The reflection formula therefore gives, pathwise,
$$
R_t=W_t+V_t-\min\left(0,\inf_{s\le t}(W_s+V_s)\right)
\ge W_t-\min\left(0,\inf_{s\le t}W_s\right)=:R_t^0.
$$
Indeed $0\le V_s\le V_t$ bounds the subtracted minimum by
$\min(0,\inf_{s\le t}W_s)+V_t$. No independence of $V$ and $W$
is needed. Reflected Brownian motion $R_t^0$ has the one-time law
$|B_t|$, so for $t>0$,
$$
\Pr(|X_t|\le\delta)
\le\Pr(|B_t|\le\delta)
\le\sqrt{\frac2\pi}\frac{\delta}{\sqrt t}.
$$
Splitting at $|X_t|=\delta$, using spatial monotonicity outside and
Cauchy–Schwarz with $\mathbb E a(t,X_t)^2=1$ inside, yields
$$
0\le m(t)\le a(t,\delta)
              +\sqrt{\Pr(|X_t|\le\delta)}.
$$
First let $t\uparrow1$ and then $\delta\downarrow0$. This proves
$m(1-)=0$. The compact-strip identity for $m$ now gives total mass one
and the tail formula. Tonelli and (6.1) give the stated expression for
$P_*$. In particular,
$$
0\le P_*-I(T)=\int_T^1(s-T)\gamma(s)\,ds
\le(1-T)m(T)=o(1-T).
$$
No numerical rate for $m(T)\to0$ or for the required spatial derivative
bounds follows from this proof.

The endpoint distinction also matters when comparing editions.
[Sellke, arXiv:2105.03506v4, Lemma 2.9, equation (2.7)](https://arxiv.org/abs/2105.03506v4)
states the expected-Hessian tail equality. In
[version 6, Lemma 2.9, equation (2.8), and Remark 2.10](https://arxiv.org/html/2105.03506v6#S2.SS3),
the general statement is an inequality, and the remark obtains equality
indirectly for even mixtures. The specialized zero-field proof above
supplies its own endpoint argument from the shape and contact identities.
It never substitutes $t=1$ into a compact-strip formula, and it avoids the
older equality in its full stated generality.
<!-- /passage:coefficient-budget -->

<a id="finite-gaussian-approximation"></a>

<a id="finite-gaussian-correlations"></a>

## 7. A finite Gaussian spin and its adjacent correlations

<!-- passage:gaussian -->
**Proposition 7.1 (finite Gaussian approximation).** For a fixed minimizer
and its diffusion, the construction below has positive active normalizers
for all sufficiently fine meshes at every fixed $0<T<1$. Its adjacent
correlation sum converges to $\int_0^T\mathbb E a(t,X_t)\,dt$, and this
integral tends to $P_*$ as $T\uparrow1$. The limits are taken in that order.

We keep the minimizer and its diffusion from §6 and construct a finite
Gaussian sign $F_N$ with two correlation vectors whose *adjacent*
pairing
$$
S_{T,N}=\sum_{j=0}^N A_jB_{j+1}
\longrightarrow\int_0^T\mathbb E a(t,X_t)\,dt.
$$
The shift by one is essential for the rooted energy identity in §8.1, and
[§10.4](#one-step-example) computes a case where the unshifted pairing has
a different value.

We first fix $0<T<1$, then a mesh $N\ge1$,
$\Delta=T/N$, $t_j=j\Delta$.
On independent standard normals $Z_1,\ldots,Z_{N+1}$, set
$$
Y_0=0,\quad Y_j=Y_{j-1}+\sqrt\Delta Z_j+
 \Delta\gamma(t_{j-1})u(t_{j-1},Y_{j-1}),
$$
$$
c_j=(\mathbb E a(t_{j-1},Y_{j-1})^2)^{-1/2},\qquad
g_j=c_ja(t_{j-1},Y_{j-1})Z_j.
$$
For sufficiently large $N$, all $c_j$, $1\le j\le N$, are finite
and positive. Couple the increments to Brownian motion. On the fixed
strip, the gradient and curvature are bounded and Lipschitz. Monotonicity
gives, for the left-grid projection $\pi_N$,
$$
\int_0^T|\gamma(s)-\gamma(\pi_N(s))|\,ds
 \le\Delta(\gamma(T)-\gamma(0)).
$$
Drift error estimates and discrete Gronwall yield
$\mathbb E\max_{j\le N}|Y_j-X_{t_j}|^2\to0$, hence
$\max_j|c_j-1|\to0$. The $g_j$ are orthonormal in $L^2$:
they have unit variance and conditional mean zero given their past.
Their predictable multipliers may depend on earlier coordinates, so
orthonormality is weaker than independence.

With $F_{\rm G}$ the standard Gaussian distribution function, define
$$
F_N=\operatorname{sgn}(u(T,Y_N)+2F_{\rm G}(Z_{N+1})-1).
$$
$$
A_j=\mathbb E[F_Ng_j],\quad B_i=\mathbb E[F_NZ_i],\quad A_0=0.
$$
The extra coordinate is independent and its distribution function is uniform.
For any $v\in[-1,1]$, the random sign of $v+2U-1$ has mean $v$.
Therefore
$\mathbb E[F_N\mid Z_1,\ldots,Z_N]=u(T,Y_N)$ exactly.
A deterministic sign of the gradient would not preserve this conditional
mean. Rounding preserves correlations with functions of the first $N$
coordinates, which is exactly what the two coefficient vectors need.
It does not make the whole rounded spin close in $L^2$ to that conditional
mean. Indeed, with $m_N=u(T,Y_N)$,
$$
 \mathbb E[(F_N-m_N)^2]=1-\mathbb E m_N^2\longrightarrow1-T.
$$
Here $m_N\to u(T,X_T)$ in $L^2$ at fixed $T$ and full support gives
$\mathbb E u(T,X_T)^2=T$. The independent rounding variance disappears
only in the later endpoint limit.

The approximation
$$
M_N=\sum_{j=1}^N\sqrt\Delta\,a(t_{j-1},Y_{j-1})Z_j
 =u(T,Y_N)+o_{L^2}(1)
$$
gives a common residual for both coefficient families. In
$\mathbb E[M_NZ_i]$, a summand with $j<i$ is independent of centered
$Z_i$, and one with $j>i$ vanishes on conditioning on the first $j-1$
coordinates. The $j=i$ term is
$\sqrt\Delta\,\mathbb E a(t_{i-1},Y_{i-1})$.
Also $\mathbb E[M_Ng_j]=\sqrt\Delta/c_j$ by orthonormality.

Since the union of the two families need not be orthogonal, we apply
Bessel's inequality separately to the orthonormal family $(g_j)$ and to
the family $(Z_i)$. This gives the vector estimates
$$
(A_j)_{j=1}^N=(\sqrt\Delta/c_j)_{j=1}^N+o_{\ell^2}(1),
$$
$$
(B_i)_{i=1}^N=
(\sqrt\Delta\,\mathbb E a(t_{i-1},Y_{i-1}))_{i=1}^N+o_{\ell^2}(1).
$$
A sum of growing length needs these vector estimates, since coordinatewise
errors would not control it. Put
$$
e_N=\|u(T,Y_N)-M_N\|_2,\quad
\alpha_j=\sqrt\Delta/c_j,\quad
\beta_j=\sqrt\Delta\,\mathbb E a(t_{j-1},Y_{j-1})
\quad(1\le j\le N).
$$
The fixed-$T$ Euler argument gives $e_N\to0$. Bessel's inequality gives
$\|A_{1:N}-\alpha\|_2,\|B_{1:N}-\beta\|_2\le e_N$.
Also $\|B_{1:N}\|_2\le1$ and $\|\alpha\|_2\le C_T$ for sufficiently
large $N$, with $C_T$ independent of the mesh. Cauchy–Schwarz gives
$$
\begin{aligned}
\left|\sum_{j=1}^{N-1}(A_jB_{j+1}-\alpha_j\beta_{j+1})\right|
&\le\|A-\alpha\|_2\|B_{2:N}\|_2
 +\|\alpha_{1:N-1}\|_2\|B-\beta\|_2\\
&\le(1+C_T)e_N.
\end{aligned}
$$
The unindexed differences in this inequality use coordinates $1,\ldots,N$,
and restricting to the indicated subvector only decreases their norms.
The rounding coordinate contributes at most
$|A_NB_{N+1}|\le e_N+\sqrt\Delta/c_N$. Thus the
dimension-free estimate is
$$
\left|\sum_{j=0}^NA_jB_{j+1}
 -\Delta\sum_{j=1}^{N-1}\frac{\mathbb E a(t_j,Y_j)}{c_j}\right|
\le(2+C_T)e_N+\frac{\sqrt\Delta}{c_N}\longrightarrow0.
$$
The index shift selects the Hessian at $t_j$, and $Z_{N+1}$ serves the
rounding, with no Euler increment of its own. This estimate is a
fixed-strip scalar correspondence, and the circuit energy identity comes
only in §8.1. Hence
$$
\sum_{j=0}^NA_jB_{j+1}
 =\Delta\sum_{j=1}^{N-1}\frac{\mathbb E a(t_j,Y_j)}{c_j}+o(1)
 \longrightarrow\int_0^T\mathbb E a(t,X_t)\,dt.
$$

<a id="unshifted-scalar-comparison"></a>

The same vector estimates also give an unshifted sum with this limit:
$$
\left|\sum_{j=1}^N A_jB_j-\sum_{j=1}^N\alpha_j\beta_j\right|
\le(1+C_T)e_N\longrightarrow0,
$$
$$
\sum_{j=1}^N A_jB_j
=\Delta\sum_{j=1}^N
  \frac{\mathbb E a(t_{j-1},Y_{j-1})}{c_j}+o(1)
\longrightarrow\int_0^T\mathbb E a(t,X_t)\,dt.
$$
The latter is a left Riemann sum, since $c_j\to1$ uniformly and the Euler
coupling replaces $Y_{j-1}$ by $X_{t_{j-1}}$. Equal continuum limits
therefore leave open which finite expression the circuit needs. The rooted
energy calculation in §8.1 selects the adjacent pairing, and the two
finite pairings can differ, as [§10.4](#one-step-example) shows.

Only afterward let $T\uparrow1$, obtaining $P_*$.
The mesh count $N$ is unrelated to the spin-system size $n$.
As in [OpenAI (2026a), Proposition 7.2][fs-cons], these coefficients
describe a finite Gaussian function. The circuit on finite SK disorder is
built only in §8.

The QAOA scalar input also uses the tail estimate
$0\le P_*-\int_0^T\mathbb E a(t,X_t)\,dt\le1-T$, from nonnegative
curvature and $\mathbb E a^2=1$. This estimate and the fixed-strip
regularity facts must be carried beside the coefficient limit, since some
of them lie outside the selected finite-Gaussian statement. They enter
the construction in [OpenAI (2026b), §3][q-classical].

Thus, for every accuracy, the same chosen minimizer supplies a cutoff $T$
and then one finite function $F_N$ with $|S_{T,N}-P_*|$ below that accuracy.
This is the finite scalar input. Turning it into spins for a finite SK
matrix uses the separate AMP theorem in §9.4, and turning it into a common
QAOA word uses §§8.1–8.6. Both operations lie outside the Gaussian limit.
<!-- /passage:gaussian -->

<!-- passage:gaussian-error-budget -->

<a id="scalar-error-budget"></a>

### 7.1 Error estimates at a fixed cutoff

We fix the minimizer first and then $T<1$. The constants in the bounds
below depend on that minimizer and strip. They give an $N^{-1/2}$ rate
there, with no uniformity as $T\uparrow1$. The derivative bounds are
finite, but no numerical values for them are available as inputs to an
algorithm that chooses a mesh.

The calculation is a deduction from the source's strip estimates and goes
beyond the qualitative selected formal statement. Write
$S_{T,N}=\sum_{j=0}^N A_jB_{j+1}$ and
$f(t)=\mathbb E a(t,X_t)$, so $0\le f(t)\le1$. On the fixed strip let
$K_T=\sup_{t\le T,x}|a(t,x)|$ and choose a spatial Lipschitz constant
$L_T$ for $a$. Define
$$
 \delta_N=\max_{0\le j\le N}\|Y_j-X_{t_j}\|_2,\qquad
 d_N=\max_{1\le j\le N}
 \left|\mathbb E a(t_{j-1},Y_{j-1})^2-1\right|.
$$
These errors have a rate at each fixed strip, even if $\gamma$ has jumps.
Put $\Gamma=\gamma(T)$ and $V=\gamma(T)-\gamma(0)$. Bounded spatial
derivatives and the PDE give temporal Lipschitz constants
$$
J=\tfrac12\|u_{xx}\|_\infty+\Gamma K_T,\qquad
H=\tfrac12\|a_{xx}\|_\infty+\Gamma(K_T^2+L_T)
$$
for $u$ and $a$, respectively, on this strip. No derivative of $\gamma$
is used. Brownian coupling gives
$\|X_s-X_{\pi_N(s)}\|_2\le\sqrt\Delta+\Gamma\Delta$.
Subtract the Euler and diffusion equations and split the drift difference
on $[t_j,t_{j+1}]$ as
$$
\gamma(s)\bigl(u(s,X_s)-u(t_j,Y_j)\bigr)
 +\bigl(\gamma(s)-\gamma(t_j)\bigr)u(t_j,Y_j).
$$
Put $A_T=\int_0^T\gamma(s)\,ds\le1$ and
$w_j=\int_{t_j}^{t_{j+1}}\gamma(s)\,ds$. Taking a pathwise maximum
over grid times, Minkowski and discrete Gronwall give, with
$$
R_N=A_TK_T\sqrt\Delta+
 [A_T(J+\Gamma K_T)+V]\Delta,
$$
the bound
$$
\delta_N\le\left\|\max_{j\le N}|Y_j-X_{t_j}|\right\|_2
\le e^{K_TA_T}R_N\le e^{K_T}R_N.
$$
The recurrence weights the preceding maximal norm on interval $j$ by
$K_Tw_j$, and these weights sum to $K_TA_T$. The term $V\Delta$ is the monotone
quadrature bound for the coefficient. For the
martingale residual, Itô isometry and the same increment estimate give
$$
e_N\le K_T\delta_N+
\sqrt T\{L_T\delta_N+L_T\sqrt\Delta+(H+\Gamma L_T)\Delta\}.
$$
Thus $\delta_N,e_N=O_{\gamma,T}(\sqrt\Delta)$.
These are deductions from the compact-strip bounds and stochastic
identities. The selected formal statement contains no such quantitative
clause.

Since $a$ is bounded and Lipschitz,
$d_N\le2K_TL_T\delta_N\to0$. Whenever $d_N\le1/2$,
$$
 \sqrt{2/3}\le c_j\le\sqrt2,\qquad
 |1/c_j-1|\le d_N,\qquad
 \|\alpha\|_2\le\sqrt{T(1+d_N)}.
$$
These estimates give an explicit sufficient condition for all the required
normalisers to be positive, and they assume nothing about coarse meshes.

The first-moment identity in §6 makes $f=m$ nonincreasing, with
$f(0)=1$. Its right Riemann sum therefore satisfies
$$
0\le I(T)-\Delta\sum_{j=1}^N f(t_j)
\le\Delta\bigl(f(0)-f(T)\bigr).
$$
Removing the last term adds $\Delta f(T)$, so
$0\le I(T)-\Delta\sum_{j=1}^{N-1}f(t_j)\le\Delta$.
The difference between the shifted interior sum and the integral is bounded by
$$
 \left|\Delta\sum_{j=1}^{N-1}
       \frac{\mathbb E a(t_j,Y_j)}{c_j}-\int_0^T f(t)\,dt\right|
 \le TK_Td_N+TL_T\delta_N+\Delta.
$$
The first two terms replace the normalisers and Euler positions, and the
last is the preceding monotone quadrature estimate.
Combining this with the vector and rounding-endpoint estimate gives
$$
 \begin{aligned}
 |S_{T,N}-P_*|
 \le{}&(1-T)m(T)
   +\bigl(2+\sqrt{T(1+d_N)}\bigr)e_N
   +\sqrt{\Delta(1+d_N)}\\
 &+TK_Td_N+TL_T\delta_N+\Delta .
 \end{aligned}
$$
In particular, for each fixed minimizer and $T<1$ there are finite
$C_{\gamma,T}$ and $N_0(\gamma,T)$ such that
$$
|S_{T,N}-P_*|\le(1-T)m(T)+C_{\gamma,T}\sqrt{T/N}
\qquad(N\ge N_0(\gamma,T)).
$$
The constants are not uniform as $T\uparrow1$. The
[coefficient budget](#coefficient-budget) bounds $\int\gamma$ and
$\gamma$ on each fixed strip, but the argument gives no numerical bounds
for the strip derivative constants and no effective mesh threshold.
The scalar tail bound $(1-T)m(T)\le1-T$ uses only monotonicity of
the first moment, and Proposition 6.2 further makes it $o(1-T)$.

All terms after this scalar tail tend to zero at each fixed $T<1$. Given a
scalar
tolerance $\zeta>0$, first choose $T$ with $1-T<\zeta/2$, then choose
$N_0(\gamma,T,\zeta)$ so that their sum is below $\zeta/2$ for every
$N\ge N_0$. This gives the required finite choice of a Gaussian spin.
The formula exposes the errors, but it gives no numerical values of
$K_T,L_T,\delta_N,e_N$ and no effective mesh threshold.
[OpenAI (2026a), Lemma 2.1][fs-reg];
[OpenAI (2026a), Proposition 7.2][fs-cons].

These thresholds also let us choose a diagonal. For prescribed
$T_m\uparrow1$, choose increasing $N_m$ large enough to make the
mesh-dependent part of the fixed-$T_m$ bound less than $1/m$. An arbitrary increasing mesh schedule
does not follow from the iterated limit alone. For example, the scalar
array
$$
 h(T,N)=\min\{1,[N(1-T)]^{-1}\}
$$
tends to zero in $N$ at every fixed $T<1$, but
$h(1-1/N,N)=1$ for every $N\ge2$. The array only illustrates the
quantifier issue and asserts no error law for the SK construction.
<!-- /passage:gaussian-error-budget -->

<a id="qaoa-construction"></a>

## 8. From the scalar formula to one ordinary QAOA word

<!-- passage:qaoa -->
On a single spin, let $Z=\operatorname{diag}(1,-1)$ and
$X=\begin{pmatrix}0&1\\1&0\end{pmatrix}$. The operators $Z_i$ and $X_i$
act on spin $i$ and as the identity on the other spins. Also set
$Y=\begin{pmatrix}0&-i\\i&0\end{pmatrix}$ and
$|\pm\rangle=(|0\rangle\pm|1\rangle)/\sqrt2$.
Then $X|\pm\rangle=\pm|\pm\rangle$ and $Z|+\rangle=|-\rangle$. The cost and mixer are
$$
C_{n,J}=n^{-1/2}\sum_{i<j}J_{ij}Z_iZ_j,\qquad B_n=\sum_iX_i.
$$
The two evolutions act differently:
$$
e^{-i\theta C_{n,J}}|\sigma\rangle
 =e^{-i\theta H_n(\sigma)}|\sigma\rangle,
\qquad
e^{-i\beta B_n}=\prod_{i=1}^n(\cos\beta\,I-i\sin\beta\,X_i).
$$
The cost evolution changes phases while preserving spin-basis probabilities,
and the mixer combines amplitudes of different spin configurations. Cost evolution
alone leaves the initial uniform probabilities unchanged, and a mixer applied
first only multiplies $|+\rangle^{\otimes n}$ by $e^{-in\beta}$.
Alternating them lets the cost-dependent phases change subsequent measurement
probabilities. [Farhi, Goldstone, Gutmann and Zhou (2022), §2,
equations (1)–(6)](https://arxiv.org/abs/1910.08187v4).

A depth-$p$ state, with $p$ counting cost–mixer pairs, is
$$
|\psi_{n,p,J}(\boldsymbol\gamma,\boldsymbol\beta)\rangle
=e^{-i\beta_pB_n}e^{-i\gamma_pC_{n,J}}\cdots
 e^{-i\beta_1B_n}e^{-i\gamma_1C_{n,J}}|+\rangle^{\otimes n}.
$$
Here the real angle vectors $\boldsymbol\gamma,\boldsymbol\beta$ are
deterministic, and $\boldsymbol\gamma$ is separate from the Parisi
coefficient $\gamma(t)$. A *word* means a finite ordered product of cost
and mixer exponentials. Its pattern and angles are fixed independently of
$n,J$, while its cost operator depends on $n,J$. Define
$$
v_p(\boldsymbol\gamma,\boldsymbol\beta)
=\lim_{n\to\infty}\frac1n\mathbb E_J
 \langle\psi_{n,p,J}|C_{n,J}|\psi_{n,p,J}\rangle,
\qquad Q_p=\sup_{\boldsymbol\gamma,\boldsymbol\beta\in\mathbb R^p}v_p.
$$
The fixed-depth, fixed-angle infinite-size limit is the input from
Farhi, Goldstone, Gutmann and Zhou, and the later depth conclusion lies
beyond it; see [Farhi, Goldstone, Gutmann and Zhou (2022)](https://arxiv.org/abs/1910.08187v4).
The source target is $\lim_{p\to\infty}Q_p=P_*$. Its needed lower
bound has the literal quantifiers
$$
\forall\varepsilon>0\quad\exists p\in\mathbb N,
\boldsymbol\gamma,\boldsymbol\beta\in\mathbb R^p:
\quad\lim_{n\to\infty}\frac1n\mathbb E_J
\langle\psi_{n,p,J}|C_{n,J}|\psi_{n,p,J}\rangle
\ge P_*-\varepsilon.
$$
The same finite list of gates and angles serves every $n,J$, and the
conclusion concerns the limiting expected energy. A cost expectation is bounded above by its
maximum eigenvalue, and zero-angle padding makes $Q_p$ nondecreasing.
It therefore suffices to construct one finite word for every accuracy.
Fixed-instance convergence at increasing depth is a different quantifier
order. Farhi, Goldstone and Gutmann's original QAOA argument obtains that
convergence by adiabatic evolution and its Trotter approximation, with depth
allowed to depend on the instance. That argument gives no
size-independent word of the kind required here. [Farhi–Goldstone–Gutmann (2014), equation (10) and
§VI, pp. 10–11](https://arxiv.org/abs/1411.4028v1);
[OpenAI (2026b), Theorem 1.1][q-intro].

The construction first describes local operations on a large tree, where
neighbouring subtrees supply Gaussian inputs. Temporary random labels let
those operations implement the finite classical function of §7. The hard
steps are to implement its aggregations using cost and mixer gates, and
then to replace every use of the labels by ordinary gates while preserving
the energy. Only the resulting ordinary word is transferred to SK.

The choices are nested: fix $\varepsilon$, choose $T<1$, choose a finite
mesh, fix the bounded classical formula and helper probabilities, then
choose the finite gate approximations and remove the seeds. The constants
in a replacement may depend on the surrounding word and on those positive
helper probabilities. [§8.6](#qaoa-assembly-detail) collects the error
allocation before the final $n\to\infty$ limit.
We state the principal circuit inputs as four local lemmas. Their
proofs are in the cited manuscript, and §§8.2–8.5 reconstruct the
mechanisms and show where each hypothesis enters.

**Smooth replacement.** Fix the height, a finite expression and channel
list, and the required derivatives. Uniform recursive weighted bounds,
convergence of operators and adjoints, and strongly commuting signals with
the stated joint exponential bounds permit substitution inside a fixed
surrounding sequence in the exclusion-gate class defined in §8.2. Constants depend on that fixed family. This is the input
used in §8.3 to replace classical aggregates by finite gates; see
[the foundations, smooth topology and gate calculus][q-gate-calculus].

**Helper contraction and synthesis.** Fix a bounded, finite, type-excluded
odd formula, zero on helpers, and positive helper masses. The contraction
converges smoothly at every cut and for the specified responses. For each
$\eta>0$, finite seeded synthesis gives
$$
\mathcal E(v)\ge\mathcal E(m)-2\eta-2\sqrt{p_H}.
$$
Section 8.4 starts from this one fixed seeded word. The source inputs are the
[synthesis statements][q-synthesis-statement] and
[helper theorem][q-helper-statement].

**Seed removal.** Fix that finite seeded word and an energy tolerance.
Reuse the same Gaussian at every occurrence on a vertex. Packet control and
the one-particle comparison then produce one finite ordinary word with the
required energy error. The [seed-removal theorem][q-seed-statement] supplies
an energy conclusion, which is stronger than convergence of seed marginals,
and so makes the ordinary-word transfer in §8.5 applicable.

**Fixed-word transfer.** Fix one ordinary word and all its angles before
either size or degree varies. Its positive tree scaled-edge limit equals
its SK expected-energy limit, with the convention conversion in §8.5.
This [fixed-word transfer][q-transfer] closes the thermodynamic-first
theorem in §8.6.

<!-- /passage:qaoa -->

<!-- passage:qaoa-depth-one-pauli -->
<a id="depth-one-calibration"></a>

**One cost–mixer pair.**
A single cost–mixer pair fixes the energy normalization before we
introduce the rooted representation. Use the positive cost $C_{n,J}$ of §8 and
$U=e^{-i\beta B_n}e^{-i\theta C_{n,J}}$, where the symbol $\theta$
is a cost angle. Put $c=\cos(2\beta)$
and $s=\sin(2\beta)$. The Pauli identities are
$$
e^{i\beta X}Ze^{-i\beta X}=cZ+sY,\qquad
e^{i\alpha Z_iZ_j}Y_je^{-i\alpha Z_iZ_j}
=Y_j\cos(2\alpha)+Z_iX_j\sin(2\alpha).
$$
The mixed term $Z_iY_j$ therefore has cost-state expectation
$$
\sin(2\theta J_{ij}/\sqrt n)
\prod_{k\ne i,j}\cos(2\theta J_{jk}/\sqrt n).
$$
The other mixed term exchanges $i$ and $j$. The $Z_iZ_j$ term
has zero expectation. The $Y_iY_j$ term can be nonzero on an individual
complete graph, but its expectation after the cost gate is independent
of $J_{ij}$, because $Y_iY_j$ commutes with that edge cost. Multiplying by
$J_{ij}$ and averaging therefore removes it, even in the presence of
triangles. For a standard Gaussian $J$,
$$
\mathbb E\cos(aJ)=e^{-a^2/2},\qquad
\mathbb E[J\sin(aJ)]=ae^{-a^2/2}.
$$
Averaging the mixed terms and summing the $n(n-1)/2$ unordered edges gives
$$
\frac1n\mathbb E_J\langle C_{n,J}\rangle
=\frac{n-1}{n}\,\theta\sin(4\beta)
  e^{-2\theta^2(n-1)/n}.
$$
This agrees with [Farhi et al., equation (62)][fggz]. With energy
maximization, $\theta=1/2$, $\beta=\pi/8$ gives the limiting value
$1/(2\sqrt e)$, and their minimization example uses $\beta=-\pi/8$
and the negative value.

For the positive cost $D^{-1/2}\sum_{\{i,j\}}Z_iZ_j$ on the
$(D+1)$-regular tree, each endpoint has $D$ other neighbours. The same
calculation now has no common-neighbour contribution and yields
$$
\langle Z_oZ_v\rangle
=\sin(4\beta)\sin(2\theta/\sqrt D)
                 \cos^D(2\theta/\sqrt D).
$$
Multiplying by $\sqrt D/2$ and letting $D\to\infty$ gives
$\theta e^{-2\theta^2}\sin(4\beta)$, the same limit as finite SK.
This confirms the factor $1/2$ and the degree convention in §8.5.

We recover the value just obtained once more from the
[rooted insertion in §8.1.1](#depth-one-rooted), which shows which part
of the limiting state carries the energy.
<!-- /passage:qaoa-depth-one-pauli -->

<!-- passage:qaoa-rooted -->

<a id="rooted-energy"></a>

### 8.1 Rooted realisation and the observable it preserves

The fixed-depth spin–boson representation has an antecedent in
[Boulebnane et al., Theorem 2 (informal), equations (10)–(11), p. 3][spin-boson]:
as the degree grows with depth fixed, the symmetric tree state converges
to a spin coupled to finitely many bosonic modes. The QAOA manuscript
credits that representation and then proves the occupation, insertion and
surrounding-circuit estimates needed below. These estimates control the
energy observable and substitutions inside a circuit, which goes beyond
identifying a limiting state. [OpenAI (2026b), §2][q-fock].

We now realise the finite Gaussian spin from §7 in a space where its
shifted correlation sum becomes an energy observable. We identify that sum
with a rooted quadratic form, and then relate the same form to the scaled
edge energy of a fixed circuit. Section 8.5 supplies the final transfer
to SK.

For an ordinary depth-$p$ word, this tree calculation uses only finitely
many spins. Conjugate $Z_iZ_j$ backwards through the circuit. A mixer
rotates the spins already in its support without enlarging that support.
In a cost layer all edge factors commute, so factors disjoint from the
current support commute past the other factors and cancel. The remaining
factors enlarge support by at most one graph step. After $p$ pairs, the
edge expectation therefore depends only on its radius-$p$ backward
light cone. On the $(D+1)$-regular tree this contains at most
$2\sum_{r=0}^pD^r$ vertices, and the spins outside factor out of the
product-state expectation.

Thus fixed depth gives fixed tree height,
although the number of children grows with $D$. This locality explains
the finite tree calculation, and the equality with dense SK is the
separate transfer theorem of §8.5. [Farhi, Goldstone and Gutmann (2014), §II,
equations (14)–(26)](https://arxiv.org/abs/1411.4028v1).

On a rooted tree, factor a circuit as $U=(\bigotimes_jU_j)u$,
where each $U_j$ contains operations wholly in child subtree $j$.
Conjugating a root observable $P$ reduces to $u^*Pu$, and its insertion
vector is $u^*Pu\,\mathrm{vac}$, centered when its vacuum mean is zero.
This factorization explains how child operations enter the recursion below.

The first change of setting preserves a specified quadratic form. At
height zero the pointed space is
$\mathsf k=\mathbb C^2\otimes L^2(\text{seed})$, with distinguished vector
$\omega=|+\rangle\otimes1$. Recursively,
$$
\mathcal H_0=\mathsf k,\qquad
\mathcal H_{h+1}=\mathsf k\otimes\Gamma_s(\mathcal H_h^\circ),
$$
where $\mathcal H_h^\circ$ is the orthogonal complement of the
vacuum and $\Gamma_s$ is symmetric Fock space. A particle here is one
child subtree whose state differs from its vacuum. It is not one
physical spin. The occupation embedding identifies a symmetric state
with $m$ nonvacuum children among $D$ available children with its
$m$-particle vector, using the factor $\binom Dm^{-1/2}$.
Adding empty descendants gives compatible embeddings between heights.
[OpenAI (2026b), §2, pointed spaces and occupation embeddings][q-fock].

For a child direction $f$, write
$q(f)=a^*(f)+a(f)$ and $W(f)=e^{iq(f)}$. Explicitly,
$\Gamma_s(H)=\bigoplus_{m\ge0}H^{\otimes_s m}$, with vacuum $\Omega$
in sector zero. On sector $m$,
$a^*(f)\psi=\sqrt{m+1}\operatorname{Sym}(f\otimes\psi)$, and $a(f)$
is its adjoint. The number operator is multiplication by $m$, and more
generally $d\Gamma(A)$ sums $A$ over those $m$ child factors. The
vacuum characteristic function
$\langle\Omega,e^{itq(f)}\Omega\rangle=e^{-t^2\|f\|^2/2}$ fixes the
Gaussian normalization.

For any finite list of fixed centred real child vectors, their parent
positions $\mathsf G(f)=q(f)$ are jointly centred Gaussian, and for $f,g$,
$$
\mathbb E[\mathsf G(f)\mathsf G(g)]=\langle f,g\rangle,
\qquad
\|[\mathsf G(f)-\mathsf G(g)]\mathrm{vac}\|=\|f-g\|.
$$
The directions are fixed elements of the child space, even when their
values depend on the child's Gaussian coordinates.
The receiving root's own seed is a separate, independent tensor factor.

The root shift
$$
\mathsf Lf=\omega\otimes a^*(f)\Omega
$$
is an isometry on centred vectors, and is extended by zero on the
vacuum. Its adjoint extracts the root-vacuum, one-child component,
including a constant receiving-root seed. For a classical function of
orthonormal Gaussian positions, this is its first Gaussian chaos.
The next calculation uses the covariance, independence and projection
properties in distinct roles.
[OpenAI (2026b), §2, root shift and classical first chaos][q-root-energy]
[OpenAI (2026b), §2, classical Gaussian positions][q-gaussian].

We start with the standard root seed $g_0=S$ and recursively set
$Z_{j+1}=\mathsf G(g_j)$. Suppose $g_0,\ldots,g_{j-1}$ are centred
orthonormal vectors. The joint Gaussian law of their lifts and the covariance
identity give independent standard normals $Z_1,\ldots,Z_j$. In the scalar
construction of §7, the multiplier
of $Z_j$ in $g_j$ depends only on its predecessors. Conditional centring
therefore makes $g_j$ orthogonal to $g_1,\ldots,g_{j-1}$, and the chosen
normaliser $c_j$ gives unit norm. Independence of the receiving-root
seed gives orthogonality to $g_0$. This closes the induction. Compatible
height embeddings ensure that the earlier variables are the same
variables when the formula is embedded at greater height.

Oddness of $u$ and evenness of $a$ make $Y_j$ and $g_j$ odd under
simultaneous seed reversal. The extra centred rounding variable
$2F_{\rm G}(Z_{N+1})-1$ is odd as well, so $F_N$ is centred. It is
independent of the receiving-root seed, hence
$\langle F_N,g_0\rangle=0=A_0$. With $A_j,B_{j+1}$ as in §7,
first-chaos projection gives
$$
\mathsf L^*F_N=\sum_{j=0}^NB_{j+1}g_j,\qquad
\mathcal E(F_N):=\operatorname{Re}\langle F_N,\mathsf LF_N\rangle
 =\sum_{j=0}^NA_jB_{j+1}.
$$
Thus the construction identifies both the scalar law and the required
rooted energy. The Gaussian covariance rule does not declare distinct
calculations on shared descendant subtrees independent.
[OpenAI (2026b), §3, “A finite Gaussian formula and its rooted energy”][q-rooted-classical].

Connecting this energy to an edge observable needs a separate statement.
Let $f_D$ be
centred, invariant under rooted permutations of complete child
subtrees, and supported at a fixed finite radius. Suppose its occupation
embedding converges strongly to $v$, with a uniformly bounded expected
top-generation occupation $\langle f_D,\mathsf Nf_D\rangle$. Put
$M=D+1$ and $R_M=\sqrt{1-\mathsf N/M}$.

To obtain the decomposition, we delete the distinguished edge and project
onto the vacuum on either resulting half-tree. There are four pairings.
The vacuum–vacuum pairing is zero by centring. If only one side is
excited, the other side forces the root and all remaining branches into
their vacuum, and root symmetry gives the term
$M^{-1/2}\langle\mathsf L^*f_D,R_Mf_D\rangle$ or its conjugate.
In the pairing where both sides are excited, each projected vector has
squared norm $M^{-1}\langle f_D,\mathsf Nf_D\rangle$ by symmetry
among the $M$ branches. Cauchy–Schwarz bounds their pairing by that same
quantity. Finally, deleting one vacuum child multiplies the $m$-particle
sector by $\sqrt{\binom{M-1}{m}/\binom{M}{m}}=\sqrt{1-m/M}$,
which identifies $R_M$. Thus the exact edge decomposition is
$$
\langle f_{D,i},f_{D,j}\rangle
 =\frac2{\sqrt M}\operatorname{Re}\langle\mathsf L^*f_D,R_Mf_D\rangle
   +\varepsilon_D,
\qquad
|\varepsilon_D|\le\frac{\langle f_D,\mathsf Nf_D\rangle}{M}.
$$
Vacuum deletion also gives
$\|R_Mf_D-f_D\|^2\le M^{-1}\langle f_D,\mathsf Nf_D\rangle$.
Consequently,
$$
\lim_{D\to\infty}\frac{\sqrt D}{2}
 \langle f_{D,i},f_{D,j}\rangle=\mathcal E(v).
$$
Strong convergence alone does not control the scaled remainder, and the
occupation estimate for the relevant finite circuits comes later. Norm
convergence of the vectors cannot replace the occupation assumption.
For orthogonal unit vectors $e_m$ in sector $m$,
$\psi_k=\sqrt{1-1/k}\,\Omega+k^{-1/2}e_{k^2}$ tends to $\Omega$
in norm, whereas $\langle\psi_k,\mathsf N\psi_k\rangle=k$.

The elementary continuity bound
$$
|\mathcal E(v)-\mathcal E(w)|
 \le(\|v\|+\|w\|)\|v-w\|
$$
controls every later approximation once the observable has been
identified. [OpenAI (2026b), §2, “Energy as a root shift,” including
its exact finite-degree decomposition][q-root-energy].

The three identifications use different information:

First, passing from the finite Gaussian formula to the rooted classical
sector preserves the joint Gaussian law and the same $F_N$. This uses
covariance, conditional centring and independence of the receiving-root seed.

Second, the root-vacuum and first-chaos projection identify the energy itself:
$\mathcal E(F_N)=S_{T,N}$.

Third, for a finite-degree circuit insertion tending to $v$, root invariance,
fixed radius, strong convergence and uniform occupation control imply that
the scaled edge expectation tends to $\mathcal E(v)$.

We gave the first two calculations above. The third uses the source's
edge decomposition, and the circuit estimates it requires enter in §8.2.
<!-- /passage:qaoa-rooted -->

<!-- passage:qaoa-depth-one -->
<a id="depth-one-rooted"></a>

#### 8.1.1 The same energy in the rooted representation

We return to the value $\theta e^{-2\theta^2}\sin(4\beta)$ obtained
in [the finite-SK and tree calculation](#depth-one-calibration).
We keep its cost angle $\theta$, mixer angle $\beta$,
$c=\cos(2\beta)$ and $s=\sin(2\beta)$.
Here $\Omega$ is the Fock vacuum, $\mathrm{vac}=|+\rangle\otimes\Omega$,
and $q(e)=a^*(e)+a(e)$. The root shift $\mathsf L$ defined in §8.1
creates one child excitation and puts the receiving root in its vacuum.
Let $e=Z\mathrm{vac}=|-\rangle\otimes\Omega$ and
$Q=q(e)$ in the rooted description. Its vacuum law is standard normal.
The depth-one root operation is $u=e^{-i\beta X}e^{-i\theta ZQ}$.
Using $Y|+\rangle=-i|-\rangle$, its insertion is
$$
v=u^*Zu\,\mathrm{vac}
=|-\rangle\otimes[c-is\cos(2\theta Q)]\Omega
 +|+\rangle\otimes s\sin(2\theta Q)\Omega.
$$
The root shift retains only the plus-root, one-particle component.
The first-chaos coefficient of the sine is
$\mathbb E[G\sin(2\theta G)]=2\theta e^{-2\theta^2}$, so
$$
\mathsf L^*v=2\theta s e^{-2\theta^2}e,\qquad
\operatorname{Re}\langle v,\mathsf Lv\rangle
=2\theta sc e^{-2\theta^2}
=\theta e^{-2\theta^2}\sin(4\beta).
$$
Here $\operatorname{Re}\langle e,v\rangle=c$ supplies the final factor.
Thus the observable itself, as well as the Gaussian law, agrees in all
three representations. The example uses the [ordinary root updates][q-ordinary-updates]
and [root shift][q-root-energy]. It calibrates one depth, and at higher
depth the occupation and replacement hypotheses remain necessary.
<!-- /passage:qaoa-depth-one -->

<!-- passage:qaoa-regularization -->

<a id="circuit-domains"></a>

### 8.2 Finite regularisation and the calculus needed for compilation

We fix $T,N$ first. The independent continuous rounding coordinate makes
the event at the sign discontinuity null. Bounded continuous odd
approximations to that sign therefore converge in $L^2$ by dominated
convergence. If the bounded output is $m^*$, shrink it to
$(1-\kappa)m^*$, with fixed $0<\kappa<1$, before taking
$\theta_\kappa=\tfrac12\arcsin((1-\kappa)m^*)$. This margin keeps
the derivative of inverse sine bounded on its used range, and the output
loss is at most $\kappa$ in $L^2$.

We then truncate the transmitted functions and approximate the resulting bounded maps by trigonometric
polynomials. The source makes these choices from output
toward leaves: each outer Lipschitz function is fixed before its inputs
are approximated. The Gaussian isometry above converts an error in a
transmitted child vector into the same $L^2$ error in its receiving
position. Energy continuity then bounds the accumulated loss. The
bounded output has the form $m=\sin(2\theta)$. The transmitted columns are
bounded, while the Gaussian fields themselves remain unbounded. [OpenAI (2026b), §3, bounded trigonometric approximation][q-bounded-classical].

**The formulas to be synthesized.** A mask is an event in the local seed
space invariant under seed reversal. Fix a finite partition into masks,
and write $p_R$ for the probability of mask $R$. Probabilities used as
denominators are positive. For disjoint masks $R,S$, aggregation on the
finite regular tree means
$$
(A_{RS}f)_i=\mathbf1_R(i)D^{-1/2}
             \sum_{j\sim i}\mathbf1_S(j)f_j.
$$
A formula has a finite expression tree, with bounded measurable functions
of the local seed as leaves. Internal operations are such aggregations or
pointwise functions of previously formed quantities. Pointwise functions
are smooth in their aggregate arguments, with polynomial bounds on all
derivatives, uniformly in the seed. Seed dependence itself need not be
smooth. Transmitted quantities are odd under simultaneous seed reversal,
and sent nonlinear functions and output angles are bounded, while aggregate
Gaussian fields may be unbounded.

An aggregation from $S$ to $R$ excludes $R$ if its transmitted formula
uses no vertex of type $R$ below that call. All ancestor exclusions are
retained, so a mask cannot recur on one active expression branch. It may
recur on a different branch, and distinct branches may share descendants,
since exclusion asserts no independence between them. Gate targets are likewise
excluded from the cross-edge dependencies of their parameters. This makes
the physical all-neighbour calculation agree with the intended descendant
calculation.

For synthesis, we reserve disjoint helper masks of positive mass,
one per positive aggregation depth and absent from the output
dependencies. The bounded odd output angle vanishes on their union.
[OpenAI (2026b), Definition 3.3][q-formula-definition];
[Theorem 4.1][q-synthesis-statement].

The types can be encoded together
with a Gaussian seed in one raw Gaussian variable: its sign supplies
the odd symmetry, while disjoint binary digits of a uniform transform
of its absolute value supply independent coordinates. The resulting
finite type label is even and the Gaussian seed is odd under reversal.
The associated pullback preserves vacuum, parity and $\mathsf L$.
Thus adding types preserves the energy identification and needs no
additional circuit resource. [OpenAI (2026b), §3, colours,
exclusions and helpers][q-colors].

For the already fixed formula, the exclusion estimate is
$$
\|m^{\rm col}-m\|^2\le C(\alpha+p_H),
$$
where $\alpha$ is the largest nonhelper type mass and $p_H$ is the total
helper mass. Its constant is independent of the number of types. The
proof uses orthogonality of the disjoint masks before applying the
Lipschitz estimate in the original, fixed list of formula coordinates.
Treating the refined types as an arbitrary growing list of independent
arguments would not provide this bound. The resulting classical
near-optimality proposition supplies a finite bounded odd formula,
zero on helpers, with energy as close to $P_*$ as required.
[OpenAI (2026b), §3, “Type exclusion at small cost” and “Classical
near-optimality in synthesis form”][q-colors].

**The topology of the later substitutions.**
The gate calculus uses recursive weights
$\mathsf W_{h+1}=\mathsf W_0+d\Gamma(\Lambda_h)$, where
$\Lambda_h$ is the previous weight restricted to the centred child
space. The local weight includes the unbounded raw seed when needed.
Its smooth domain is the intersection of all weighted domains. Smooth
bounds mean polynomial bounds, with a finite loss of weight powers,
for the maps, their adjoints and every specified parameter derivative.
The finite height, fixed expression and fixed number of channels matter
in these estimates.

Strong convergence together with uniform stronger
weighted bounds gives convergence in each weaker weighted norm by a
spectral cutoff. This is the topology in which one replaces a gate
inside a fixed surrounding circuit.
[OpenAI (2026b), §2, weights and the topology of convergence][q-weights].

More precisely, the imported smooth-convergence contract is: for each
output power $m$ and parameter multi-index $\alpha$, there are an input
power $r$, exponent $k$ and constant $C$ such that
$$
\|\mathsf W_D^m\partial_x^\alpha F_D(x)\xi\|
\le C(1+|x|)^k\|\mathsf W_D^r\xi\|,
\qquad \xi\in\bigcap_{q\ge0}\operatorname{Dom}(\mathsf W_D^q).
$$
The bound is uniform in degree $D$ and parameter $x$. The same bound
and strong convergence on smooth inputs are required for the adjoints
and all specified parameter derivatives, with compatible limiting
adjoints. Constants may depend on the fixed expression, height, channel
list and derivative order. Bounded adjoint derivatives alone do not
supply their convergence. At bounded local seeds $\mathsf W_0=1$, and
for polynomial Gaussian growth the source uses the vacuum-preserving
weight $P_0+Q_0(1+|s|)Q_0$, where $P_0$ projects to the local vacuum
and $Q_0=1-P_0$. These recursive weights control every generation.

They differ from the top-generation number weight in §8.4's packet
argument. No uniformity in growing word length, vanishing helper mass
or the later seed-removal parameter is implicit in this contract.
[OpenAI (2026b), §2, smooth-family bound][q-weights].

One of the essential inputs retains a term beyond the Gaussian field.
Let $K_D(x)$ be a smooth unitary family with $K_D(0)=1$, uniformly
controlled derivatives and adjoint derivatives that converge smoothly
in the stated weighted topology to those of a limiting family $K$,
and centred first vacuum derivatives. Write
$K_\alpha=\partial_\alpha K(0)=iT_\alpha$ and
$K_{\alpha\lambda}=\partial_\alpha\partial_\lambda K(0)$.
For a fixed finite parameter list $a$, the source proves
$$
K_D(a/\sqrt D)^{\otimes D}\longrightarrow
\exp\!\left\{
 i\sum_\alpha a_\alpha q(T_\alpha\mathrm{vac})
 +\frac12\sum_{\alpha,\lambda}a_\alpha a_\lambda
 \left\langle K_{\alpha\lambda}
             +\tfrac12\{T_\alpha,T_\lambda\}\right\rangle_0
\right\}.
$$
The second term is imaginary by unitarity. It survives because a
second-order change of size $D^{-1}$ occurs in each of $D$ children.
Its scalar coefficients become a possibly nonconstant controlled
operator when commuting signals replace $a$. For example,
$K(x)=e^{ix^2}$ has zero first derivative at zero but
$K(a/\sqrt D)^D=e^{ia^2}$ exactly. Keeping only the first derivative
would miss even this elementary contribution. The example explains
why the term is necessary, and the source's weighted product lemma supplies
the general convergence and domain bounds.
[OpenAI (2026b), §2, “Products of small unitaries”][q-unitary].

The smooth-substitution lemma makes the required joint control explicit.
For a fixed finite list of strongly commuting self-adjoint signals
$S_D=(S_{D,1},\ldots,S_{D,b})$, their polynomials must converge smoothly,
and their joint exponentials must converge strongly on smooth inputs and
satisfy, for every output weight $m$,
$$
\|\mathsf W_D^m e^{it\cdot S_D}\xi\|
 \le C_m(1+|t|)^{k_m}\|\mathsf W_D^{r_m}\xi\|,
$$
uniformly in $D$ and $t\in\mathbb R^b$. The coefficient family $F_D(x)$
and its adjoint obey the smooth bounds above and converge with all their
jets on compact parameter sets. Each coefficient must commute with the
signals' spectral projections. For unbounded coefficients this means
commutation in the closed-operator sense. Seed-dependent substitution also
requires strong commutation with the receiving seed multiplication algebra.
These hypotheses permit the joint spectral and Fourier calculation of
$F_D(S_D)$, including its derivatives and adjoints. A formal commutator
identity on a single domain would not supply them. No nonsingular covariance
matrix or inverse Gram matrix is required.
[OpenAI (2026b), Lemma 2.10, “Smooth commuting substitution”][q-substitution].

For the formula fixed here, its finite expression tree supplies the finite
channel list. Exclusion supplies the commuting physical control algebra,
and normalized-sum limits and the preceding product estimates supply the signal
limits and weighted bounds. Conjugation by the already fixed preceding
circuit preserves strong commutation, including when signals share
descendants. The gate-calculus theorem applies these facts inductively to
the formula and its adjoints, at each cut and inside the fixed surrounding
sequence. It is this theorem, with these constructed inputs, that licenses
the later replacements. Its conditional root-seed bounds also give the
masked-column estimates used in §8.3.
[OpenAI (2026b), Theorem 2.12][q-gate-calculus].

The exact cut recursion records the effect of such shared lower
calculations. Let $y$ be the incoming boundary data, $V(y)$ the root
matrix, and $m(y)$ its outgoing signals in the current frame. Factoring
out the operations wholly inside each child gives
$$
E^{[d]}(y)=V(y)\prod_{j\ \mathrm{child}}
 K_j^{[d-1]}(D^{-1/2}m(y)),\qquad
K^{[d]}(y)=u^*E^{[d]}(0)^*E^{[d]}(y)u.
$$
Every ratio $K^{[d]}$ equals one at zero. Nevertheless $m(0)$ generally
does not vanish, because removing the parent input leaves the root's
own signals. The actual zero-boundary root update is
$u_{\rm new}=E^{[d]}(0)u$. It is not recovered by setting the entire
ratio to one and discarding the baseline operation. Applying the
unitary product lemma to this recursion, and then commuting
substitution, is the source's smooth gate calculus. Its replacement
statement covers a fixed finite surrounding sequence and its required
parameter derivatives. It asserts no uniformity for arbitrary sequences
whose length or channel count grows.
[OpenAI (2026b), §2, smooth gate calculus and its cut recursion][q-gate-calculus].

<!-- /passage:qaoa-regularization -->

<!-- passage:qaoa-synthesis -->

<a id="helper-contraction"></a>

### 8.3 Seeded synthesis: helper contraction and nested responses

We temporarily add longitudinal seed gates
$\exp(-it\sum_i s_iZ_i)$, using an independent Gaussian $s_i$ at each
vertex and reusing that same variable whenever a later seed gate acts
there. The synthesis theorem turns the finite classical formula into
a finite word of these gates, costs and mixers. Its ideal local
$Y$ rotations have $Z$ insertion
$$
v_{\rm cl}=|+\rangle m+|-\rangle n,\qquad m=\sin(2\theta).
$$
Here $n$ is the remaining seed-dependent coefficient of the $|-\rangle$
component, not the number of spins. Both coefficients may depend on
descendant seeds.
The second term does not contribute to $\mathcal E$: $\mathsf L$ puts
the root spin in $|+\rangle$, and shifting the $|-\rangle$ term puts a
spin excitation into a child, orthogonal to the classical sector.
Thus $\mathcal E(v_{\rm cl})=\mathcal E(m)$.
[OpenAI (2026b), §2, classical rotations][q-root-energy]
[OpenAI (2026b), §4, “Classical synthesis”][q-synthesis-statement].

There are two different approximation operations. First, seed
conjugations of $X$ generate $\cos(2ts)X$ and $\sin(2ts)Y$.
Trigonometric approximation and matrix commutators give the required
bounded measurable masked spin controls. Sign averaging of masked
$X$ flips isolates an interaction between chosen types. Since the cost
terms commute, this isolation is exact. This sign averaging is the
refocusing method of [Jones and Knill (1999), §II, p. 2](https://arxiv.org/abs/quant-ph/9905008v1):
orthogonal sign rows cancel unwanted Ising couplings, while identical
rows retain a selected coupling. The source applies it to the finite
partition of seed masks.

Second, finite commutator products
then approximate the desired pointwise matrices with smooth bounds.
They can be substituted after the surrounding circuit has been fixed.
The local masks are measurable functions of the seed. Smoothness is
required only in the continuous aggregate inputs, and the mask labels
need none.
[OpenAI (2026b), §4, elementary masked controls and matrix products][q-masked-controls].

The second operation implements an aggregation. Let $I$ be the
receiver region, $J$ the sender region, and $K$ an unused helper region
of fixed positive probability $p_K$. They are disjoint. The lower
seed-only formula $f$ excludes these regions along its cross-edge
paths. With $A_{ij}=D^{-1/2}1_{i\sim j}$, the desired propagation has
generator
$$
H=\sum_{i\in I,j\in J}P_iA_{ij}f_j.
$$
Partition $J$ into parts $J_l$ of probabilities $p_l$. Two levels of
spin commutators generate the commuting Hamiltonian
$$
\widetilde H_{\mathcal P}
 =\frac1{p_K}\sum_l\sum_{i\in I,\,j,j'\in J_l,\,k\in K}
 P_iA_{ij}X_jA_{jk}X_kA_{kj'}X_{j'}f_{j'}.
$$
For a fixed receiver–sender edge $i\sim j$, the terms with $j=j'$
contain $X_j^2=1$, and $f_j$ is seed-only and commutes with these spin factors.
On helper-pinned states $X_k=1$, so their sum is exactly
$$
\frac1{p_K}P_iA_{ij}f_j
 \sum_{k\in K}A_{jk}A_{kj}
=P_iA_{ij}f_j\frac{N_K(j)}{Dp_K},
$$
where $N_K(j)$ counts helper neighbours of $j$. Conditional on the receiver
edge, its endpoint $i\in I$ is not a helper and the other $D$ neighbours
have mean helper count $Dp_K$. Its binomial variance is $Dp_K(1-p_K)$,
so the ratio tends to one in $L^2$ at fixed $p_K>0$.
Terms with $j\ne j'$ are simple paths and must disappear under
partition refinement. The factor $1/p_K$ is harmless only after a
positive helper mass has been fixed.
[OpenAI (2026b), §4, helper generators and contraction][q-helper-statement].

Helper pinning is a statement about every helper spin in the light
cone. Its finite-degree projection is
$$
\Pi_{K,V}=\prod_{v\in V}
 \left[1-1_K(s_v)\frac{1-X_v}{2}\right].
$$
It retains the seed variables and helper-labelled descendant spaces
while fixing the helper spins in $|+\rangle$. The ideal propagation,
the exact Hamiltonian $\widetilde H_{\mathcal P}$, and the surrounding
gates that avoid $K$ preserve this compressed space. Multiplying an
insertion by $1_{K^c}(s_o)$ only masks the root label and does not impose
this descendant condition. Individual finite commutator factors may
leave the compressed space. Accordingly, the proof first establishes
the exact contraction there, then chooses finite commutator products
on the full space to approximate it. [OpenAI (2026b), §4,
“Simultaneous helper contraction”][q-helper-statement].

The estimate controlling a growing partition is
$$
\rho^2=\sum_l p_l^2\le p_J\max_l p_l\longrightarrow0.
$$
At a helper root, each part supplies two commuting fields with columns
$d_l,e_l$ on disjoint sender masks, of weighted norm at most
$C\sqrt{p_l}$. The remaining correction is a sum of centred quadratics
$$
\sum_l\bigl[q(d_l)q(e_l)-\langle d_l,e_l\rangle\bigr].
$$
Normal ordering and disjoint-mask orthogonality bound its quadratic
kernels by $C\rho$ in Hilbert–Schmidt norm. The source separately
bounds the exponentials in the weighted topology, uniformly in the
number of parts. Smallness of the generator on smooth vectors alone
would not justify inserting its exponential into the surrounding
circuit. [OpenAI (2026b), §4, “Quadratics with many orthogonal
columns”][q-helper-quadratics].

The helper estimate alone does not control all cuts. Put
$\epsilon_{\mathcal P}=\max_l\sqrt{p_l}$. The receiver and helper
corrections have weighted size $O(\rho)$, and a sender correction is
$O(\sqrt{p_l})$ conditional on its label. The remaining lower-formula
responses have size $O(\epsilon_{\mathcal P})$. Both quantities tend
to zero with $f$, the number of lower channels and $p_K>0$ fixed.

There remains a response inside the lower formula $f$. A cut can pass
through a descendant used to compute $f$, rather than through one of
the displayed three-edge paths. The source first writes the exact
difference of the affected Hamiltonian terms before applying the
preceding subtree circuit. That circuit is fixed with respect to the
new cut input. Exclusions ensure that the downward coefficients are
independent of that active boundary argument. The resulting additive
cut-response rule retains both first derivatives and mixed second
derivatives:
$$
\sum_\alpha a_\alpha q(H_\alpha\mathrm{vac})
 +\frac12\sum_{\alpha,\lambda}a_\alpha a_\lambda
       \langle H_{\alpha\lambda}\rangle_0.
$$
Here $H_\alpha$ is centred by parity, so no limit of the extensive
zero-input Hamiltonian is required. The self-adjoint generator is
obtained from the limiting time group of cut ratios.
[OpenAI (2026b), §2, “Additive cut responses”][q-additive-response].

After substituting the signals of a lower formula, a boundary
difference contains, in particular,
$$
\begin{aligned}
&\sum_\alpha[m_\alpha(y)-m_\alpha(0)]q(r_\alpha)\\
&\quad+\frac12\sum_{\alpha,\lambda}
 [m_\alpha(y)m_\lambda(y)-m_\alpha(0)m_\lambda(0)]
 c_{\alpha\lambda}.
\end{aligned}
$$
The mixed coefficients $c_{\alpha\lambda}$ must remain even when two
channels share descendants. A bounded scalar example makes the issue
visible: for a shared standard normal $Q$, take
$H(y_1,y_2)=\cos(Q+y_1+y_2)-\cos Q$. At zero both first derivatives
are $-\sin Q$ and are centred, but
$\mathbb E H_{12}=-e^{-1/2}$. The two cross terms in the half-sum
therefore contribute $-e^{-1/2}a_1a_2$. Centred first responses do not
remove a mixed second response. This example checks the algebraic reason
for retaining the term, and the cut-response theorem supplies the operator
and domain justification.

At the actual root there is no parent
input to subtract: the corresponding unsubtracted $m(0)$ terms enter
the root update. The source propagates small terminal responses
through the fixed depth and fixed list of lower channels. It refines
the calling sender mask while keeping those lower formulas fixed.
This is what controls all cuts in a simultaneous replacement of
several calls. A root-seed scalar phase may remain on helper labels, where
it cancels in insertions and child conjugations.
[OpenAI (2026b), §4, lower-formula responses and completion of the
contraction proof][q-helper-responses].

Compilation now proceeds from outer aggregations to inner ones.
Reserve a separate unused helper type for each aggregation rank. At
a given rank, establish the simultaneous exact contraction, fix the
sender partitions, and only then choose finite commutator
approximations. Subsequent inner calls use their own helpers, which
remain unused. Each inner accuracy is chosen after its surrounding
finite sequence is fixed. This order avoids assuming uniformity over
all helper masses, ranks or circuit lengths.
[OpenAI (2026b), §4, order of compilation][q-compilation].

If the resulting insertion $v$ obeys
$\|1_{H^c}(v-v_{\rm cl})\|<\eta$, seed commutation and unitarity give
$\|1_Hv\|=\sqrt{p_H}$. First omit the helper part of $v$, at energy cost at most
$2\sqrt{p_H}$. Compare the two insertions projected off helpers.
Their norm difference is below $\eta$, so their energies differ by
at most $2\eta$. The projected classical insertion still has energy
$\mathcal E(m)$ because $m=0$ on helpers. Thus
$$
|\mathcal E(v)-\mathcal E(m)|\le2\eta+2\sqrt{p_H}.
$$
All projected insertion vectors have norm at most one. This completes the
interface from a bounded classical formula to a finite seeded word.
It uses the source's smooth gate and helper-contraction theorems, and
neither the scalar theorem nor a finite-degree central limit theorem
alone implies it.
[OpenAI (2026b), §4, synthesis statement and its proof][q-synthesis-statement].

<!-- /passage:qaoa-synthesis -->

<!-- passage:qaoa-seed -->

<a id="seed-removal"></a>

### 8.4 Removing the seed controls without changing their joint law

Seed removal must simulate one Gaussian variable per vertex, reused
at all seed gates. Replacing each occurrence with a fresh independent
Gaussian would change the circuit. The source proves a theorem for
any already fixed finite seeded word and any energy tolerance, producing
one finite ordinary word. Its proof works in the fixed-word tree
limit and never takes a degree limit simultaneously with its small
parameter. [OpenAI (2026b), §5, “Removal of seed controls”][q-seed-statement].

**A small component supplies a large response.**
Fix $b>0$ and $0<r<1$, and put
$$
R_*=e^{-i\pi X/4},\qquad \delta_r=\arcsin r,\qquad
\gamma_r=b/r,\qquad e=Z\mathrm{vac},\qquad
\ell_r=Y e^{-2i\gamma_r Zq(e)}\mathrm{vac}.
$$
Apply, in chronological order, cost $\gamma_r$, mixer $\delta_r$,
cost $-\gamma_r$ and mixer $\pi/4$. Just before the inverse cost the
insertion is $v_{2,r}=\cos(2\delta_r)e+\sin(2\delta_r)\ell_r$. Define
$$
d_{0,r}=\gamma_r(v_{2,r}-e)
       =-2br\,e+2b\sqrt{1-r^2}\,\ell_r,
\qquad W_Z(f)=e^{iZq(f)}.
$$
Expanding the first mixer and using the Weyl identity gives the exact split
$$
u_r=a_{0,r}U_{O,r}+a_{1,r}U_{S,r},\qquad
U_{O,r}=R_*W_Z(d_{0,r}),\quad
U_{S,r}=R_*XW_Z(-2\gamma_re-d_{0,r}),
$$
where
$$
a_{0,r}=\cos\delta_r\,e^{i\vartheta_r},\quad
 a_{1,r}=-ir e^{-i\vartheta_r},\quad
\vartheta_r=\gamma_r^2\operatorname{Im}\langle v_{2,r},e\rangle.
$$
Both component operators are unitary. The special component
$\psi_{S,r}=a_{1,r}U_{S,r}\mathrm{vac}$ therefore has norm exactly $r$.
Its normalized contribution to the insertion is
$$
w_r=u_r^*Z\psi_{S,r}/r,\qquad \|w_r\|=1.
\tag{8.1}
$$
This is the direction amplified in the echo below. The two components
need not be orthogonal. The ordinary component has a bounded displacement
$d_{0,r}=2b\ell_r+o(1)$, whereas the special component contains a displacement
of order $1/r$. Subsequent unitaries preserve each component's norm.
Only the undisplaced packet tails have bounded occupation moments.
Unitarity carries the norm remainders, and no such bounds are assigned
to them.

The source's high-direction lemma gives that $\ell_r,w_r$ are centred
unit vectors, become mutually orthogonal, and escape every fixed
occupation sector. Splitting a uniformly bounded family of occupation
moments at a fixed sector cutoff shows that its pairings with these
directions vanish: first let $r\downarrow0$, then send the cutoff to
infinity. The special component must nevertheless be retained because
the later echo magnifies its insertion contribution by $1/r$.
[OpenAI (2026b), §5.2, exact initialization and high directions][q-seed-packets].

A packet has the form $W(d_r/r)\eta_r$, where the scaled centre $d_r$
is bounded and the tail $\eta_r$ has uniformly bounded occupation
moments. Packets with centres separated by a fixed positive amount
have overlap smaller than every fixed power of $r$. A short probe
along a Pauli axis $P$ sees a finite rotation determined by
$-2\operatorname{Im}\langle v_P,d_r\rangle$, as well as a small
Weyl displacement. The direction $v_P=u^*Pu\,\mathrm{vac}$ is the
actual current insertion. Toggling a cost rotates both its root Pauli
axis and this direction, so the two cannot be selected independently.
[OpenAI (2026b), §5, packet estimates][q-seed-packets].

For the finite list of nonzero special centres $h_j\in\mathbb R^2$,
the source constructs the same rotation $e^{-i\theta X}$ at every
centre, with *zero clocks*: the signed pulse durations on each of two
perpendicular axes in the $ZY$ plane sum exactly to zero. Choose the
axes so that $c_{j,a}=h_j\cdot P_a\ne0$, and the square pairs
$(c_{j,1}^2,c_{j,2}^2)$ are distinct except for equal or opposite centres.
Brackets have zero clock components and generate $iX$ coefficients
$c_{j,1}c_{j,2}F(c_{j,1}^2,c_{j,2}^2)$. Polynomial interpolation makes
all these coefficients one. Equal or opposite centres have the same
product, so they cause no obstruction to this common rotation.

This does not give arbitrary independent rotations: opposite centres
obey $U_{-h}=XU_hX$. Exact finite reachability of the required common
rotation follows by taking finitely many reachable conjugates spanning
the generated Lie algebra. Their product map contains an identity
neighbourhood, and subdivision reaches the whole one-parameter path.
Closure alone would not give exact cancellation. The Lie-bracket and
polynomial method follows [Li and Khaneja (2005), §II, Examples 1–2,
pp. 2–3](https://arxiv.org/abs/quant-ph/0510012v1); the finite-family,
equal/opposite-centre and exact-clock assertions are proved in the
QAOA source itself.

The distinction is exact: zero net probe clocks and the common target
rotation are assertions about the finite ensemble of limiting centres.
A finite pulse list realizing those assertions is fixed first. Its action
on the actual finite-$r$ packets is approximate. The separate accuracy
lemma then follows the actual insertion drift
through every probe prefix. Its ordinary-component error is $o(r)$, and
the error on the normalised special component is $o(1)$ and hence
also contributes $o(r)$ to the state. These bounds are uniform over
the bounded-tail families needed for insertions, and their constants
refer to the chosen finite pulse list.
[OpenAI (2026b), §5, “Selective rotations with zero clocks” and
“Accuracy of selective probes”][q-seed-selective].

The shared seed family is essential even for a single spin. Two gates
$e^{-itSZ}$ applied to $|+\rangle$, with the same standard Gaussian $S$,
give mean $X$ expectation $\mathbb E\cos(4tS)=e^{-8t^2}$.
Replacing the second seed by an independent copy gives $e^{-4t^2}$.
Agreement of marginal seed laws does not preserve a circuit's value.

The error scale needed below follows from the source's quantitative
probe estimate. For a fixed finite pulse list, separated packet centres
and bounded designated tail moments, write $\varepsilon_r=o(r)$ for
the discarded vacuum remainder. At each probe prefix $j$, with current
insertion $v_{P,r}^{(j)}$ and huge-cost direction $v_r$, the proof gives
$$
\max_j|\operatorname{Im}\langle v_{P,r}^{(j)},v_r\rangle|
 \le C_M(r^2+\varepsilon_r+r^M).
$$
Choose a fixed $M>1$. The right side is $o(r)$: within each ordinary
packet the relevant Hermitian expectation is real, distinct packet
centres have overlap $O(r^M)$, and terms containing only special
packets have size $O(r^2)$. Unitarity carries the norm remainder, which
receives no unproved occupation bound. The exact zero clocks then
cancel the first-order displacement on ordinary packets, leaving
$o(r)$ error there. The prescribed action on a normalized special
packet has $o(1)$ error, multiplied by its $O(r)$ amplitude.
These are the unequal accuracy requirements that the next echo uses.
[OpenAI (2026b), §5, accuracy of selective probes][q-seed-selective].

**The estimates must hold on the next input.**
The repeated echo uses more than the initial special norm in (8.1).
Let $\mathsf N$ count particles in the top child Fock space. Call a family
$(\eta_r)$ bounded smooth when
$\sup_r\|(1+\mathsf N)^k\eta_r\|<\infty$ for every fixed $k$.
At each stage between simulated gates, the source propagates a linear
ordinary/special decomposition on every such family, with
$$
\begin{aligned}
 u_r\eta_r&=\Psi_{O,r}(\eta_r)+\Psi_{S,r}(\eta_r),\\
 \|\Psi_{S,r}(\eta_r)\|&=r\|\eta_r\|,\\
 \Psi_{O,r}(\eta_r)&=\zeta_r A_r\eta_r+o(r).
\end{aligned}
$$
Here $|\zeta_r|=1$, and the low unitary $A_r$ and its adjoint preserve
all these seminorms uniformly. A separate part of the induction transports
the special insertion:
$$
 u_r^*Z\psi_{S,r}/r=w_r+o(1),
$$
with the negative sign between a selective flip and its undoing.
The norm remainders remain separate from the designated smooth tails,
and no occupation estimate is imposed on them.

This forward estimate gives the insertion approximation used at the next
gate. For $P=Y,Z$, the vector
$\widehat v_{P,r}=A_r^*PA_r\mathrm{vac}$ is itself a bounded smooth
family. Since $A_r\widehat v_{P,r}=PA_r\mathrm{vac}$, unitarity gives
$$
\begin{aligned}
\|u_r^*Pu_r\mathrm{vac}-\widehat v_{P,r}\|
&=\|Pu_r\mathrm{vac}-u_r\widehat v_{P,r}\|\\
&\le\|P(u_r-\zeta_r A_r)\mathrm{vac}\|\\
&\quad+\|(u_r-\zeta_r A_r)\widehat v_{P,r}\|=O(r).
\end{aligned}
$$
The second term explains why an estimate on the vacuum alone would not
suffice: its input changes with $r$. The scalar phase cancels in this
comparison. This $O(r)$ estimate supplies smooth approximants for actual
insertions and is never divided by $r$. The singular echo instead uses the
more accurate difference derived next.
[OpenAI (2026b), §5, packet invariant and insertion-approximation lemma][q-seed-echo].

**The echo amplifies a controlled difference.**
Consider a forward cost $t/r$, a selective flip of the special
component, the reverse cost $-t/r$, and a selective undoing of that
flip. The packet invariant transports the special component so that
its contribution to the insertion changes sign. For the actual
insertion directions before and after the first flip, it gives
$$
v'_r-v_r=-2r\,w_r+o(r).
$$
The Weyl product of the two costs is formed with these actual
directions before making any approximation. Their large centres
cancel, leaving a bounded displacement
$$
\frac tr(v'_r-v_r)=-2t\,w_r+o(1).
$$
Approximating each direction by a low component with merely $O(r)$
error before this subtraction would leave an uncontrolled order-one
error. The remaining scalar Weyl phase is independent of the root
spin and cancels in insertions, so it need not itself converge. The
second flip restores the transported special component so that the
same construction can be used again. Mixers receive analogous
selective compensations. [OpenAI (2026b), §5, packet invariant,
insertion approximation and “One echo”][q-seed-echo].

In the limit, $\ell_r$ and $w_r$ become two orthogonal external real
Gaussian modes at each root. The initialization uses $q(\ell)$ and
every simulated seed gate uses the same $q(w)$: choosing
$t=\sigma/2$ implements $e^{-i\sigma Zq(w)}$. Different vertices
have independent local copies. This is the required joint law of the
seeded word. A finite induction uses the Weyl identity
$$
\langle W(f)\Omega,W(g)\Omega\rangle
 =\exp\!\left(-\tfrac12\|g-f\|^2
                  +i\operatorname{Im}\langle f,g\rangle\right)
$$
to match the relevant finite Gram matrices. Escaping high directions
supply precisely the vanishing pairings between the new external
modes and endogenous child directions.
[OpenAI (2026b), §5, external-mode limit][q-seed-gram].

<a id="seed-energy-pairing"></a>

**The energy pairing after seed removal.** Matching the seed law and its
finite Gram matrices is not yet enough. For a centred low insertion
$h=\sum_j s_j\otimes W(f_j)\Omega$, the one-particle coefficient of
$W(f_j)\Omega$ is $i e^{-\|f_j\|^2/2}f_j$. Hence
$$
\langle h,\mathsf Lh\rangle
 =-i\sum_j\langle s_j,+\rangle e^{-\|f_j\|^2/2}
                      \langle f_j,h\rangle.
$$
In the seeded representation, $\mathsf L$ places the root-seed factor
in its vacuum and creates $h$ only in the endogenous child space.
External components therefore contribute zero to the last pairing.
The high-direction estimate yields the same zeros in the actual
limit. The actual insertion differs from its low insertion by $O(r)$,
so energy continuity completes the comparison. Matching Gram
matrices without this one-particle calculation would not identify
$\mathsf L$ or its energy. [OpenAI (2026b), §5, “Finite Gram
induction,” including the energy calculation][q-seed-gram].

**Why the selective rotations are available at every required stage.**
The limiting projection of each special centre must be nonzero. The
source imposes finitely many visibility conditions on the angles of
the formal seeded word. At each relevant standard stage they have
form
$$
-2\operatorname{Im}\langle v_Z^{\rm form},
        -2bz\,e+yD^{\rm form}\rangle\ne0,
\qquad z,y\in\{\pm1\},
$$
where $e=Z\mathrm{vac}$ and $D^{\rm form}$ accumulates echoes whose
reverse costs have already occurred. These functions are real
analytic by the finite Weyl formulas. Setting all target angles to
zero, while retaining initialization, gives
$4bz e^{-8b^2}\ne0$, so none is identically zero. An arbitrarily small
angle perturbation avoids their finite union of zero sets and the
zero echo-time hyperplanes. The huge-cost stages need a separate
transverse-projection estimate, after which their limiting projection pairs
also have a nonzero coordinate.

Finite pulse lists can now be chosen
successively from limits already established at earlier stages.
No list is chosen from a limit that assumes its own successful
implementation. [OpenAI (2026b), §5, huge-stage transverse projection,
visibility and construction order][q-seed-visibility].

The choices are therefore nested. We make $b$ small to control
initialization, perturb the fixed target angles to ensure visibility and
choose all finite probe lists. Finally we choose $r$ small. A value of
$r$ fixes every large angle and every finite product. The theorem
allows those angles to be large or negative. It makes no claim of
uniform bounds as the target word length grows or as the tolerance
tends to zero. [OpenAI (2026b), §5, proof of seed removal][q-seed-visibility].

<!-- /passage:qaoa-seed -->

<!-- passage:qaoa-transfer -->

<a id="fixed-word-transfer"></a>

### 8.5 Transfer of the resulting ordinary word

After all these choices there is one finite ordinary word $W$, fixed
independently of degree, size and disorder. On the $(D+1)$-regular
tree use
$C_D=D^{-1/2}\sum_{\{i,j\}\in E}Z_iZ_j$, counting each edge once.
The edge expectation is defined on the finite light cone described in
§8.1. The source applies the fixed-word transfer theorem
$$
v_W=\lim_{D\to\infty}\frac{\sqrt D}{2}
 \langle+|W_D^*Z_iZ_jW_D|+\rangle.
$$
The [full preprint of Basso et al., arXiv:2110.14206v3](https://arxiv.org/abs/2110.14206v3),
Theorem 1, equates the limiting tree recursion with the infinite-size
SK recursion for every fixed depth and fixed angle vector. Its proof
is in §6, especially (6.6)–(6.31), with the symmetry statement proved
in Appendix B and the iteration stabilisation in Appendix A.4. The
shorter proceedings edition states the theorem but refers to this
full preprint for the proof.

Basso et al. use a negative tree cost in (2.3). Let
$V=\prod_{i\in A}X_i$ on one bipartition of the finite light cone.
Every edge has exactly one endpoint in $A$, hence
$$
VZ_iZ_jV=-Z_iZ_j,\qquad VB_DV=B_D,\qquad
V|+\rangle=|+\rangle.
$$
Consequently, if $C_D^-=-C_D$, then $VW(C_D^-,B_D)V=W(C_D,B_D)$
for the same fixed word and angles, while its measured edge expectation
changes sign. This converts the negative cut convention in their (2.6)
to the positive energy above.
Their SK cost (6.1) already has the positive convention used here.
The result applies after seed removal and does not directly transfer an
arbitrary seeded or controlled word. It also supplies no joint
limit in which depth or angles change with $D$ or $n$.
[OpenAI (2026b), §2, “Fixed-word transfer”][q-transfer].

For an ordinary cost or mixer the root updates are, respectively,
$$
u\longmapsto e^{-itZq(v_Z)}u,
\qquad
u\longmapsto e^{-i\beta X}u,
\qquad v_Z=u^*Zu\,\mathrm{vac}.
$$
The source proves smooth convergence for these updates, their
adjoints and fixed angle derivatives. The resulting occupation
bounds supply the hypothesis of the root-shift energy proposition
in §8.1. Thus its energy is the scaled edge limit required for
transfer. Finally merge consecutive gates of the same type and
insert zero-angle gates where needed to express $W$ in alternating
QAOA form. [OpenAI (2026b), §2, “Ordinary root updates”][q-ordinary-updates].

<!-- /passage:qaoa-transfer -->

<!-- passage:qaoa-assembly -->

<a id="qaoa-assembly-detail"></a>

### 8.6 Fixing the word before the thermodynamic limit

<a id="finite-formula-energy-error"></a>

**Approximating a specified finite observable.** First fix $T,N$ and the
positive finite normalizers in §7. The rooted construction in §8.1 has
energy exactly $S_{T,N}$. Let $z$ be a smoothed version of its rounded
spin $F_N$, let $m=\sin(2\theta)$ be the bounded formula's spin output,
and let $m^{\mathrm{col}}$ be its color-excluded approximation. Write
$$
\delta_s=\|z-F_N\|_2,\qquad
\delta_b=\|m-z\|_2,\qquad
\delta_c=\|m^{\mathrm{col}}-m\|_2.
$$
These measure errors in the spin outputs themselves. The rooted
shift is a contraction, so
$$
|\mathcal E(f)-\mathcal E(g)|
\le(\|f\|_2+\|g\|_2)\|f-g\|_2.
$$
All these outputs have norm at most one. Apply synthesis with tolerance
$\eta_s$ and total helper probability $p_H$, then remove the seeds with
energy tolerance $\eta_r$. The source's absolute synthesis bound and
seed-removal theorem, followed by fixed-word transfer, give one ordinary
word $W$ with
$$
|v(W)-S_{T,N}|
\le2(\delta_s+\delta_b+\delta_c)
       +2\eta_s+2\sqrt{p_H}+\eta_r.
\tag{8.2}
$$
The bounded formula is chosen before the colors and helper law. The
excluded receiver and ancestor colors prevent a dependency from returning
through a traversed branch. Reserved helper colors are absent from every
dependency, and the output is zero on helpers. Color refinement and small
positive helper probabilities jointly control $\delta_c$ and $p_H$.
Only then is synthesis performed with that law fixed, using the same
sitewise Gaussian seed at every seed gate. Each term in (8.2) can thus be
made smaller than its allotted tolerance in this order.
[OpenAI (2026b), Lemmas 3.1–3.4][q-classical];
[synthesis energy bound][q-synthesis-statement];
[seed removal][q-seed-statement].

This gives an absolute approximation of each specified finite
$S_{T,N}$ in the construction, whether or not it is already near $P_*$.
It preserves the energy and makes no claim about the entire classical
output law. The cutoff and mesh choices then give near-optimality.

We allocate the errors after choosing the desired $\varepsilon>0$, and
apply the
classical near-optimality proposition with the two tolerances
$\varepsilon/4$ and $\eta_H=(\varepsilon/8)^2$. It supplies one
formula and its helper law together, with classical loss below
$\varepsilon/4$ and $0<p_H<\eta_H$. In its proof the uncoloured
formula is fixed first, then the small positive helper probabilities
and the excluded-colour approximation are selected. Thus changing a
helper law does not silently preserve an already fixed formula. With
$\eta=\varepsilon/8$, synthesis gives
$$
\mathcal E(v)\ge\mathcal E(m)-2\eta-2\sqrt{p_H}
 >P_*-3\varepsilon/4.
$$
Seed removal uses less than $\varepsilon/4$. The resulting fixed
ordinary word therefore satisfies $v_W>P_*-\varepsilon$ after the
fixed-word transfer. Its thermodynamic limit is taken only after
all approximation choices have been made. Since $Q_p\le P_*$ and
zero padding makes $Q_p$ nondecreasing, this proves the source's
thermodynamic-first conclusion from its stated analytic inputs.
[OpenAI (2026b), §6][q-assembly].

The choices come in a fixed order. We choose the accuracy, then the
minimizer and a fixed $T<1$, then the finite scalar mesh. Next come the
smoothing and bounded approximations, the finite types and positive
helper masses, and the exact helper contractions and finite commutator
products, after which one seeded word is fixed. We then choose the
initialization size, a generic angle perturbation, finite selective pulse
lists and a sufficiently small $r$, which fixes one ordinary word. Only
then do we take its $n\to\infty$ limit.

The degree limits identify the rooted calculus and the transferred value
for each already fixed word, and they never allow that word to change with
the finite instance.

The reconstruction makes the source's circuit mechanisms explicit and
keeps their evidence classes separate. The scalar construction
and its formal interfaces in §§6–7 and 12 do not by themselves establish the
weighted unitary calculus, simultaneous helper contraction or seed
removal. Those analytic theorems are additional inputs from the QAOA
manuscript, with the domain, commutation and uniformity hypotheses
specified above. No quantitative depth bound, finite-size rate,
efficient angle-selection procedure, or independent formal
verification of this entire circuit construction follows from the
selected formal statements.

This completes the main support-to-circuit proof.

The later comparisons have separate purposes.
[§11.1](#positive-temperature-route) gives another classical input, and
[§11.2](#fixed-word-probability) adds a second-moment sampling theorem.
[§§11.3–11.5](#depth-quantifiers) compare quantifiers, other models and
later finite-size estimates. None of them is a premise of the proof just
completed.
<!-- /passage:qaoa-assembly -->

<a id="further-support-consequences"></a>

## 9. Density and finite-disorder consequences

<!-- passage:density -->
<a id="smooth-density-consequence"></a>

### 9.1 Smoothness of the minimizing coefficient

A density for the Stieltjes measure needs an argument beyond full support. In the source manuscript it uses the PDE again, after the gap contradiction has established
$\mathbb E a(t,X_t)^2=1$. We reconstruct this argument because it gives a further conclusion of the original paper, outside the selected full-support formal statement.

Put $f_k=\partial_x^k\Phi_\gamma$, so $f_1=u$ and $f_2=a$. All formulas in this subsection are on a fixed compact time strip, where every finite collection of spatial derivatives is bounded. Differentiating the PDE and applying Itô's formula along the associated diffusion gives
$$
df_k(t,X_t)=
-\frac{\gamma(t)}2\sum_{j=1}^{k-1}\binom{k}{j}
 f_{j+1}(t,X_t)f_{k-j+1}(t,X_t)\,dt
+f_{k+1}(t,X_t)\,dB_t.
$$
The sum is empty when $k=1$. This identity needs no derivative of $\gamma$.

For an initially nonsmooth coefficient, we apply it to finite-step approximations and pass to the integrated identity. There the coefficient converges in $L^1$, the spatial factors converge uniformly, and the stochastic integrals converge by the Itô isometry. [OpenAI full-support manuscript, smooth-density proof, lines 35–50][fs-density].

Set $b=f_3$. The case $k=2$ gives
$da=-\gamma a^2dt+b\,dB_t$, and consequently
$$
0=\frac d{dt}\mathbb E a^2
 =\mathbb E b^2-2\gamma(t)\mathbb E a^3
\quad\text{for almost every }t<1.
$$
Convexity gives $a\geq0$, and the already established second moment gives
$\mathbb E a^3\geq(\mathbb E a^2)^{3/2}=1$. Thus the quotient
$$
\widetilde\gamma(t)=\frac{\mathbb E b(t,X_t)^2}
                         {2\mathbb E a(t,X_t)^3}
$$
is well defined and continuous, before any continuity of $\gamma$ is known. It equals $\gamma$ almost everywhere.

Approaching an arbitrary $t<1$ from the right through the full-measure equality set, continuity of $\widetilde\gamma$ and right continuity of $\gamma$ give equality at that $t$ as well. This proves continuity of $\gamma$. At zero, $X_0=0$, and evenness of $\Phi(0,\cdot)$ gives $b(0,0)=0$; hence $\gamma(0)=0$. [OpenAI full-support manuscript, lines 52–66][fs-density]; [Chen, Proposition 4.6, pp. 25–26][chen26].

The bootstrap must treat all polynomial moments simultaneously. For a polynomial $P$ in finitely many variables, write
$$
M_P(t)=\mathbb E P(f_1(t,X_t),\ldots,f_m(t,X_t)).
$$
The preceding stochastic identities and Itô's product rule provide polynomials $Q_P,R_P$, involving only finitely many additional derivatives, such that
$$
M_P(t)-M_P(0)=\int_0^t\bigl(M_{Q_P}(s)+\gamma(s)M_{R_P}(s)\bigr)\,ds.
$$
Explicitly, if
$$
D_k(y)=-\tfrac12\sum_{j=1}^{k-1}\binom{k}{j}y_{j+1}y_{k-j+1},
$$
then
$$
Q_P=\tfrac12\sum_{k,l}\partial_{kl}P\,y_{k+1}y_{l+1},
\qquad R_P=\sum_k\partial_kP\,D_k.
$$
All these moments are continuous to begin with. Once $\gamma$ is continuous, the integral formula makes every moment $C^1$, and the quotient formula makes $\gamma$ itself $C^1$. If all moments and $\gamma$ are $C^r$, the same argument makes them $C^{r+1}$. Induction proves $\gamma\in C^\infty([0,1))$, with derivatives at zero taken from the right. Since $\gamma(0)=0$, there is no atom at zero. Its Stieltjes measure is $\gamma'(t)\,dt$, and monotonicity gives $\gamma'\geq0$. The source credits this polynomial-moment method to Auffinger–Chen and to Chen's zero-temperature treatment. [OpenAI full-support manuscript, lines 68–101][fs-density]; [Auffinger and Chen, proof of Theorem 2(ii), p. 16][ac-properties]; [Chen, Appendix B.6, p. 38][chen26].

The exact arXiv edition of Auffinger–Chen's polynomial-moment proof has
a local display discrepancy: p. 16 of arXiv:1303.3573v1 prints a square
in $F_3$ below (40), whereas Proposition 3, p. 9, has the cube.
The identity $d(a^2)=(-2\gamma a^3+b^2)dt+2ab\,dB$ requires the
cube used above, and the moment induction itself is unchanged.

<!-- /passage:density -->

<!-- passage:density-positive -->

<a id="positive-density"></a>

### 9.2 Strict positivity of the density away from zero

Strict positivity is a further consequence of full support, and neither full support nor §6 uses it.

A smooth increasing function can have zero derivative at an interior point. For example,
$g(t)=(t-\tfrac12)^3+\tfrac18$ on $[0,1)$ is nonnegative and strictly increasing, and $dg$ has full support, although $g'(\tfrac12)=0$. The example is no minimizing coefficient, and it refutes only the inference from strict increase to a positive derivative. For the minimizing density we need a further estimate.

The smoothness just proved permits (4.2), now with $q(t)=t$.
Writing $b=\Phi_{xxx}$ and $d=\Phi_{xxxx}$, it becomes
$$
2\gamma'(t)\mathbb E a^3=\mathcal C_\gamma(t).
$$
Fix $t\in(0,1)$ and set $c=\gamma(t)>0$, using full support and
$\gamma(0)=0$. The coefficient is now smooth, so Proposition 4.1 applies
at this time to the same diffusion. Its proof uses nonminimizing local
constant replacements, with compact-strip bounds independent of the
widths of the inserted plateaus. Combining (4.1) with the preceding
identity gives
$$
2\gamma'(t)\mathbb E a^3
\geq\frac{\gamma(t)^2}{1000}\mathbb E a^4>0,
\qquad 0<t<1.
$$
The final strict inequality uses $\mathbb E a^4\geq(\mathbb E a^2)^2=1$. Cauchy–Schwarz also gives
$\mathbb E a^3\le\sqrt{\mathbb E a^2\mathbb E a^4}
=\sqrt{\mathbb E a^4}\le\mathbb E a^4$. Therefore the same estimate proves
the quantitative ordinary consequence
$$
\gamma'(t)\ge\frac{\gamma(t)^2}{2000}>0\quad(0<t<1),\qquad
\mu_\gamma((r,s])\ge\frac{s-r}{2000}\gamma(r)^2
\quad(0<r<s<1).
$$
The second inequality follows by integration and monotonicity. Its lower
bound depends on this minimizer and the left endpoint, so it gives no
uniform mass estimate for a family of finite-temperature measures.

At the left endpoint the argument proves only $\gamma'(0)\geq0$. It makes no claim that the minimizing coefficient itself is constant on a neighbourhood, and it passes none of the approximants' third time derivatives to the limit. Only a spatial-derivative expectation is transferred. [OpenAI full-support manuscript, lines 102–152][fs-density].

The density result is a separate addition to the scalar approximation in
[§7](#finite-gaussian-correlations). The next consequence concerns the
geometry of near-ground spin configurations.
<!-- /passage:density-positive -->

<a id="near-ground-overlaps"></a>

### 9.3 Near-ground configurations at a prescribed overlap

<!-- passage:overlap-pairs -->
We choose one minimizer supplied by the attained Parisi formula and apply
the full-support theorem to that same coefficient. The external landscape
theorem below then applies at every point of its Stieltjes support. For every fixed $r\in[0,1)$ and $\varepsilon,\eta>0$,
there is a constant $K=K(r,\varepsilon,\eta)>0$, independent of $n$, such that
$$
\mathbb P\left(\exists\sigma,\tau\in\{-1,1\}^n:
\left|\frac1n\sum_{i=1}^n\sigma_i\tau_i-r\right|<\varepsilon,
\quad \frac{H_n(\sigma)}n,\frac{H_n(\tau)}n\ge P_*-\eta\right)
\ge1-Ke^{-n/K},\qquad n\ge1.
$$
The probability is over the SK couplings, and the configurations may depend
on them. The display specializes [Auffinger–Chen–Zeng, Proposition 1,
equation (5), p. 5][acz] to the full-support minimizer above, since their
theorem applies at each point of its Stieltjes support.

Their proof, on pp. 15–16,
uses the coupled free energy constrained to a neighbourhood of that
overlap, passes to the coupled maximum and applies Gaussian concentration.
The coupled maximum approaches $2P_*$, while the unrestricted maximum
approaches $P_*$. These two facts force both members of a maximizing pair
to be near ground energy.

The covariance normalization needs one small adjustment. Writing
$R=n^{-1}\sum_i\sigma_i\tau_i$, our Hamiltonian has covariance
$nR^2/2-1/2$. Adjoin a common independent Gaussian $C$ of variance $1/2$.
Then $\widetilde H_n(\sigma)=H_n(\sigma)+C$ has covariance
$n\xi(R)$ with $\xi(t)=t^2/2$, as in the cited mixed-spin theorem.
Its expected maximum equals that of $H_n$, so its ground-state limit is
the same $P_*$. Apply the theorem with energy tolerance $\eta/2$.
On $|C|\le\eta n/2$, the resulting pair also satisfies the displayed
energy inequalities for $H_n$. The omitted event has probability at most
$2e^{-\eta^2n^2/4}$, which can be absorbed into $Ke^{-n/K}$ by enlarging
$K$. Since the conclusion for $H_n$ depends only on its couplings,
this also removes the auxiliary Gaussian from the probability statement.

The pointwise result also gives a simultaneous statement at any fixed
positive resolution. Given $\varepsilon,\eta>0$, choose finitely many
centres $r_j\in[0,1)$ whose open $\varepsilon/2$-neighbourhoods cover
$[0,1]$. Apply the preceding result at each centre with overlap tolerance
$\varepsilon/2$ and energy tolerance $\eta$. A finite union bound gives
constants $C,c>0$, depending only on these tolerances, such that
$$
\mathbb P\!\left(\forall r\in[-1,1]\ \exists\sigma,\tau\in\{-1,1\}^n:
 |R(\sigma,\tau)-r|<\varepsilon,\quad
 H_n(\sigma)/n,H_n(\tau)/n\ge P_*-\eta\right)
 \ge1-Ce^{-cn}.
\tag{9.1}
$$
Indeed, on the event where every centre has a suitable pair, choose a
nearest centre and use the triangle inequality. For negative targets,
reverse one spin configuration: the pure two-spin Hamiltonian preserves
its energy and changes the overlap's sign. The finite net covers the
endpoints as well, with centres strictly below one.

<a id="exact-finite-maximizers"></a>

At a fixed finite size the picture is different. If $\sigma,\tau$
differ at $k$ sites, precisely the $k(n-k)$ edges crossing that set change
their spin product. Their independent Gaussian contributions give
$$
\operatorname{Var}(H_n(\sigma)-H_n(\tau))=\frac{4k(n-k)}n.
$$
For $0<k<n$ this is positive, so the two energies tie with probability
zero. A finite union over pairs, together with
$H_n(-\sigma)=H_n(\sigma)$, shows that almost surely the exact maximum
has one maximizing pair $\{\sigma_*,-\sigma_*\}$. The overlap conclusion
above concerns nearly optimal pairs at intermediate separations. Its
threshold $P_*-\eta$ is deterministic, and setting $\eta=0$ would not turn
it into exact maximization of a finite sample. This elementary deduction
uses only the displayed Gaussian Hamiltonian and is separate from the
external asymptotic landscape theorem.

The overlap deduction uses fixed tolerances and target-dependent pairs. It gives
neither exact arbitrary real overlaps at finite $n$, a typical-pair law,
a search procedure, nor a guarantee for tolerances shrinking with $n$.
It uses the external landscape theorem, full support and attainment, and
lies outside the two selected formal statements in §12. The separate
algorithmic consequence is in §9.4.
<!-- /passage:overlap-pairs -->

<!-- passage:algorithm -->
<a id="classical-algorithm-consequence"></a>

### 9.4 A classical optimization algorithm

The scalar Gaussian construction is not itself an algorithm on the matrix of SK couplings. We obtain a finite-disorder algorithm from the theorem of El Alaoui, Montanari and Sellke. In their cited preprint, Assumption 2 asks for a strictly increasing minimizer of the zero-temperature Parisi functional on $\mathcal U$. The support theorem gives this property, and the imported attainment argument supplies the minimizer.

Their Corollary 2.2 then returns, for each fixed $\varepsilon>0$, a randomized spin configuration within $\varepsilon n$ of the finite-instance maximum with probability tending to one. Its operation count is the one in the displayed Theorem 3 on p. 7, which the preprint's running references call “Theorem 3.2.” [El Alaoui, Montanari and Sellke, arXiv:2001.00904v1, pp. 7–8][eams20].

To match that theorem's Gaussian input exactly, form a symmetric matrix with
$$
A_{ij}=J_{ij}/\sqrt n\quad(i<j),\qquad
A_{ii}=\sqrt{2/n}\,G_i,
$$
where the $G_i$ are additional independent standard Gaussians. For $\widehat H_n(x)=\tfrac12x^{\mathsf T}Ax$,
$$
\mathbb E\widehat H_n(\sigma)\widehat H_n(\tau)
=\frac{(\sigma^{\mathsf T}\tau)^2}{2n},
\qquad
\widehat H_n(\sigma)=H_n(\sigma)+\frac1{\sqrt{2n}}\sum_iG_i.
$$
The final term is the same for all spin configurations. Therefore the maximizing spin and every optimization gap are unchanged. The extra diagonal randomness is included in the algorithm's randomness. The input convention thus matches, and the target optimization problem is unchanged. [OpenAI full-support manuscript, algorithm application, lines 30–47][fs-algorithm].

The imported operation count is
$C(\varepsilon)(\chi+n)+n\chi_1$, where $\chi$ is the cost of evaluating the full gradient and $\chi_1$ is the cost of one coordinate of the distinct-index gradient. Here $\nabla\widehat H_n(x)=Ax$ costs $O(n^2)$, and one coordinate of $\nabla H_n(x)$ costs $O(n)$. The resulting count is $C(\varepsilon)n^2$, with $C(\varepsilon)$ independent of $n$.

Real sums and multiplications have unit cost in the cited model, and the scalar functions and normalizers prescribed for the chosen accuracy are part of that construction. This statement supplies neither a bound for their numerical preprocessing nor a finite-precision bit-complexity bound. [El Alaoui, Montanari and Sellke, Remark 2.1 and Theorem 3, p. 7][eams20]; [OpenAI full-support manuscript, lines 1–11 and 49–59][fs-algorithm].

A distinct classical construction is [Jekel, Sandhu and Shi,
*Potential Hessian Ascent*, version 3, Assumption 1.4 and
Theorem 1.5](https://arxiv.org/abs/2408.02360v3).
Its hypothesis is finite-temperature interval support
$\operatorname{supp}\mu_\beta=[0,q_\beta^*]$, with $q_\beta^*\to1$.
Its inputs include oracles for the Parisi PDE, its derivatives and the
minimizer's distribution function. For each fixed accuracy the stated
runtime is polynomial in $n$, with an exponential dependence on inverse
accuracy, and it excludes the cost of implementing those oracles.

Its Ginibre
Hamiltonian has covariance $nR^2$: on spins it has the law
$\sqrt2 H_n$ plus a Gaussian term independent of the configuration. Thus its energy
and accuracy must be divided by $\sqrt2$ to compare with this paper's
normalization. This is a separate algorithm theorem with its own
analytic inputs, and it gives no new reading of the finite Gaussian spin
in §7.

The consequence is classical. It supplies no efficient construction of QAOA angles, no quantum depth bound and no clause of either selected formal interface. The proof of the separate QAOA statement still needs its own rooted energy, synthesis, seed-removal and transfer arguments.
<!-- /passage:algorithm -->

<a id="independent-scalar-comparison"></a>

## 10. An independent scalar comparison

<!-- passage:formal-contact -->
<a id="formal-scalar-contact"></a>

### 10.1 Contact from convexity on constant intervals

A second route obtains $q=t$ from the variational inequalities. It uses
convexity on intervals where the tail potential is positive, without first
proving full Stieltjes support. The scalar implementation uses this
argument, which is independent of the formal full-support theorem.

In the scalar model put $q(t)=\mathbb E[u(t,X_t)^2]$,
$f(t)=q(t)-t$ and $G(t)=\int_t^1f(s)\,ds$. The source first obtains
$G\ge0$, $G(1)=0$, $q(0)=0$ and $G'=-f$. Compact-mass variations
also show that on any interval where $G>0$ the coefficient is constant:
if its cumulative mass increased between two points there, integration
of the strictly positive $G$ against that positive increment would
contradict the mass-removal variational inequality. This is the content
of the [scalar contact-integral argument][scalar-contact-integral].

Suppose now that $G(x)>0$ for some $0<x<1$. Choose the component
$(a,b)$ of this positivity set containing $x$, allowing $a=0$ or
$b=1$. Its coefficient is constant, say $c\ge0$, and $G(b)=0$.
At the left endpoint, $f(a)=0$: this is the contact derivative when
$a>0$, and $q(0)=0$ when $a=0$.

If $c>0$, one also has $G(a)=0$. The only extra case is $a=0$:
right continuity gives $\gamma(0)=c>0$, so the mass-removal condition
at that initial atom forces $G(0)=0$. The mean-value theorem on
$[a,x]$ and $[x,b]$ now provides points $r<s$ with
$f(r)<0<f(s)$. There is a later point $z\le b$, with $z<1$, for
which $f(z)\le f(s)$. For $b<1$ take $z=b$, since contact gives
$f(b)=0$. For $b=1$, choose $z$ sufficiently close to one and use
$f(z)=q(z)-z\le1-z<f(s)$.

The three intervals $[a,r]$, $[r,s]$ and $[s,z]$ therefore have
secant slopes for $f$ of signs respectively negative, positive and
nonpositive. Applying the mean-value theorem once more gives three
ordered points where $f'$ has those signs. This is impossible when
$f'=q'-1$ is convex: its middle value cannot exceed the larger of
its two exterior values. The scalar source supplies precisely this
nonstrict convexity on a constant interval. See
[the positive-hump argument and its endpoint condition][scalar-hump]
and [the curvature-moment convexity lemma][scalar-convexity].

If $c=0$, the source instead gives convexity of $f=q-t$. A mean-value
argument between $x$ and $b$ gives $f(s)>0$, and the same right-endpoint
argument produces $z>s$ with $f(z)\le f(s)$. Since $f(a)=0$,
$$
\frac{f(s)-f(a)}{s-a}>0\ge
\frac{f(z)-f(s)}{z-s},
$$
contradicting monotonicity of adjacent secant slopes of a convex
function. This case does not need $G(a)=0$. See
[the zero-coefficient hump lemma][scalar-zero-hump].

Thus $G$ cannot have a positive component, and $G\equiv0$ on
$(0,1)$. Differentiation gives $q(t)=t$ there, with the origin supplied
separately. The [scalar assembly][scalar-contact-assembly] then obtains
the unit curvature moment, finite Gaussian approximation and terminal
value identities from this contact equality. Nonstrict convexity suffices
for this scalar route. The strict curvature estimate in Sections 4–5
has the further job of excluding gaps in Stieltjes support. These are
source-level proof reconstructions, not a fresh Lean acceptance or a
claim that the two formal models have been composed.

<!-- /passage:formal-contact -->

<!-- passage:formal-lower-overview -->
<a id="independent-scalar-lower-bound"></a>

### 10.2 A universal bound for predictable correlations

The scalar implementation obtains the value equality by two inequalities.
Its finite-spin limit follows from superadditivity and a uniform upper
bound. A finite-profile Gaussian comparison supplies
$P_*\le\mathcal P(\gamma)$ for every admissible coefficient. The reverse
inequality at a minimizer comes from a more general finite-dimensional
fact: predictable orthonormal Gaussian functions paired with any spin
have a shifted correlation sum at most $P_*$. We reconstruct that fact
below. It supplies the inequality that §7's mesh limit consumes, without
invoking the literature's Parisi formula used in §6.

Write $e_n=n^{-1}\mathbb E\max_\sigma H_n(\sigma)$, with the Hamiltonian
of §1. The finite-spin convergence and upper comparison are separate
inputs to this subsection, and their precise suppliers are
[the superadditive limit][scalar-spin-limit] and
[the upper comparison][scalar-upper]. The proof below concerns only the
lower comparison, and it builds neither a finite-disorder algorithm nor a
circuit.

**Proposition 10.1 (predictable Gaussian comparison).** Fix $K<\infty$, independent standard normals $z_0,\ldots,z_K$,
and real functions $f_j=f_j(z_0,\ldots,z_j)$ for $0\le j<K$.
Assume measurability, square integrability and
$\mathbb E[f_if_j]=\delta_{ij}$. For any measurable
$S=S(z_0,\ldots,z_K)\in\{-1,1\}$, the conclusion is
$$
\sum_{j=0}^{K-1}\mathbb E[f_jS]\,\mathbb E[S z_{j+1}]\le P_*.
$$
These assumptions allow nonconstant, dependent $f_j$, and even
$\mathbb E f_j=0$ is unnecessary. Their order of dependence ensures
$\mathbb E[f_i z_{j+1}]=0$ whenever $i\le j$. The next fresh coordinate,
rather than centring of $f_i$, is what will remove the projection terms.
[Scalar implementation, predictable-coefficient theorem][scalar-predictable].
<!-- /passage:formal-lower-overview -->

<!-- passage:gaussian-column-split -->
<a id="gaussian-column-split"></a>

#### 10.2.1 Revealing one Gaussian column

For $w\in\mathbb R^n$, put $Q(w)=w\otimes w/\sqrt2$.
An independent standard Gaussian matrix $G$ gives a centered field
$\langle Q(w),G\rangle$ with covariance
$\langle w,w'\rangle^2/2$. Given a unit vector $v$, define
$$
R_vw=w-\langle v,w\rangle v,\qquad
C_v(w)=\langle v,w\rangle R_vw+
       \frac{\langle v,w\rangle^2}{\sqrt2}v.
$$
Write $\alpha=\langle v,w\rangle$, $\beta=\langle v,w'\rangle$
and $r=\langle R_vw,R_vw'\rangle$. For an independent standard Gaussian
column $Z$ and matrix $G'$, the covariance of
$\langle C_v(w),Z\rangle+\langle Q(R_vw),G'\rangle$ is
$$
\alpha\beta r+\frac12\alpha^2\beta^2+\frac12r^2
=\frac12(r+\alpha\beta)^2
=\frac12\langle w,w'\rangle^2.
$$
Thus the new field has the old joint Gaussian law. The same statement
holds for $v=0$: then $R_v$ is the identity and $C_v=0$.
Allowing that case will avoid assuming empirical linear independence.

For a finite nonempty family of candidates $w_s$ and offsets $d_s$, define
$$
\mathcal V(w,d)=\mathbb E_G\max_s
  \{\langle Q(w_s),G\rangle+d_s\}.
$$
The covariance identity and Fubini give the exact recursion
$$
\mathcal V(w,d)=\mathbb E_Z\mathcal V
 \bigl((R_vw_s)_s,(d_s+\langle C_v(w_s),Z\rangle)_s\bigr).
$$
Also $d_s\le\mathcal V(w,d)$ for each $s$, since the residual Gaussian
term has mean zero. These are the two facts needed for adaptive choices.
The source derives the equality by Gaussian comparison in both directions
and proves the required integrability in
[the one-column decomposition][scalar-column].
<!-- /passage:gaussian-column-split -->

<!-- passage:gaussian-adaptive-lower -->
<a id="adaptive-gaussian-lower"></a>

#### 10.2.2 Adaptive directions and a finite-spin inequality

We choose each zero-or-unit direction measurably from columns already
revealed. Conditional on that history, the current residual candidate
vectors and offsets are fixed, so the preceding recursion applies.
More precisely, conditional on the revealed history, the expected next
continuation value, including its accumulated offsets, equals the current
one. Iteration therefore preserves the initial expected maximum. For any measurable
choice of a candidate from the final history, its accumulated offset is
bounded by the remaining conditional expected maximum. Hence
$$
\mathbb E\,d_{s(\mathrm{history})}^{\rm final}
\le\mathcal V(w,0).
$$
Contraction of each residual, the finite candidate set and Gaussian
column moments justify the expectations. Adaptivity does not turn a
future column into part of the conditioning history.
[Scalar implementation, adaptive recursion][scalar-adaptive].

Take all spins $w_s\in\{-1,1\}^n$ as candidates. The full quadratic
matrix has off-diagonal coefficients $(G_{ij}+G_{ji})/\sqrt2$, which
are independent standard normals, and a centered diagonal contribution
common to every spin. Thus
$$
\mathcal V(\mathrm{spins},0)=\sqrt n\,\mathbb E\max_\sigma H_n(\sigma).
$$
Dividing the offset inequality by $n\sqrt n$ gives a bound by $e_n$.
The implementation proves this normalization through equality of Gaussian
increment distances in [the finite adaptive SK comparison][scalar-finite-adaptive].

To obtain the desired coefficients, take independent rows
$z^{(1)},z^{(2)},\ldots$, each with the law of $(z_0,\ldots,z_K)$,
and put
$$
\widehat g_n=n^{-1/2}(g(z^{(1)}),\ldots,g(z^{(n)})).
$$
Let $v_{n,j}$ be the normalized Gram–Schmidt vector obtained from
$\widehat f_{0,n},\ldots,\widehat f_{j,n}$, where a zero remainder is
normalized to zero. Start the adaptive procedure with a zero direction
to reveal column $z_0$. After seeing columns through $z_j$,
$v_{n,j}$ is available, and the next column is $z_{j+1}$.
This explains both the extra column and the shift.

Let $R_{n,j}$ project away $v_{n,0},\ldots,v_{n,j-1}$ and set
$$
E_{n,j}=\langle C_{v_{n,j}}(R_{n,j}\widehat S_n),
                       \widehat z_{j+1,n}\rangle.
$$
The raw spin and fresh column are each $\sqrt n$ times their empirical
versions, and $C_v$ is quadratic in its input. Consequently their revealed
energy is exactly $n\sqrt n E_{n,j}$. Transposing the array between
independent rows and independent columns preserves its Gaussian law.
The finite comparison therefore gives, for every $n>0$,
$$
\mathbb E\sum_{j<K}E_{n,j}\le e_n.
$$
This is an actual choice of the adaptive directions and final rowwise
spin, not an uninstantiated conditional inequality.
[Empirical directions and normalization][scalar-empirical-directions];
[row–column transposition and finite bound][scalar-transpose].
<!-- /passage:gaussian-adaptive-lower -->

<!-- passage:gaussian-empirical-limit -->
<a id="empirical-gaussian-limit"></a>

#### 10.2.3 Orthogonalization preserves the limiting correlations

We keep $K$ fixed while $n\to\infty$. The strong law applies to the
finitely many needed products, which are integrable by Cauchy–Schwarz.
The empirical Gram matrix tends to the identity. Inductively in
Gram–Schmidt, every projection coefficient tends to zero and every
remaining norm tends to one; hence
$\|v_{n,j}-\widehat f_{j,n}\|\to0$ almost surely. In particular,
$$
\langle v_{n,j},\widehat S_n\rangle\to a_j:=\mathbb E[f_jS],
\qquad
\langle\widehat S_n,\widehat z_{j+1,n}\rangle
   \to b_{j+1}:=\mathbb E[S z_{j+1}],
$$
while $\langle v_{n,i},\widehat z_{j+1,n}\rangle\to0$ for $i\le j$.
The last limit uses precisely the fresh-coordinate cancellation stated
at the start of §10.2. [Gram–Schmidt limits][scalar-gram-schmidt].

Expand the revealed energy before taking its limit:
$$
\begin{aligned}
E_{n,j}={}&\langle v_{n,j},\widehat S_n\rangle
 \langle R_{n,j+1}\widehat S_n,\widehat z_{j+1,n}\rangle\\
&+\frac{\langle v_{n,j},\widehat S_n\rangle^2}{\sqrt2}
 \langle v_{n,j},\widehat z_{j+1,n}\rangle.
\end{aligned}
$$
All terms introduced by projecting away earlier directions vanish,
as does the second line. The remaining product tends to $a_jb_{j+1}$.
Thus $\sum_{j<K}E_{n,j}\to\sum_{j<K}a_jb_{j+1}$ almost surely.
This establishes the random-variable limit, but an expectation bound
still requires the next step.
[Scalar implementation, empirical energy limit][scalar-empirical-limit].
<!-- /passage:gaussian-empirical-limit -->

<!-- passage:gaussian-expectation-limit -->
<a id="gaussian-expectation-limit"></a>

#### 10.2.4 Passing the expectation bound and returning to the value

Since $S^2=1$, $\|\widehat S_n\|=1$ for $n>0$, and every residual
projection contracts. The column formula gives
$|E_{n,j}|\le1+\|\widehat z_{j+1,n}\|^2$.
Define
$$
c_n=\sum_{j<K}(1+\|\widehat z_{j+1,n}\|^2).
$$
Then $\sum_jE_{n,j}+c_n\ge0$, $\mathbb E c_n=2K$ exactly, and
$c_n\to2K$ almost surely. Fatou's inequality, the finite bound and
$e_n\to P_*$ imply
$$
\sum_{j<K}a_jb_{j+1}+2K
\le\liminf_n\mathbb E\bigl[\sum_{j<K}E_{n,j}+c_n\bigr]
\le P_*+2K.
$$
Cancelling $2K$ proves the claimed predictable-coefficient inequality.
The source's constant-limit helper establishes the same nonnegative
comparison by clipping at the limiting constant and applying dominated
convergence. It shifts to dimensions $n+1$ to avoid dimension zero.
It does not apply Fatou directly to the signed energies or presume
uniform integrability. [Column bound and correction][scalar-empirical-bound];
[expectation limit and final theorem][scalar-predictable].

For the Euler construction in §7, take $K=N$, relabel its $Z_{j+1}$ as
$z_j$, and set $f_j=g_{j+1}$. Each normalized increment depends only
on $z_0,\ldots,z_j$ and the increments are orthonormal when their raw
second moments are positive. The rounded sign is spin-valued almost
surely: its tie event is null by the independent last coordinate.
The scalar implementation uses a measurable Bool spin equal to that
sign almost surely. It follows that
$$
S_{T,N}=\sum_{j=1}^N A_jB_{j+1}\le P_*.
$$
This finite inequality assumes admissibility, $0\le T<1$, $N>0$ and
positive active second moments, but no minimizer or diffusion.
[Literal scalar specialization][scalar-literal-lower].

At a minimizer, the separate contact and diffusion arguments establish
the fixed-strip inputs and eventual positivity. First take the mesh
limit to obtain $I(T)\le P_*$, then let $T\uparrow1$. The scalar
terminal assembly identifies that last limit with $\mathcal P(\gamma)$.
Together with $P_*\le\mathcal P(\gamma)$ this proves equality.
The third scalar endpoint constructs its own diffusion before making
this application. These sources therefore supply an independent route
to the equality, alongside the imported Parisi formula in §6.
[Contact assembly][scalar-contact-assembly]. The inspection here does
not change the research replay outcome in §13.
<!-- /passage:gaussian-expectation-limit -->

<!-- passage:gaussian-sanity-check -->

<a id="zero-coefficient-example"></a>

### 10.3 The zero coefficient

The admissible coefficient $\gamma_0=0$ makes the use of minimality in
the normalization step explicit. Its diffusion is $X_t=B_t$, and the
[zero-coefficient heat formula][fs-reg] gives

$$
u_0(t,x)=2F_{\mathrm G}\!\left(\frac{x}{\sqrt{1-t}}\right)-1,
\qquad
a_0(t,x)=\sqrt{\frac2\pi}\frac{e^{-x^2/(2(1-t))}}{\sqrt{1-t}}.
$$

Since $\mathbb E e^{-\lambda B_t^2}=(1+2\lambda t)^{-1/2}$,

$$
\mathbb E a_0(t,B_t)=\sqrt{\frac2\pi},\qquad
\mathbb E a_0(t,B_t)^2=\frac2{\pi\sqrt{1-t^2}},\qquad
q_0(t)=\frac2\pi\arcsin t.
$$

The last identity follows by integrating $q_0'=\mathbb E a_0^2$
from $q_0(0)=0$. This heat and late-step variation argument is the
zero-field specialization of [Chen, Handschy and Lerman, Theorem 6
and its proof, arXiv:1609.04368v4, pp. 13–16](https://arxiv.org/abs/1609.04368v4).
Strict convexity of $\arcsin$ gives $q_0(t)<t$
for $0<t<1$. Thus $G_0(s)=\int_s^1(q_0(t)-t)\,dt<0$ for
every $s<1$. Inserting an atom at $s$ is admissible, and its
[first variation][fs-var] is $G_0(s)/2<0$. In particular, the derivative for inserting
mass at zero is $1/4-1/\pi<0$. Hence this coefficient does not
minimize.

Here the Euler scheme is exact: $Y_j=B_{t_j}$ under the Brownian
coupling. Nevertheless its normalizers are

$$
c_j=\sqrt{\frac\pi2}\,(1-t_{j-1}^2)^{1/4},
$$

so they approximate this nonconstant function of time, rather than
converging uniformly to one. Already $c_1=\sqrt{\pi/2}$ for every
mesh. The directions $g_j$ remain orthonormal by construction and the
rounding still preserves conditional means. The missing input to the
value limit is the minimizing identity $\mathbb E a(t,X_t)^2=1$.

The terminal spin also survives:
$U_1=\operatorname{sgn}(B_1)$ almost surely. But its correlation is

$$
\mathbb E[U_1B_1]=\sqrt{\frac2\pi}=\mathcal P(\gamma_0),
$$

the value of this nonminimizing candidate. The negative first variation
proves that this value is strictly above $P_*$. Admissibility, exact
Euler sampling, and a spin-valued terminal martingale therefore do not
supply the minimizing normalization or identify the ground-state value.
<!-- /passage:gaussian-sanity-check -->

<!-- passage:gaussian-one-step-example -->

<a id="one-step-example"></a>

### 10.4 An exact one-step calculation

To test the definitions on a finite example, we use the admissible heat
coefficient $\gamma_0\equiv0$, whose solution has
$$
 u_0(t,x)=2F_{\rm G}\bigl(x/\sqrt{1-t}\bigr)-1,\qquad
 a_0(0,0)=\sqrt{2/\pi}.
$$
This coefficient is not a minimiser. The calculation therefore tests the
construction and its indices, and leaves the minimiser-dependent
convergence theorem aside. Take $T=4/5$ and $N=1$. Then
$$
 Y_1=\frac2{\sqrt5}Z_1,\qquad
 c_1=\sqrt{\pi/2},\qquad g_1=Z_1,\qquad
 m_1=u_0(T,Y_1)=2F_{\rm G}(2Z_1)-1.
$$
Strict monotonicity and symmetry of the Gaussian distribution function give
$$
 F_1=\operatorname{sgn}\bigl(F_{\rm G}(2Z_1)+F_{\rm G}(Z_2)-1\bigr)
     =\operatorname{sgn}(2Z_1+Z_2)
$$
almost surely, so the additional Gaussian is visibly the rounding
coordinate. Set $V=(2Z_1+Z_2)/\sqrt5$, a standard normal. Gaussian
regression gives $\mathbb E[Z_1\mid V]=2V/\sqrt5$ and
$\mathbb E[Z_2\mid V]=V/\sqrt5$. Since
$\mathbb E|V|=\sqrt{2/\pi}$,
$$
 A_0=0,\qquad A_1=B_1=2\sqrt{\frac{2}{5\pi}},\qquad
 B_2=\sqrt{\frac{2}{5\pi}}.
$$
Consequently
$$
 \sum_{j=0}^1A_jB_{j+1}=\frac4{5\pi}
 \quad\hbox{whereas}\quad A_1B_1=\frac8{5\pi}.
$$
Dropping the index shift changes the answer by a factor of two.

At $N=1$ the interior Riemann sum is empty, and the entire product comes
from the rounding endpoint. That endpoint term disappears only
asymptotically, and at a finite mesh it need not vanish. The displayed
coefficients follow from exact Gaussian regression, so the example needs
no numerical quadrature or finite computation.
<!-- /passage:gaussian-one-step-example -->

<!-- passage:gaussian-all-mesh-example -->
<a id="heat-all-mesh"></a>

#### 10.4.1 The same test at every mesh size

For the nonminimizing coefficient $\gamma=0$, Euler sampling is exact:
$Y_j=B_{t_j}$. The curvature moment in §10.3 gives
$$
c_j=\sqrt{\pi/2}(1-t_{j-1}^2)^{1/4}.
$$
Monotonicity and symmetry of the Gaussian CDF turn the rounded spin into
$$
F_N=\operatorname{sgn}(B_T+\sqrt{1-T}\,Z_{N+1})
\quad\hbox{almost surely}.
$$
The argument of this sign is standard normal. Gaussian regression gives
$B_i=\sqrt{2/\pi}\sqrt\Delta$ for $1\le i\le N$ and
$B_{N+1}=\sqrt{2/\pi}\sqrt{1-T}$.
To calculate $A_j$, we condition on Brownian motion through $t_j$ and use
the heat martingale. Gaussian integration by parts then gives
$$
\mathbb E[Z_j u_0(t_j,x+\sqrt\Delta Z_j)]
=\sqrt\Delta\,a_0(t_{j-1},x).
$$
Multiplication by $c_ja_0(t_{j-1},B_{t_{j-1}})$ and expectation yield
$A_j=c_j\sqrt\Delta\,\mathbb E a_0(t_{j-1},B_{t_{j-1}})^2
=\sqrt\Delta/c_j$. Consequently, for every $N\ge1$,
$$
S_{T,N}=\sqrt{\frac2\pi}
 \left(\Delta\sum_{j=1}^{N-1}\frac1{c_j}
             +\frac{\sqrt{\Delta(1-T)}}{c_N}\right).
$$
The one-step value above is its $N=1$ case. At fixed $T<1$, its limit is
$$
\frac2\pi\int_0^T(1-t^2)^{-1/4}\,dt,
$$
whereas the unnormalized candidate covariance is $I_0(T)=T\sqrt{2/\pi}$.
The subsequent $T\uparrow1$ limit is
$4\Gamma(3/4)/(\sqrt\pi\,\Gamma(1/4))\approx0.76275976$, distinct from
$\mathcal P(0)=\sqrt{2/\pi}\approx0.79788456$.
The general predictable comparison of [§10.2](#independent-scalar-lower-bound)
turns this limit into a lower bound for $P_*$, as shown next. It
determines neither $P_*$ nor the value of the zero coefficient.

Exact Euler simulation, orthonormal predictable increments and correct
rounding all hold here. The missing identity is
$\mathbb E a_0(t,B_t)^2=1$, and consequently the normalizers do not
approach one. The calculation extends the heat and rounding derivations of
§§10.3–10.4 and isolates the role of minimality in the convergence theorem.
<!-- /passage:gaussian-all-mesh-example -->

<!-- passage:heat-lower-bound -->
<a id="heat-lower-bound"></a>

**Corollary 10.2 (the heat test).** In this normalization,
$$
P_*\ge\frac{4\Gamma(3/4)}{\sqrt\pi\,\Gamma(1/4)}
       \approx0.7627597635.
\tag{10.2}
$$
For each fixed $T<1$ and $N\ge1$, the heat normalizers $c_j$ are
positive. The increments $g_j$ have unit second moments and are
orthogonal, since their multipliers depend only on preceding Gaussians.
The rounded $F_N$ is a measurable spin after an arbitrary choice on its
null tie event. Thus Proposition 10.1 applies with
$K=N$, $z_j=Z_{j+1}$ and $f_j=g_{j+1}$, giving $S_{T,N}\le P_*$.
First let $N\to\infty$ at fixed $T$, using the displayed exact sum,
and then let $T\uparrow1$. The integrand is nonnegative and its endpoint
singularity has exponent $1/4<1$. Substitution $s=t^2$ yields
$$
P_*\ge\frac2\pi\int_0^1(1-t^2)^{-1/4}\,dt
 =\frac1\pi B\!\left(\frac12,\frac34\right)
 =\frac{4\Gamma(3/4)}{\sqrt\pi\,\Gamma(1/4)}.
$$
The bound follows from the heat calculation and the general Gaussian
comparison. Minimality is unnecessary for the inequality. It is what made
the different, optimizing construction converge to $P_*$ itself.
<!-- /passage:heat-lower-bound -->

<a id="further-circuit-comparisons"></a>

## 11. Alternative constructions and circuit limits

<!-- passage:alternative -->

<a id="positive-temperature-route"></a>

### 11.1 The separate positive-temperature route

The positive-temperature appendix supplies a separate classical input to
the same synthesis and seed-removal theorems. Fix $\beta_0>1$, let
$\mu_{\beta_0}$ be its Parisi measure, and write
$q=q_{\beta_0}=\max\operatorname{supp}\mu_{\beta_0}$.
Here $\gamma(t)=\beta_0\mu_{\beta_0}([0,t])$,
$m=\phi_{\beta_0,x}$, $a=\phi_{\beta_0,xx}$ and $X$ belong to the
positive-temperature PDE, whose terminal value is
$\beta_0^{-1}\log(2\cosh(\beta_0x))$. They are distinct from the
zero-temperature objects in §§1–8. The input credited to
[Lopatto, arXiv:2607.11756v3, Theorem 1.1](https://arxiv.org/pdf/2607.11756v3)
gives $\operatorname{supp}\mu_{\beta_0}=[0,q]$ for $\beta_0>1$, and
its smooth density below $q$ gives the regularity used on a fixed
$[0,T]$ with $0<T<q$.

The normalization can be checked directly against Lopatto's PDE.
Writing $\widehat\Phi_{\beta_0}(t,y)$ for his solution with terminal
value $\log\cosh y$ and covariance factor $\beta_0^2$, set
$$
\phi_{\beta_0}(t,x)
 =\frac{\widehat\Phi_{\beta_0}(t,\beta_0x)+\log2}{\beta_0},
\quad m=\widehat\Phi_y(t,\beta_0x),\quad
 a=\beta_0\widehat\Phi_{yy}(t,\beta_0x).
$$
His diffusion $\widehat X$ has noise coefficient $\beta_0$, hence
$X=\widehat X/\beta_0$ has unit noise and drift
$\beta_0\mu_{\beta_0}([0,t])m$. His identity
$\beta_0^2\mathbb E\widehat\Phi_{yy}(t,\widehat X_t)^2=1$
is exactly $\mathbb E a(t,X_t)^2=1$. This is a fixed-temperature
change of variables, not convergence of either Parisi measure as
$\beta_0\to\infty$. [Lopatto, arXiv:2607.11756v3, equations
(1.3), (2.1)–(2.2) and Lemma 2.2](https://arxiv.org/abs/2607.11756v3).

The source derives $\mathbb E a(t,X_t)^2=1$ on $[0,q]$ and
$1-q\le\beta_0^{-1}$. The value estimate is
$$
\int_0^T\mathbb E a(t,X_t)\,dt
\ge P_*-\frac{1/4+\log2}{\beta_0}-(q-T).
$$
To recover its constants, martingale contact first gives
$\int_0^t\mathbb E a(s,X_s)^2ds=t$ on $[0,q]$, and continuity yields
unit second moment including the endpoints. On $[q,1]$, $\gamma=\beta_0$,
so Cole–Hopf gives
$$
\phi(q,x)=\beta_0^{-1}\log(2\cosh\beta_0x)+\beta_0(1-q)/2,
\quad m(q,x)=\tanh(\beta_0x),\quad a(q,x)=\beta_0\operatorname{sech}^2(\beta_0x).
$$
The endpoint identity $\mathbb E a(q,X_q)^2=1$ is also needed.
Cauchy–Schwarz gives
$$
1-q=\mathbb E\operatorname{sech}^2(\beta_0X_q)
\le\bigl(\mathbb E\operatorname{sech}^4(\beta_0X_q)\bigr)^{1/2}
=\beta_0^{-1}.
$$
This is the endpoint argument in [OpenAI (2026b), Appendix A,
Lemma “A positive-temperature value bound”][q-positive].
Itô's product rule and the value evolution give
$$
\int_0^q\mathbb E a=\mathbb E[X_qm(q,X_q)]-\int_0^qt\gamma(t)dt,
\quad
\mathbb E\phi(q,X_q)=\phi(0,0)+\tfrac12\int_0^qt\gamma(t)dt.
$$
To compare this integral with the spin optimum, define the expected
free energy in energy units by
$$
F_{n,\beta_0}=(n\beta_0)^{-1}\mathbb E\log\sum_\sigma
e^{\beta_0H_n(\sigma)},\qquad
F_{\beta_0}=\lim_{n\to\infty}F_{n,\beta_0}.
$$
The fixed-temperature limit is supplied by the positive-temperature
Parisi formula cited below. The finite-size soft-maximum inequality
gives $F_{\beta_0}\ge P_*$. Using
$x\tanh(\beta_0x)\ge\beta_0^{-1}\log(2\cosh\beta_0x)
-\beta_0^{-1}\log2$ and retaining the penalty on $[q,1]$ yields
$$
\int_0^q\mathbb E a\ge F_{\beta_0}
-\frac{\beta_0(1-q)^2}{4}-\frac{\log2}{\beta_0}.
$$
The coefficient $1/4$ comes from the two tail contributions, since
$\gamma=\beta_0$ above $q$:
$$
\frac{\beta_0}{2}\int_q^1 t\,dt
 -\frac{\beta_0}{2}(1-q)
=-\frac{\beta_0}{4}(1-q)^2.
$$
The cutoff removes at most $q-T$, because $0\le\mathbb E a\le1$.
This proves the displayed estimate. Choose first a sufficiently large
finite $\beta_0$, then $T<q$ close enough to $q$. This route obtains the existence and value of the
ground-state limit from the positive-temperature Parisi formula of
[Talagrand (2006), Theorem 1.1, pp. 222–223](https://doi.org/10.4007/annals.2006.163.221).
At each fixed $\beta_0$, his covariance hypotheses (1.3)–(1.5) hold with
$\xi(t)=\beta_0^2t^2/2$ and error $c(n)=\beta_0^2/(2n)$.
His partition sum (1.14) is unnormalized, and dividing its logarithmic
free energy by $\beta_0$ gives the convention here, including the
$\log2/\beta_0$ entropy term. The finite-size inequality used above is
$$
F_{n,\beta_0}-(\log2)/\beta_0\le n^{-1}\mathbb E\max H_n\le F_{n,\beta_0}.
$$
It does not invoke the zero-temperature companion
to justify that step. Write $M_n=n^{-1}\mathbb E\max H_n$. The same
uniform entropy bound gives
$$
|F_{n,\beta_0}-P_*|\le |M_n-P_*|+\frac{\log2}{\beta_0}.
$$
Thus arbitrary joint $n,\beta_0\to\infty$ are permitted for these scalar
free-energy values. This says nothing about joint convergence of measures,
PDE derivatives, message calculations or circuit words.
[OpenAI (2026b), Appendix A][q-positive].

The remaining limits have distinct tasks. At fixed
$\beta_0,T,\delta,L$, with $\delta>0$, integer $L\ge1$ and $L\delta\le T$, the degree limit $D\to\infty$ identifies the
finite message law and its edge value. Let $U_0,U_1,\ldots$ be independent
standard normals. Put
$$
X_0^\delta=\sqrt\delta\,U_0,\qquad
X_{r+1}^\delta=X_r^\delta+\delta\gamma(r\delta)m(r\delta,X_r^\delta)
                   +\sqrt\delta\,U_{r+1},\qquad
c_{\delta,r}=(\mathbb E a(r\delta,X_r^\delta)^2)^{-1/2}.
$$
The finite-degree output is a message calculation on the $(D+1)$-regular
tree, where $\partial i$ is the set of neighbours of vertex $i$.
Write $t_r=r\delta$ and $b(t,x)=\gamma(t)m(t,x)$.
Initialise independent standard normal labels $g_i$, with
$u^0_{i\to j}=u_i^0=g_i$, $A^{-1}_{i\to j}=1$ and
$x^0_{i\to j}=x_i^0=\sqrt\delta\,g_i$. For $0\le r<L$ set
$$
u^{r+1}_{i\to j}=D^{-1/2}\!\sum_{\substack{v\in\partial i\\v\ne j}}
 A^{r-1}_{v\to i}u^r_{v\to i},\qquad
u^{r+1}_i=(D+1)^{-1/2}\!\sum_{v\in\partial i}
 A^{r-1}_{v\to i}u^r_{v\to i}.
$$
Each $x$ then takes the Euler step
$x^{r+1}=x^r+\delta b(t_r,x^r)+\sqrt\delta\,u^{r+1}$ with its corresponding
directed-edge or vertex message. The directed-edge law fixes the common
normaliser:
$$
c_{D,r}=(\mathbb E a(t_r,x^r_{i\to j})^2)^{-1/2},\quad
A^r_{i\to j}=c_{D,r}a(t_r,x^r_{i\to j}),\quad
B_i^r=c_{D,r}a(t_r,x_i^r),\quad
z_i^L=\sqrt\delta\sum_{r=1}^{L}B_i^{r-1}u_i^r.
$$
The one-step lag in $A^{r-1}u^r$ gives conditional centring from the new
outer Gaussian layer, although successive finite-degree messages need not
be independent. At fixed $\delta,L$, the joint vertex messages converge to
the independent $U_r$ above, and $z_i^L$ converges to
$\sqrt\delta\sum_{r=1}^Lc_{\delta,r-1}a(t_{r-1},X^\delta_{r-1})U_r$.
The source's separate edge calculation gives
$$
\lim_{D\to\infty}\frac{D+1}{2\sqrt D}\mathbb E[z_o^Lz_v^L]
  =\delta\sum_{r=1}^{L-1}c_{\delta,r}\mathbb E a(r\delta,X_r^\delta),
\qquad o\sim v.
$$
Only afterwards take $\delta\downarrow0$ with $L=\lfloor T/\delta\rfloor$.
The source bounds the discrepancy from
$\int_0^T\mathbb E a(t,X_t)\,dt$ by $C(\beta_0,T)\sqrt\delta$.
A separate norm-convergence argument with a uniform first occupation
moment identifies this edge limit with the rooted energy. A
finite-dimensional central limit theorem alone does not identify that
observable. [OpenAI (2026b), Appendix A][q-positive].

<!-- /passage:alternative -->

<!-- passage:weighted-tree -->
<a id="eams-finite-comparison"></a>

**Source comparison for the finite weighted-tree formula.** This
auxiliary comparison concerns the positive-temperature route, and a reader
following the circuit argument can go directly to
[§11.1.1](#temperature-route-comparison) or [§11.2](#fixed-word-probability).

The comparison is with
[El Alaoui–Montanari–Sellke, arXiv:2111.06813v2, 3 February 2023](https://arxiv.org/abs/2111.06813v2).
Proposition 2.1 (p. 7) sums the adjacent-time term over
$\ell=2,\ldots,L$. Applying Lemma 4.3, equation (4.6) (p. 14),
replaces the expectation in that summand by the coefficient at
$r=\ell-1$, so the resulting range is $r=1,\ldots,L-1$.
The displayed ranges in (4.5) (p. 13) and (4.18) (p. 16) are instead
$2,\ldots,L$.

Separately, the same-time display (2.21) in Lemma 2.8
(p. 9) omits a factor $1/k$ that is present in its proof and in
Proposition 2.1. The QAOA manuscript's Appendix A already uses the latter proposition,
the reindexed sum and spatial Lipschitz bounds for both $a$ and $b$.
The two following calculations establish the finite discrepancies and
their precise scope. They do not compare formulae in the journal edition.

**The test weights expose the missing factor.** Work on the $k$-regular
tree, $k\ge2$, and let all initial labels $g_i$ be independent standard
Gaussians. In particular, the adjacent vertices $o,v$ have distinct
labels $g_o,g_v$. Set $u_i^0=u_{i\to j}^0=g_i$,
$A_{i\to j}^{-1}=1$, and use constant algorithmic update functions
$F_r=H_r=1$. These satisfy EAMS Assumption 2 and its later boundedness
Assumption 3. For the auxiliary tests in Lemma 2.8 choose
$Y_o^r=g_o$ and $\widetilde Y_v^r=g_v$ for every $r\ge0$.
Each test is measurable with respect to its required initial-label
sigma-field, and all moments are finite. The tests are separate from
the constant update functions.

At $t=1$, the recursion gives
$$
u_o^1=k^{-1/2}\sum_{x\sim o}g_x,\qquad
u_v^1=k^{-1/2}\sum_{y\sim v}g_y.
$$
Hence
$$
\mathbb E[u_o^1u_v^1g_og_v]
=\frac1k\sum_{x\sim o,\,y\sim v}\mathbb E[g_xg_yg_og_v]
=\frac1k.
$$
Only $x=v,y=o$ survives. The printed right side of (2.21) is
$\mathbb E[g_v^2g_o^2]=1$, and multiplying it by $1/k$ agrees with the
calculation. Both initial labels matter. With both test weights equal
to one, both sides are zero. With just the neighbouring test omitted,
the left side is an odd Gaussian moment and is again zero.

A bounded version uses $Y_o^r=\operatorname{sgn}(g_o)$ and
$\widetilde Y_v^r=\operatorname{sgn}(g_v)$. Only the same edge pair
survives, so
$$
\mathbb E[u_o^1u_v^1\operatorname{sgn}(g_o)\operatorname{sgn}(g_v)]
=\frac1k(\mathbb E|g|)^2=\frac{2}{\pi k},
$$
whereas the printed right side is $2/\pi$. These tests are bounded
and initial-label measurable. Both witnesses concern the auxiliary
weighted identity, and neither substitutes arbitrary test weights into a
Parisi output theorem.

**The shifted endpoint violates the printed uniform finite bound.**
For a separate test retain the initial labels and set
$$
\delta=\varepsilon=\frac14,\qquad L=2,\qquad b(t,x)=0,
\qquad
a(t,x)=
\begin{cases}
1,&0\le t<3/8,\\
-1,&3/8\le t<1.
\end{cases}
$$
Use the source's actual initial states
$x_i^0=x_{i\to j}^0=\sqrt\delta\,g_i$ and its definitions
$A_{i\to j}^r=a(r\delta,x_{i\to j}^r)$,
$B_i^r=a(r\delta,x_i^r)$. The state initialization has not been set
to zero. It drops out of these coefficients because $a$ is independent
of $x$. This fixed function is bounded by $K=1$, has spatial
Lipschitz constant zero, and satisfies $a^2=1$, so all directed-edge
normalizers are exactly one. The drift has Lipschitz constant $M=0$,
and $L\delta=1/2\le1-\varepsilon$. It therefore meets the printed
hypotheses of Theorem 4.1 and Assumptions 2–3. That finite theorem
does not require temporal continuity or positivity of $a$.

At the mesh points, $a_0=a_1=1$ and $a_2=-1$. Thus
$$
u_i^2=\frac1{\sqrt{k(k-1)}}
 \sum_{w\sim i}\sum_{y\sim w,\,y\ne i}g_y,
\qquad z_i^2=\sqrt\delta\,(u_i^1+u_i^2).
$$
The equal-time covariances across the edge are zero. Each adjacent-time
covariance is $\sqrt{k-1}/k$: exactly $k-1$ labels occur in both sums,
and their product normalization is $k\sqrt{k-1}$. Therefore
$$
k\mathbb E[z_o^2z_v^2]=2\delta\sqrt{k-1},\qquad
\frac{k}{2\sqrt{k-1}}\mathbb E[z_o^2z_v^2]=\delta.
$$
The target printed in (4.5) uses $a(2\delta,x_o^2)=-1$, so it is
$-2\delta\sqrt{k-1}$. Its absolute error is
$4\delta\sqrt{k-1}$, which is unbounded as $k\to\infty$ while all
arguments of the claimed constant $C(M,K,\delta,\varepsilon)$ remain
fixed. This refutes auxiliary Theorem 4.1 as printed. The example is
not a Parisi curvature: it changes sign. Consequently it is not by
itself a counterexample to the specialized assertion (4.18) in
Proposition 4.7, and the endpoint correction there comes from the exact
proposition and reindexing explained above. Neither test refutes the
main EAMS Theorem 1.7.

**A finite estimate with the required hypotheses.** The source argument
does provide a usable replacement. Fix $\delta>0$, $L\ge1$ and
$L\delta<1$. Allow the coefficients $a_k,b_k$ to depend on degree, but
require bounds independent of $k$ at the finitely many mesh times:
$|a_k|\le K$ with $K\ge1$, and spatial Lipschitz constants $L_a,L_b$
for $a_k,b_k$. Use the original initialization and the same normalized
function for node and directed-edge coefficients, with
$\mathbb E(A_{i\to j}^r)^2=1$. Then
$$
\left|k\mathbb E[z_o^Lz_v^L]
 -2\sqrt{k-1}\,\delta\sum_{r=1}^{L-1}\mathbb E B_o^r\right|
 \le C(K,L_a,L_b,\delta,L).
$$
The constant is independent of $k$, and the sum is empty when $L=1$.
The coefficient expectation on the right uses the vertex law, while its
normalization was defined using the directed-edge law. The following
proof keeps the two laws distinct. It reconstructs the finite estimate
from EAMS Proposition 2.1 and Lemmas 3.2–3.4 and 4.2–4.3. It does not
claim that changing an index alone repairs every hypothesis of the
printed theorem.

Put $D=k-1$. Conditional on the earlier cavity labels, the new
message is centered Gaussian with variance $(\tau^r)^2$. Its recursion is
$$
(\tau_{i\to j}^{r+1})^2=\frac1D
\sum_{\substack{w\in\partial i\\w\ne j}}(A_{w\to i}^{r-1})^2(\tau_{w\to i}^r)^2,
\qquad \tau^0=1.
$$
This conditional Gaussianity gives
$\Delta_r=\mathbb E[((\tau^r)^2-1)^2]\le V_r/D$, where
$$
V_0=0,\qquad V_{r+1}=2K^4V_r+2(K^4+1).
$$
Indeed, split the conditional-variance recursion into the average of
$(A^{r-1})^2((\tau^r)^2-1)$ and the average of $(A^{r-1})^2-1$.
Jensen bounds the first square, while independence of child cavities
and exact coefficient normalization bound the second by a constant
over $D$. The moment recursion also gives
$\mathbb E(u_{i\to j}^r)^2\le K^{2r}$.

Splitting off the missing neighbour as in the source gives
$\mathbb E|u_i^r-u_{i\to j}^r|^2\le4K^{2r}/k$.
Subtract the node and directed-edge Euler recursions. Their equal
initial states and the drift Lipschitz bound give
$$
\mathbb E|x_i^r-x_{i\to j}^r|^2\le d_r/k,
\quad d_0=0,\qquad
d_{r+1}=2(1+\delta L_b)^2d_r+8\delta K^{2(r+1)}.
$$
The bound
$\|B_i^r-A_{i\to j}^r\|_2\le L_a\sqrt{d_r/k}$
now uses spatial Lipschitzness of $a$. This property is invoked in the
proof of EAMS Lemma 4.2 but omitted from the printed hypotheses of
Theorem 4.1. It is available in the QAOA manuscript's compact-strip application.

For $s=\ell-1$, the adjacent-time summand of Proposition 2.1 is
$$
Q_s=\mathbb E[A_{v\to o}^{s-1}(u_{v\to o}^s)^2B_v^{s-1}B_o^s].
$$
Let $\mathcal G_{i\to j}^{r,+}$ be the labels within distance $r$
of $i$ in its cavity after removing edge $ij$. Condition on
$\mathcal G_{v\to o}^{s-1,+}\vee\mathcal G_{o\to v}^{s,+}$.
The three coefficient factors are measurable there: their vertex balls
split into these two disjoint cavity balls. The newest layer driving
$u_{v\to o}^s$ is outside this field. Adding the initial label at $v$
to the source's minus field does not change the conditional law, since
that message and its variance exclude that label. Conditioning therefore replaces $(u_{v\to o}^s)^2-1$ by $(\tau_{v\to o}^s)^2-1$.
Its contribution is bounded by $K^3\sqrt{V_s/D}$. Replacing the two
node coefficients by directed-edge coefficients then costs at most
$K^2L_a(\sqrt{d_{s-1}}+\sqrt{d_s})/\sqrt k$.
The two directed-edge trees are independent, and exact normalization
reduces their product expectation to $\mathbb E A_{o\to v}^s$.
One final comparison with $\mathbb E B_o^s$ proves
$$
|Q_s-\mathbb E B_o^s|\le\frac{C_s}{\sqrt D},\quad
C_s=K^3\sqrt{V_s}+K^2L_a\sqrt{d_{s-1}}
                         +(K^2+1)L_a\sqrt{d_s}.
$$
For the same-time remainder in that proposition,
$|R_\ell|\le K^{2\ell+2}$ by boundedness and Cauchy–Schwarz. Hence
the asserted finite bound holds, for example, with
$$
C=2\delta\sum_{s=1}^{L-1}C_s
                  +\delta\sum_{\ell=1}^LK^{2\ell+2}.
$$
No constant uniform in increasing horizon, temperature or terminal
cutoff has been asserted.

**Why the continuum value survives the correction.** The QAOA manuscript already checks
the convergence and uniform boundedness of its finitely many
degree-dependent normalizers, and supplies both spatial Lipschitz
bounds, before taking the degree limit. Its limiting coefficient
$m_r=c_{\delta,r}\mathbb E a(t_r,X_r^\delta)$ satisfies $|m_r|\le1$
by Cauchy–Schwarz and the definition of $c_{\delta,r}$. The shifted
and corrected sums differ exactly by
$$
\delta\sum_{r=2}^Lm_r-\delta\sum_{r=1}^{L-1}m_r
=\delta(m_L-m_1),\qquad |\delta(m_L-m_1)|\le2\delta.
$$
Thus the error in the printed unnormalized finite theorem can grow
like $\sqrt{k}$ at fixed mesh even though the correction disappears
in the subsequent normalized continuum limit. The fixed-temperature
coupling and separate strong realization in the QAOA manuscript justify that limiting
application. These checks neither reject nor independently certify
the whole EAMS or QAOA theorem. [OpenAI (2026b), Appendix A][q-positive].
<!-- /passage:weighted-tree -->

<!-- passage:alternative-assembly -->
After clipping and bounded approximation, the result enters the same
synthesis and seed-removal route as the primary construction. Its source
dependencies and fixed-temperature choice of parameters remain separate
from the retained zero-temperature formal interfaces.
[OpenAI (2026b), Appendix A][q-positive].

<a id="temperature-route-comparison"></a>

### 11.1.1 Where the temperature routes meet

Both routes produce a bounded finite classical calculation whose rooted
energy is close to $P_*$. They reach that common input by different
theorems and different finite approximations.

**Variational input and support.** The zero-temperature route fixes an
attained minimizer $\gamma\in\mathcal U$ with terminal value $|x|$ and
uses full relative Stieltjes support on $[0,1)$. The positive-temperature
route fixes a finite $\beta_0>1$ and its Parisi probability measure, with
terminal value $\beta_0^{-1}\log(2\cosh\beta_0x)$. It uses Lopatto's
interval-support result on $[0,q_{\beta_0}]$.

**Scalar error.** Before discretization, the zero-temperature loss is at
most $1-T$ for $T<1$. The positive-temperature loss is at most
$$
(1/4+\log2)/\beta_0+(q_{\beta_0}-T),\qquad T<q_{\beta_0}.
$$

**Finite energy.** In the zero-temperature route, a first-chaos projection
identifies the explicit Gaussian coefficient sum with
$\operatorname{Re}\langle F_N,\mathsf L F_N\rangle$. The alternative
route first calculates finite-degree messages and edge energy, then uses
a rooted norm and occupation argument.

**Order of approximation.** At zero temperature, fix $T$ before refining
the scalar mesh, and choose a finite mesh before regularization and synthesis.
At positive temperature, fix $\beta_0,T,\delta,L$ for the message degree
limit, then refine $\delta$ at fixed $\beta_0,T$.

Both routes then use the same bounded approximation, exclusions, seeded
synthesis, seed removal and fixed-word transfer. The quantum construction
is shared, and its classical input has been supplied independently.

Here the displayed positive-temperature estimate is exactly the one
derived above, rather than an assertion of convergence of zero-temperature
supports. One can choose $\beta_0$ and $T$ finitely for the required
error and then stop. The later circuit argument needs that finite
choice and never asks for a joint limit of the temperature, message
mesh, graph degree and circuit depth. [OpenAI (2026b), Appendix A,
Proposition “The independent positive-temperature construction”][q-positive];
[§6][q-assembly].
<!-- /passage:alternative-assembly -->

<!-- passage:gibbs-comparison -->

### 11.1.2 Finite-temperature values and exact Gibbs samples

This comparison samples the exact Gibbs distribution. The QAOA Born law
in §11.2 is a separate distribution, and none of the following bounds
identifies the two. For one fixed instance, let $L_n=\max_\sigma H_{n,J}(\sigma)/n$,
$\mathcal F_{n,\beta}=(n\beta)^{-1}\log\sum_\sigma e^{\beta H_{n,J}(\sigma)}$,
and let $\mathcal U_{n,\beta}$ be the mean energy per spin under the
Gibbs law proportional to $e^{\beta H_{n,J}(\sigma)}$. Then
$$
0\le\mathcal F_{n,\beta}-L_n\le\frac{\log2}{\beta},\qquad
0\le L_n-\mathcal U_{n,\beta}\le\frac{\log2}{\beta}.
$$
The first is the log-sum bound. The second uses
$\mathcal F=\mathcal U+S/(n\beta)$, where the Gibbs entropy satisfies
$0\le S\le n\log2$, together with $\mathcal U\le L_n\le\mathcal F$.
Thus both disorder averages approach $P_*$ along every joint
$n\to\infty$, $\beta\to\infty$, by the already established
convergence of $\mathbb E L_n$.

There is also a finite probability estimate. Every configuration has
Gaussian coefficient-vector norm $\sqrt{(n-1)/(2n^2)}$ after division
by $n$, and their maximum has the same Lipschitz bound. Gaussian
concentration gives, for $n\ge2$,
$$
\Pr_J(|L_n-\mathbb E L_n|>s)
\le2\exp\{-n^2s^2/(n-1)\}.
$$
For an exact Gibbs draw $\Sigma^{\rm G}$, count the at most $2^n$
configurations with deficit at least $s$ and bound the partition sum
below by its maximal term. This gives, for every instance,
$$
\Pr\{H_{n,J}(\Sigma^{\rm G})/n\le L_n-s\mid J\}
\le\min\{1,e^{n(\log2-\beta s)}\}.
$$
The lower tail below $P_*-s$ is bounded by
$\Pr_J(L_n<P_*-s/2)+\min\{1,e^{n(\log2-\beta s/2)}\}$;
the upper tail is at most the upper tail of $L_n$. Hence an exact
Gibbs sample also approaches $P_*$ in joint probability. These are
elementary scalar consequences of the Gibbs law, entropy and Gaussian
concentration. They supply neither an efficient sampler nor QAOA angles.
<!-- /passage:gibbs-comparison -->

<!-- passage:qaoa-quantifiers -->

<a id="fixed-word-probability"></a>

### 11.2 What is averaged, and a fixed-word probability consequence

The probability conclusion here uses a fixed-word second-moment theorem
in addition to the expected-value construction of §§8.1–8.6.

The theorem in §8 is an existence statement about limiting expected
energy, and two random choices lie behind that expectation. First the
couplings $J$ are sampled. Then a computational-basis measurement of the
state produces a spin string $\Sigma$. For the single word $W$ already
chosen, put
$$
E_n(J,\Sigma)=H_{n,J}(\Sigma)/n,\qquad
\overline E_n(J)=\mathbb E[E_n(J,\Sigma)\mid J].
$$
The quantity defining $v_W$ is $\mathbb E_J\overline E_n(J)$.
It is neither the energy of every measured string nor a worst-disorder
guarantee. The temporary Gaussian labels used during synthesis add no
randomness to the final angle vectors.

Farhi, Goldstone, Gutmann and Zhou supply the additional input. They prove, with
the same Hamiltonian and gate-sign convention, that for every fixed word
$$
\mathbb E_{J,\Sigma}E_n\longrightarrow v_W,
\qquad
\mathbb E_{J,\Sigma}E_n^2\longrightarrow v_W^2.
$$
Their second expression is the quantum expectation of $(C_{n,J}/n)^2$
averaged over disorder, not merely the square of the conditional mean.
Consequently
$$
\mathbb E_{J,\Sigma}(E_n-v_W)^2\longrightarrow0,
\qquad
\Pr_{J,\Sigma}(|E_n-v_W|>s)\longrightarrow0
\quad(s>0).
$$
The law of total variance separates what vanishes:
$$
\mathbb E(E_n-v_W)^2=
\mathbb E_J\operatorname{Var}(E_n\mid J)
 +\mathbb E_J(\overline E_n(J)-v_W)^2.
$$
Both terms are nonnegative and hence tend to zero. Concentration of the
conditional mean alone would not control a measured string.
This is the elementary second-moment deduction from [Farhi et al.,
arXiv:1910.08187v4, §4, equations (13)–(22)][fggz].

Now run the circuit construction with accuracy $\varepsilon/2$ and
freeze its resulting word. Since $v_W\ge P_*-\varepsilon/2$,
$$
\Pr_{J,\Sigma}\{E_n<P_*-\varepsilon\}
\le \frac{4}{\varepsilon^2}
       \mathbb E_{J,\Sigma}(E_n-v_W)^2\longrightarrow0.
$$
This follows from the expected-value construction **together
with** the external second-moment theorem, and it lies outside both
selected formal interfaces.

If
$b_n(J)=\Pr(E_n<P_*-\varepsilon\mid J)$, then for every fixed
$\tau>0$,
$$
\Pr_J\{b_n(J)>\tau\}
\le\mathbb E_J b_n(J)/\tau\longrightarrow0.
$$
Thus the same word also has small conditional measurement-failure
probability for most disorder instances, in this precise sense. We claim
nothing for every instance, for arbitrarily varying depths, or with a
quantitative size threshold.

For $K$ fresh state preparations at the same disorder instance, the
conditional failures are independent. Their joint failure probability is
therefore $\mathbb E_J b_n(J)^K$, which in general differs from
$(\mathbb E_J b_n(J))^K$.
A fraction $\alpha$ of instances that always fail gives $\alpha$ for
the former at every $K$. A bound $b_n(J)\le\tau$ on a particular
instance does give $\tau^K$ there. Repeatedly measuring an already
collapsed state is a different experiment.

<!-- /passage:qaoa-quantifiers -->

<!-- passage:sampled-finite-instance -->
<a id="sampled-finite-instance"></a>

**Comparison with the same finite instance.** The concentration of the
finite maximum in §11.1.2 strengthens this to a near-optimal sampled
configuration. For every $\varepsilon>0$, there is one finite
deterministic word $W$, independent of $n$ and $J$, such that
$$
\Pr_{J,\Sigma}\!\left\{
H_{n,J}(\Sigma)\ge\max_\sigma H_{n,J}(\sigma)-\varepsilon n
\right\}\longrightarrow1.
$$
To see this, choose the word with $v_W\ge P_*-\varepsilon/4$ and
write $r_n=\mathbb E(E_n-v_W)^2\to0$ and
$M_n=\max_\sigma H_{n,J}(\sigma)/n$. The convergence
$\mathbb E M_n\to P_*$ makes its bias at most $\varepsilon/4$
for all sufficiently large $n$. The Gaussian coefficient-vector bound
from §11.1.2 and Chebyshev's inequality for $E_n$ then give, for $n\ge2$,
$$
\Pr_J\{M_n>P_*+\varepsilon/2\}
 \le2\exp\!\left\{-\frac{n^2\varepsilon^2}{16(n-1)}\right\},
\qquad
\Pr_{J,\Sigma}\{E_n<v_W-\varepsilon/4\}
 \le\frac{16r_n}{\varepsilon^2}.
$$
Outside these two events, $M_n-E_n\le\varepsilon$. Their union
therefore has probability tending to zero. The maximum and the measured
energy use the same disorder, and the union bound needs no independence.

One fresh preparation and measurement suffices for this asymptotic
statement. Its inputs are the ordinary-word theorem, the external
fixed-word second-moment theorem and Gaussian concentration, and it lies
outside the selected formal endpoints. The unspecified sequence
$r_n$ and mean-maximum bias still give no effective size threshold or
efficient angle-selection procedure.
<!-- /passage:sampled-finite-instance -->

<!-- passage:qaoa-depth -->

<a id="depth-quantifiers"></a>

### 11.3 Depth, angle choice and the order of quantifiers

The complete witness is a finite ordered word, including every real
angle. Its existence for each error tolerance asserts uniformity in
system size and disorder that fixed-instance convergence does not provide.
We read the quantifiers literally.

The source theorem says that for every $\varepsilon>0$, there is one
finite word $W_\varepsilon$ such that
$$
\lim_{n\to\infty}\mathbb E E_n(W_\varepsilon)\ge P_*-\varepsilon.
$$
The cost operator uses $J,n$, but this word and its angles do not.

In fixed-instance QAOA convergence, by contrast, the chosen depth and
angles may depend on the fixed finite cost matrix and the requested accuracy.
The performance of a prescribed family of nonzero angles with
$p=p(n)\to\infty$ needs further estimates. A depth allowance alone can
be met by the padded existence construction below.

The distinction already occurs for the elementary array
$a_{n,p}=p/(n+p)$. For each fixed $n$, its limit as $p\to\infty$
is one, and for each fixed $p$, its limit as $n\to\infty$ is zero.
This is a quantifier example and makes no claim to model QAOA. The finite-dimensional
convergence theorem of [Binkowski, Koßmann, Ziegler and Schwonnek,
arXiv:2302.04968v1, Theorem 12][binkowski] fixes the optimization
problem before selecting a sufficiently long circuit. It does not
replace the construction of a size-independent word here.

<a id="existential-depth-schedule"></a>

**An existential schedule under a prescribed depth allowance.**

There is a narrower diagonal consequence for sampled output. Choose
$W_k$ with $P_*-1/k\le v_{W_k}\le P_*$. Fixed-word second-moment
convergence supplies $N_k$ so that, for every $n\ge N_k$,
$\mathbb E(E_n(W_k)-v_{W_k})^2\le k^{-2}$. Given any prescribed
integer depth allowance $d(n)\to\infty$, enlarge these strictly
increasing thresholds until $d(n)\ge\operatorname{depth}(W_k)$
for every $n\ge N_k$. On $N_k\le n<N_{k+1}$, use $W_k$ padded
with zero-angle layers to depth $d(n)$. Padding changes no state, and
$$
\mathbb E(E_n(W_k)-P_*)^2\le
2\mathbb E(E_n(W_k)-v_{W_k})^2+2(v_{W_k}-P_*)^2\le4/k^2.
$$
Thus a suitable schedule approaches the optimum in joint $L^2$, even
under an arbitrarily slowly diverging allowance. The word family and
thresholds are selected nonconstructively. This says nothing about an
independently prescribed family of nonzero angles, and is not an
interchange of limits or an effective depth bound.

<a id="computational-meaning"></a>

**Computational meaning of a finite word.**

Nor does a supremum over angles assert that an optimizing angle vector
exists at any particular depth. The proof constructs a sufficiently
good word directly, and monotonicity from identity padding then gives the
limit of $Q_p$. Each QAOA round contains a full cost evolution and a
mixer evolution. In the dense SK model a cost evolution involves all
$\binom n2$ pair couplings, so the round count is not itself a bound on
two-qubit gate depth, connectivity cost, precision or physical runtime.

Three computational questions remain distinct: evaluating the limiting
energy at supplied angles, finding suitable angles, and implementing and
sampling the resulting quantum circuit. The [Basso et al. iteration][bfmvz] evaluates the fixed-angle
limit in $O(p^2 4^p)$ arithmetic time, which bounds the cost of
evaluation. The search needed to find the existence theorem's angles is a
separate problem.

Spin–boson representations provide another route to evaluation. The earlier
[Boulebnane et al. mapping, arXiv:2505.07929v2][spin-boson], and the
later [mixed-spin decoupling paper, arXiv:2609.40309v1, §§II–III and
Supplement S2.1][decoupling], must be distinguished. The latter is a
separate work, distinct from the reference called BKL2026 in the original
QAOA manuscript. Its fixed-depth mapping and numerical tensor-network
performance do not supply an efficient angle-selection theorem for
the present construction. Its discussion also gives fixed-word
finite-size estimates. The absence of a finite-size rate **in the
construction here** is not a claim that the wider literature contains
none.

<a id="higher-spin-boundary"></a>

**Why the two-spin hypothesis matters.**

The statement is specific to the quadratic, zero-field SK model.
The pure even higher-spin model has a different thermodynamic-first
conclusion. For every even interaction order $q\ge4$, [Basso,
Gamarnik, Mei and Zhou, arXiv:2204.10306v2, Theorem 4, (3.15),
and its proof in §4.2](https://arxiv.org/abs/2204.10306v2) show
$$
v_p^{(q)}(\boldsymbol\theta,\boldsymbol\beta)
\le\eta_{\mathrm{OGP}}(G_q)<\eta_{\mathrm{OPT}}(G_q)
$$
for every finite $p$ and every fixed angle vector, in that source's
normalization. The same ceiling bounds their supremum over all finite
$p$ and angles. The proof combines its dense–sparse value identity
with the coupled-overlap-gap obstruction of [Chou, Love, Sandhu and Shi,
arXiv:2108.06049v3, Parameter 4.1 and Corollary 4.4](https://arxiv.org/abs/2108.06049v3).
That local-algorithm obstruction applies to sparse random signed
Max-$q$-XOR for even $q\ge4$, at sufficiently large fixed degree and
under the source's logarithmic-depth bound. The Basso–Gamarnik–Mei–Zhou
transfer above only needs the resulting obstruction at each fixed depth.
It requires no interchange of depth and system size. The exclusion
$q\ge4$ is essential, so this is not an obstruction for the SK two-spin
model treated here.

For comparison, [Goh, arXiv:2404.06087v6, Theorem 4.2 and
Corollary 4.2.1][goh], treats even interaction order at least four.
Its logarithmic-depth obstruction concerns sparse random-regular
Max-q-XORSAT, and the dense spin-glass corollary is explicitly conditional
on swapping limits. The source states on p. 12 that its sparse/dense
equivalence is only established at fixed depth, so it does not prove
a dense logarithmic-depth obstruction. These hypotheses do not include
quadratic SK. Neither a conclusion for all mixed spin glasses nor a
general quantum advantage follows from §8.

### 11.4 Limits of the regular-MaxCut consequence

The original manuscript also gives a consequence for MaxCut. For each
$\eta>0$ it selects one finite depth and normalized angles, and then
a degree threshold, such that every finite simple unweighted $(D+1)$-regular graph of
girth greater than $2p+1$ has expected cut fraction at least
$$
\frac12+\frac{P_*-\eta}{\sqrt D}
$$
for all sufficiently large $D$. The usual MaxCut cost is
$H_{\rm MC}=\frac12\sum_{\{i,j\}}(I-Z_iZ_j)$, so its angles are
$2\boldsymbol\gamma/\sqrt D$: the normalized angles are fixed,
while these physical cost angles depend on degree. The sign conversion
uses a bipartition of the relevant tree neighbourhood of each edge, so
the whole graph need not be bipartite.

The further optimality comparison is for a uniform simple labelled
$(D+1)$-regular graph $G_{n,D+1}$, along sizes with $n(D+1)$ even.
Write $\operatorname{OPT}(G)$ for the maximum cut size divided by $|E(G)|$,
$q_{p,G}$ for the circuit's expected cut fraction, and
$\Delta_D(G)=\operatorname{OPT}(G)-q_{p,G}(2\boldsymbol\gamma/\sqrt D,
\boldsymbol\beta)$. The same word can be chosen so that
$$
\lim_{D\to\infty}
\limsup_{\substack{n\to\infty\\n(D+1)\ \mathrm{even}}}
\Pr_G\{\sqrt D\,\Delta_D(G_{n,D+1})>\eta\}=0.
$$
Size goes to infinity first at fixed degree and word. The probability is
over the random graph, and the cut fraction has already averaged the
quantum measurement. For fixed $D,p$, [McKay, Wormald and Wysocka (2004),
Theorem 1](https://doi.org/10.37236/1819), gives tight counts of cycles
of lengths at most $2p+3$. Every edge with a nontree neighbourhood is
within bounded distance of one of these cycles, so their fraction tends
to zero.

The optimum is supplied separately by [Dembo, Montanari and
Sen, arXiv:1503.03923v2, Theorem 1.6(b)](https://arxiv.org/abs/1503.03923v2):
dividing their cut size per vertex by $(D+1)/2$ gives the leading cut
fraction $1/2+P_*/\sqrt{D+1}$, with an error $o(D^{-1/2})$ in the
stated iterated regime. A high-girth lower bound alone would not prove
optimality on every high-girth graph. See [OpenAI
(2026b), Corollary 1.2 and its complete proof][q-maxcut]. We include this
separate consequence to keep the source's scope in view. It adds no
selected formal endpoint and no joint size–degree limit.

<!-- /passage:qaoa-depth -->

<!-- passage:qaoa-finite-size -->

<a id="finite-size-comparison"></a>

### 11.5 A later finite-size estimate and its limited depth consequence

Boulebnane, Khan, Subrahmanya, Herman, Farhi, Villalonga and Shaydulin
later gave a quantitative comparison with the fixed-angle limit
[arXiv:2609.40309v1, Corollary S3.1, pp. 46–47][decoupling]. They use the
ordered-tuple Gaussian mixed-spin cost with covariance polynomial
$\xi(t)=\sum_q c_q^2t^q$ of finite degree.
Let $\Gamma>0$ bound every cost angle in a depth-$p$ word. Choose
$x>0$ bounding $\xi(2),\xi'(1),\xi''(33)$, with
$\Gamma^2p^2x\ge1$. The mixer angles need no bound. Set
$$
B_p=256p^4(2e\Gamma^2p^2x)^p,\qquad
E_p=70\Gamma^2p^2x+p\log2,\qquad
R_p=(2E_p/\log2)^{2p}.
$$
For
$$
0<\varepsilon\le4\Gamma px
 \min\{e^{-1/2},2^{-(8p^2+3)/100}\},
$$
the source asserts $|\nu_n-\nu_\infty|\le\varepsilon$ whenever
$$
n\ge T_p:=\max\left\{
2^{8p^2+3}B_p
 \left(\frac{2^{\deg\xi+4}\Gamma^2p^2x^2}{\varepsilon^2}\right)^{R_p}
 \left[2\log_2\frac{4\Gamma px}{\varepsilon}+\deg\xi\right]^2,
\frac{16\Gamma px}{\varepsilon}\right\}.
$$
Here $\nu_n$ is the disorder-averaged energy per spin.

The estimate applies to each near-optimal word from §8.6 separately.
Given $\zeta>0$,
choose a finite word $W_\zeta$ with
$\nu_\infty(W_\zeta)\ge P_*-\zeta/2$. Its depth $p_\zeta$ and cost
angles are finite. We pad by a zero layer if needed and take
$\Gamma_\zeta=\max\{1,\max_j|\theta_j(W_\zeta)|\}$. Choose a positive
$\varepsilon_\zeta\le\zeta/2$ satisfying the displayed small-error
condition. With these fixed parameters, the threshold gives
$$
\nu_n(W_\zeta)\ge P_*-\zeta\qquad(n\ge T_{p_\zeta}).
$$
The threshold is explicit in the chosen word and error. The existential
construction supplies no quantitative bounds for $p_\zeta$ or
$\Gamma_\zeta$ as $\zeta\downarrow0$, so it gives no bound in terms of
the accuracy alone. The fixed-word application needs no common angle
bound, while the uniform growing-depth consequence below does.

For SK take $c_2=1/\sqrt2$. Combining the two ordered off-diagonal
couplings recovers our standard Gaussian $J_{ij}$, and the diagonal adds
an independent scalar Gaussian of variance $1/2$. That scalar changes
the circuit by a global phase and has zero mean, so the expected-energy
comparison uses the same normalization. A fixed
$x\ge\max\{2,\Gamma^{-2}\}$ satisfies the above bounds for every
$p\ge1$.

For fixed mixture, $\Gamma$ and $x$, put $\varepsilon_p=e^{-p^2}$.
The small-error condition holds eventually, and direct logarithms give
$$
\log R_p=4p\log p+O(p),\qquad
\log\log T_p=4p\log p+O(p+\log p).
$$
Consequently, for any deterministic sequence of words $W_n$ with
depths $p(n)\to\infty$, one common cost-angle bound $\Gamma$,
and $p(n)\log p(n)=o(\log\log n)$, the displayed hypotheses hold
and $|\nu_n(W_n)-\nu_\infty(W_n)|\le e^{-p(n)^2}$ eventually.
This is a uniform comparison along that specified bounded-angle class.
It gives neither an efficient way to find near-optimal angles nor
concentration of a quantum measurement in this growing-depth regime.
The fixed-word second-moment theorem in §11.2 has separate quantifiers.
<!-- /passage:qaoa-finite-size -->

<a id="formal-interfaces"></a>

## 12. The formal statements and their scope

<!-- passage:formal -->

### 12.1 The support statement and the derivative bridge

The first formal statement assumes a Brownian system and an admissible
minimizer. It supplies a Stieltjes measure, an associated diffusion, full
relative support and strict increase, and the two moment identities for
every associated diffusion. Its proof assembles contact, exclusion of a
terminal plateau, curvature for positive and zero constant intervals,
and uniqueness to transfer the moments between associated diffusions.
This describes the inspected proof's mathematical route, and we claim no
successful research-target replay.

Its quantifiers are $\forall W\,\forall\gamma$, followed by the minimizer
hypothesis, and it supplies both witnesses on the given Brownian system. The moment
conclusion is universal over every diffusion satisfying the specified
equation, and covers exactly $0\le t<1$. The diffusion equation itself
holds through time one, simultaneously for all those times outside
one null set. Controls are progressive for the natural filtration of
Brownian increments and bounded by one pointwise. These are not
additional regularity or minimizer-existence assumptions. See the
[exact support statement][support-type] and the companion
[source-interface review](https://github.com/wcook04/lean-exposition/blob/21b1f9e995c4276fdad02377b378f11541a3d5e9/papers/qaoa-sk/source-review/full-support-interface.md).

The manuscript's first variation and its formal counterpart differ in one
useful respect. The formal result initially gives a derivative
along chord lengths $1/(n+1)$. For this particular control functional,
that sampled limit determines the ordinary feasible right derivative.
With a control fixed, the terminal absolute value is convex in the
coefficient while the penalty is linear. The supremum over the same
control class, and then subtraction of the linear Parisi penalty,
preserve convexity.

The secant slopes therefore vary monotonically as the chord shortens, and
each intervening chord lies between two successive
sampled chords and has the same limiting slope. If $\eta$ is an
integrable feasible direction, choose $\varepsilon_0>0$ so that
$\gamma+\varepsilon_0\eta$ is admissible and rescale the chord. This
bridge is an ordinary deduction from the definitions and the sampled
limit, and no separately named formal theorem states it.

At the origin, the manuscript smoothly extends auxiliary quotient
variables. The formal coercivity proof for the spatial diffusion density instead works on the positive
half-line, controls the zero flux at the origin and reflects. The two
must agree in the resulting integrated inequality, while auxiliary values
at zero may differ. In particular, the declared generic flux lemma has a
global derivative hypothesis and is not the supplier used here. The
finite-density proof uses the Gaussian half-line coercivity lemma and
reflection. A totalized quotient value at zero cannot silently satisfy the
generic global hypothesis. The precise source spans are in the
correspondence record.

The [fixed-time supplier `actual_coercivity`][support-fixed-time] assumes
positive time and continuity of the coefficient at that time. Its scaled
polynomial inequality is (4.1) multiplied by $c^2$, and division requires
$c>0$. This gives a direct source destination for Proposition 4.1 without
adding a minimizing or plateau premise.

<!-- /passage:formal -->

<!-- passage:formal-scalar -->
<a id="formal-scalar-statement"></a>

### 12.2 The separate scalar statement

The second formal statement has three clauses. Given a minimizer and an
associated diffusion, it establishes convergence of the finite-spin
ground-state values, a bounded martingale with a spin-valued terminal
limit, convergence of the gradient and the value identities. Under the
same hypotheses and fixed $0<T<1$, it gives eventual positivity of the
Gaussian normalizers and convergence of the finite coefficient sum.

These are the declarations `scalar_value_target` and
`finite_gaussian_target`. Their caller supplies the same minimizer and
associated diffusion. The implementation derives compact-strip bounds
from `sourceStripRegularity` and the unit curvature moment from contact
for those objects before applying its fixed-strip helper. Those internal
inputs are not additional hypotheses of the exported clauses. The
convergence clause establishes existence of the finite-spin limit.

The index match in the Gaussian statement is exact. For
$\Delta=T/N$ and $t_j=j\Delta$, the Euler path uses
$Z_1,\ldots,Z_N$ and produces $Y_0,\ldots,Y_N$.
The source normalizer indexed from zero corresponds to manuscript
$c_1,\ldots,c_N$, and its eventual positivity covers precisely those
normalizers. A separate $Z_{N+1}$ performs mean-preserving rounding.
It enters $B_{N+1}$ and is no Euler increment. Thus the finite sum is
$\sum_{j=0}^{N}A_jB_{j+1}$ with $A_0=0$, using exactly the displayed
index ranges. Total definitions at $N=0$ make no extra mathematical
claim about the construction for $N\ge1$.

More explicitly, the
zero-based formal quantities `coordinate N j`, `rawCoefficient ... j`
and `normalizer ... j` correspond to $Z_{j+1}$,
$a(t_j,Y_j)$ and $c_{j+1}$ for $0\le j<N$. The positivity quantifiers
are $\exists N_0\,\forall N\ge N_0\,\forall j<N$, simultaneously over
the growing mesh, for each already fixed $T$. The extra coordinate
with formal index $N$ is $Z_{N+1}$.

The formal martingale stops at its terminal random variable $U$ for all
later times. Then $U^2=1$ almost surely, and the gradient converges to $U$ almost surely and in $L^2$ from below
one. The internal terminal-gradient lemma proves $UX_1=|X_1|$
without requiring $X_1$ to have no atom at zero. That identity is used
inside the implementation and is no extra conjunct of the endpoint. See the [scalar statement][value-type] and its
[complete dictionary](https://github.com/wcook04/lean-exposition/blob/21b1f9e995c4276fdad02377b378f11541a3d5e9/papers/qaoa-sk/source-review/scalar-value-interface.md).

<!-- /passage:formal-scalar -->

<!-- passage:formal-parisi-clause -->
<a id="formal-parisi-clause"></a>

The third declaration, `ground_state_parisi_target`, assumes a minimizer
alone and concludes $P_*=\mathcal P(\gamma)$. Its proof chooses
$X=\texttt{diffusionProcess}\ W\ \gamma$ and proves the diffusion
predicate for that process before applying the equality with a supplied
diffusion. No diffusion is required from the caller, and none of these
clauses constructs a minimizer. The exact declarations are
[MainConclusions, lines 13–27][scalar-main-conclusions]; their analytic
suppliers and applications are in [ContactAssembly][scalar-contact-assembly].
The independent contact and lower-bound arguments are reconstructed in
[§10.1](#formal-scalar-contact) and [§10.2](#independent-scalar-lower-bound).
<!-- /passage:formal-parisi-clause -->

<!-- passage:formal-transport -->
<a id="formal-model-transport"></a>

### 12.3 Relating the two statements

The support and scalar models encode closely related control problems.
A direct formal application still needs a specified map of their objects.
The following contract lists the obligations, and it asserts no
equivalence of their complete interfaces.

**Coefficients.** Restrict the scalar real-time coefficient to $[0,1)$.
Conversely, choose an extension compatible with the scalar right-continuity
predicate. Prove equality almost everywhere of the coefficients used by each
payoff and penalty. The scalar endpoint and exterior values are unconstrained
by the shape conditions, so two extensions need not be literally equal.

**Brownian space.** Bundle or unbundle the same sample space, sigma algebra,
probability law and Brownian process. Identify the natural filtration and
every elapsed-increment filtration. Equality in law alone does not identify
admissible controls on a chosen space.

**Controls and payoffs.** Bundle an admissible scalar control as a support
control and conversely. With $s=t+r$, prove equality between the absolute
integral over $[t,1]$ using $\alpha(s-t)$ and the elapsed integral over
$[0,1-t]$ using $\alpha(r)$. Preserve progressive measurability and the
pointwise bound $|\alpha|\le1$.

**Value and minimizer.** Equality of the payoff sets gives equality of their
suprema for every $x$ and $0\le t<1$. Only then identify the first and second
spatial derivatives and the Parisi penalties. Transport minimality against
every admissible competitor, including restriction or extension of that
competitor.

**Diffusion and moments.** Reindex the same process between nonnegative and
real time, and stop it at one where required. Transport the integral equation
on one full-measure event for all $t\in[0,1]$, path continuity, initial value,
gradient and moment integrals.

The diffusion comparison has a version issue. The support predicate asks
for a globally progressive process, while the scalar predicate asks for
adaptedness of the stopped process and almost-sure continuity on $[0,1]$.
An ordinary construction gives the needed progressive version.
For $t\ge0$ and $k=0,1,\ldots$, put
$$
 X_t^{(k)}=X_{\min(2^{-k}\lfloor2^k t\rfloor,1)}.
$$
Each left-step process is progressive: on every bounded time interval,
its value on a step is measurable at that step's left endpoint. Define
$\bar X_t$ as the pointwise limit where the sequence converges, and zero
elsewhere. The convergence set and this limit are progressively measurable.
On the single event where $X$ is continuous on $[0,1]$, the equality
$\bar X_t=X_{\min(t,1)}$ holds for every $t\ge0$. Thus stopping and
reindexing preserve paths, time integrals and moments on that event, and
constructing this version needs no completion of the natural filtration. The initial value follows from the scalar equation at zero and
$B_0=0$ on their common full-measure event.

This settles the version construction at the ordinary mathematical level.
Its implementation in the two formal predicates, and the coefficient,
filtration and value identifications listed above, remain separate
obligations. In particular, transporting the derivatives needs equality of
values for every $x$, and agreement at $(0,0)$ is insufficient.

Once those obligations are proved, an application must use the same
transported minimizer and associated diffusion in all premises. No such
cross-model transport theorem is supplied here. The manuscript proof in
§§3–7 and the scalar implementation's independent argument remain
available without it. These distinctions come from the defining
[support model][support-model] and [scalar model][value-model]
themselves.
<!-- /passage:formal-transport -->

<a id="obligations"></a>

### 12.4 Logical limits of the argument

<!-- passage:obligations -->
The full-support theorem and scalar-value theorem use related but
different formal representations. Their mathematical conclusions can be
discussed together, but a direct formal implication would require the
coefficient, Brownian law, control value and diffusion translations
described in §12, with common witnesses. The finite Gaussian result
also needs its rooted energy identification before it says anything
about a QAOA circuit. The boundedness, regularization, synthesis and
transfer estimates in §8 are additional steps, with temperature-specific
hypotheses where the positive-temperature route is used.

Infinite support allows gaps. An indefinite curvature identity gives strict
convexity only once the expectation is controlled, and an algebraic
certificate leaves its analytic hypotheses to be established.
Coordinatewise convergence does not control a growing coefficient sum.
A finite-dimensional Gaussian law identifies an energy observable only
together with its operator argument. Separate witnesses cannot be merged
silently. Finally, the order
$\forall\varepsilon\ \exists\text{finite word}\ \lim_{n\to\infty}$
cannot be replaced by a fixed-instance depth limit.

The local comparison in §11.1 concerns two printed finite-step displays.
It neither replaces the cited theorem's hypotheses nor supplies the
remaining circuit estimates.
<!-- /passage:obligations -->

<a id="verification"></a>

## 13. Source checks and finite calculation

<!-- passage:verification -->
The source collection retains the two manuscripts, formal statements,
their implementations and build metadata. The derived project specifies
Lean 4.34.1 and a pinned mathlib revision. The proof-assistant and library
infrastructure is described by [de Moura and Ullrich (2021), *The Lean 4
Theorem Prover and Programming Language*](https://doi.org/10.1007/978-3-030-79876-5_37)
and [The mathlib Community (2020), *The Lean Mathematical Library*](https://doi.org/10.1145/3372885.3373824).
These references credit the software, and the retained project metadata
identifies the versions used for this source.

The retained polynomial computation
in §4 uses exact rational arithmetic. Its expansion matches
all 22 positive remainder
coefficients. The surplus of the $v^4$ coefficient above $1/1000$ is
$140507/1521000000$. This finite check does not establish the PDE
signs, integrability, boundary flux or limiting argument.

The source-byte audit verified the inspected edition and proved none of
its claims. Paper assembly checked identifiers, declared scope and
document bindings. It did not check semantic correspondence or reader
understanding. Tool-compatibility checks and synthetic fixtures exercised
only their stated cases, which include neither selected SK theorem. Reading of the mathematical and
formal sources supports the bounded comparisons reported here, without
a complete independent proof acceptance or transitive axiom audit.

We keep the retained observations separate. The original graph records
unrun plans and an earlier full-support attempt that could not start
because the Comparator executable was unavailable. A later retained
attempt, observed on 23 October 2026, built and exported the challenge,
then stopped while building the implementing solution. The retained
[partial-observation receipt](https://github.com/wcook04/lean-exposition/blob/21b1f9e995c4276fdad02377b378f11541a3d5e9/examples/qaoa-sk/verification/full-support-build-failure-20261023.json)
and [log](https://github.com/wcook04/lean-exposition/blob/21b1f9e995c4276fdad02377b378f11541a3d5e9/examples/qaoa-sk/verification/full-support-build-failure-20261023.log) preserve that event: the imported
object for `Diffusion/AbsIncrement` was missing. The challenge is an
admitted specification, so its successful build and export are not a
proof of the theorem. There was no successful solution export or
research-target comparison in that observation.

The scalar-value invocation in the same original two-target run reached
a terminal failure observed on 24 October 2026. Its challenge built and
exported, but its solution build stopped because the imported object for
`Gaussian/IntegralPositiveMulGaussian` was missing. The retained scalar
receipt and log establish no solution export, kernel acceptance or
comparison acceptance. The companion source review locates both records.

Both failures concern the original build configuration. A subsequent
repair to the derived configuration was recorded. In the subsequent
corrected replay, the full-support challenge built and its exporter was
observed running. Reconciliation on 25 October found that the controller
had stopped without a recovered terminal research result. At that date,
Comparator and research extraction outcomes were unknown, neither
rejected nor accepted. The [interruption receipt](https://github.com/wcook04/lean-exposition/blob/21b1f9e995c4276fdad02377b378f11541a3d5e9/examples/qaoa-sk/verification/corrected-replay-interruption-20261025.json)
retains this observation separately from the original build failures.

A subsequent [source-bound observation of 26 October 2026](https://wcook04.github.io/lean-exposition-site/verification/full-support-build-extraction-observation-20261026.json)
records a successful local Lake build of the Support implementation and
v2 extraction of the single conditional full-support root. Its selected
traversal contains 1,674 declarations. The root reports no sorry, unsafe
or partial flag, and its recorded transitive axiom names are
`propext`, `Classical.choice` and `Quot.sound`. The selected in-scope
traversal completed, but dependencies outside it remain unresolved
boundaries, so the extracted graph is partial. This does not establish
minimiser existence, positive density, the scalar target or QAOA convergence.

The portable observation identifies the original receipts by their hashes
and retains the source inventory, root and extraction limits. Its
internal-consistency checks are not authenticated execution provenance.
The curated paper graph records the build and documentary extraction
observation separately, and it is distinct from the extracted declaration
graph.
Comparator acceptance, standalone kernel replay and the scalar-target
results remain pending. The earlier failed and interrupted invocations
remain dated history, rather than evidence of a later rejection or
acceptance. Permitted axioms and checker settings describe the requested
comparison and cannot supply its missing result.

This editorial revision does not establish a new formal result.
The [source review through 26 October 2026](https://github.com/wcook04/lean-exposition/blob/ef46d0e72787c1dc283733b10b433d112f97ac64/papers/qaoa-sk/source-review/references-and-adversarial-review.md)
records the preceding paper revision's reading coverage, finite checks,
unsuccessful observations and remaining obligations, and viewing it
requires repository access. This remains a draft after source-based agent
review. Independent mathematical acceptance and a human reconstruction
study have not been performed.
<!-- /passage:verification -->

## References

- Auffinger and Chen (2015), “On properties of Parisi measures,”
  *Probability Theory and Related Fields* **161**, 817–850. Inspected
  preprint: [arXiv:1303.3573v1](https://arxiv.org/abs/1303.3573v1).
- Farhi, Goldstone and Gutmann (2014), *A Quantum Approximate Optimization
  Algorithm*, [arXiv:1411.4028v1](https://arxiv.org/abs/1411.4028v1).
- Binkowski, Koßmann, Ziegler and Schwonnek (2023), *Elementary Proof of
  QAOA Convergence*, [arXiv:2302.04968v1](https://arxiv.org/abs/2302.04968v1).
- Basso, Gamarnik, Mei and Zhou (2022), *Performance and limitations of
  the QAOA at constant levels on large sparse hypergraphs and spin glass
  models*, [arXiv:2204.10306v2](https://arxiv.org/abs/2204.10306v2).
- Boulebnane, Khan, Liu, Larson, Herman, Shaydulin and Pistoia (2026),
  *Spin–Boson Mapping of the Quantum Approximate Optimization Algorithm*,
  [arXiv:2505.07929v2](https://arxiv.org/abs/2505.07929v2).
- Boulebnane, Khan, Subrahmanya, Herman, Farhi, Villalonga and Shaydulin
  (2026), *Decoupling of the QAOA into independent spin-boson systems and
  high-depth performance on pure and mixed spin glasses*,
  [arXiv 2609.40309, version 1](https://arxiv.org/abs/2609.40309v1).
- Goh (2025), *The Overlap Gap Property limits limit swapping in the
  QAOA*, [preprint, version 6](https://arxiv.org/abs/2404.06087v6).

- OpenAI (2026a), *Full support of the zero-temperature
  Sherrington–Kirkpatrick order parameter*, manuscript, 27 September 2026.
- OpenAI (2026b), *QAOA attains the SK ground-state energy in the
  thermodynamic-first limit*, manuscript, 25 September 2026.
- Auffinger and Chen (2017), “Parisi formula for the ground state energy
  in the mixed $p$-spin model,” *Annals of Probability* **45**, 4617–4631.
- Auffinger and Chen (2015), “The Parisi formula has a unique minimizer,”
  *Communications in Mathematical Physics* **335**, 1429–1444.
- Auffinger, Chen and Zeng (2020), “The SK model is infinite step replica
  symmetry breaking at zero temperature,” *Communications on Pure and
  Applied Mathematics* **73**, 921–943.
- Jagannath and Tobasco (2016), “A dynamic programming approach to the
  Parisi functional,” *Proceedings of the American Mathematical Society*
  **144**, 3135–3150.
- Chen, Handschy and Lerman (2018), “On the energy landscape of the mixed
  even $p$-spin model,” *Probability Theory and Related Fields* **171**,
  53–95.
- Montanari (2025), “Optimization of the Sherrington–Kirkpatrick
  Hamiltonian,” *SIAM Journal on Computing* **54**, FOCS19-1–FOCS19-38.
- El Alaoui, Montanari and Sellke (2021), “Optimization of mean-field spin
  glasses,” *Annals of Probability* **49**, 2922–2960.
- Farhi, Goldstone, Gutmann and Zhou (2022), “The Quantum Approximate
  Optimization Algorithm and the Sherrington–Kirkpatrick Model at Infinite
  Size,” *Quantum* **6**, 759.
- Basso, Farhi, Marwaha, Villalonga and Zhou (2022), “The Quantum
  Approximate Optimization Algorithm at High Depth for MaxCut on
  Large-Girth Regular Graphs and the Sherrington–Kirkpatrick Model,”
  *LIPIcs* **232**, 7:1–7:21.
- El Alaoui, Montanari and Sellke (2023), “Local algorithms for Maximum
  Cut and Minimum Bisection on locally treelike regular graphs of large
  degree,” *Random Structures and Algorithms* **63**, 689–715.
- Chen (2026), *FRSB in the SK spin glass: convergence to full-interval
  support at zero temperature*, arXiv:2607.18032v1.
- Lopatto (2026), *Full replica symmetry breaking in the
  Sherrington–Kirkpatrick model*, arXiv:2607.11756v3.
- Zhou (2026), “Existence of full replica symmetry breaking for the
  Sherrington–Kirkpatrick model at low temperature,” *Communications on
  Pure and Applied Mathematics* **79**, 1746–1770.

- Sellke (2021), *Optimizing Mean Field Spin Glasses with External
  Field*, [arXiv:2105.03506v4](https://arxiv.org/abs/2105.03506v4).
- Sellke (2024), *Optimizing Mean Field Spin Glasses with External
  Field*, *Electronic Journal of Probability* **29**, paper 4.
  Inspected revision: [arXiv:2105.03506v6, 17 December 2023](https://arxiv.org/html/2105.03506v6).
  Lemma 2.9 and Remark 2.10 are distinguished from the preceding v4 entry.
- Jekel, Sandhu and Shi (2026), *Potential Hessian Ascent: The
  Sherrington–Kirkpatrick Model*,
  [arXiv:2408.02360v3](https://arxiv.org/abs/2408.02360v3).
- Chou, Love, Sandhu and Shi (2022), *Limitations of Local Quantum
  Algorithms on Random Max-k-XOR and Beyond*,
  [arXiv:2108.06049v3](https://arxiv.org/abs/2108.06049v3).
- Lopatto (2026), *Full replica symmetry breaking in the
  Sherrington–Kirkpatrick model*,
  [arXiv:2607.11756v4](https://arxiv.org/abs/2607.11756v4).
  This later comparison is distinct from the v3 input cited above.
- Lopatto (2026), *Parisi measure of the Sherrington–Kirkpatrick model
  with external field*, [arXiv:2610.00127v1](https://arxiv.org/abs/2610.00127v1).
- Jones and Knill (1999), “Efficient Refocussing of One Spin and Two
  Spin Interactions for NMR Quantum Computation,” *Journal of Magnetic
  Resonance* **141**, 322–325. Inspected preprint:
  [arXiv:quant-ph/9905008v1](https://arxiv.org/abs/quant-ph/9905008v1).
- Li and Khaneja (2005), *Noncommuting vector fields, polynomial
  approximations and control of inhomogeneous quantum ensembles*,
  [arXiv:quant-ph/0510012v1](https://arxiv.org/abs/quant-ph/0510012v1).
- Talagrand (2006), “The Parisi formula,” *Annals of Mathematics*
  **163**, 221–263. [Published article](https://doi.org/10.4007/annals.2006.163.221).
- Dembo, Montanari and Sen (2017), “Extremal Cuts of Sparse Random
  Graphs,” *Annals of Probability* **45**, 1190–1217. Inspected
  preprint: [arXiv:1503.03923v2](https://arxiv.org/abs/1503.03923v2).
- McKay, Wormald and Wysocka (2004), “Short Cycles in Random Regular
  Graphs,” *Electronic Journal of Combinatorics* **11**, R66.
  [Published article](https://doi.org/10.37236/1819).

- de Moura and Ullrich (2021), “The Lean 4 Theorem Prover and
  Programming Language,” *CADE 28*, 625–635.
  [Published article](https://doi.org/10.1007/978-3-030-79876-5_37).
- The mathlib Community (2020), “The Lean Mathematical Library,”
  *CPP 2020*, 367–381.
  [Published article](https://doi.org/10.1145/3372885.3373824).

- Zhou (2026), “On the Gardner Transition in the Ising Pure p-Spin Glass II”.
  [arXiv:2608.06523v1][zhou-gardner], 6 August 2026. Appendix E.3,
  (E.28), (E.37), (E.41), pp. 127–128, and p. 130.

[scope]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/docs/281.md
[support-type]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/ComparatorChallenges/SKFullSupport.lean#L81-L92
[support-main]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKSupport/Main.lean#L19-L61
[support-config]: https://raw.githubusercontent.com/openai/math/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/ComparatorChallenges/SKFullSupport.json
[value-type]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/ComparatorChallenges/SKValue.lean#L20-L173
[value-main]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Main.lean#L16-L27
[value-config]: https://raw.githubusercontent.com/openai/math/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/ComparatorChallenges/SKValue.json
[chord]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKSupport/Support/VariationIntegrals.lean#L152-L175
[parameter-chords]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKSupport/Regularity/ParameterChords.lean#L12-L56
[support-definitions]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKSupport/Model.lean#L12-L79
[shifted-pairing]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Equations/ShiftedPairing.lean#L47-L91
[density-origin]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKSupport/Density/GaussianLogDensity.lean#L106-L166
[fs-main]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/main.tex
[fs-reg]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/regularity.tex
[fs-signs]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/signs.tex
[fs-cert]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/certificate.tex
[fs-curv]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/curvature.tex
[fs-var]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/variation.tex
[fs-cons]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/consequences.tex
[q-intro]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/01-introduction.tex
[q-foundations]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex
[q-classical]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/03-classical-optimization.tex
[q-synthesis]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/04-classical-synthesis.tex
[q-seeds]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/05-seed-simulation.tex
[q-assembly]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/06-assembly.tex
[q-positive]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/07-positive-temperature.tex
[ac2017]: https://arxiv.org/abs/1606.05335v2
[acz]: https://arxiv.org/abs/1703.06872v2
[chen2026]: https://arxiv.org/abs/2607.18032v1
[method-269]: https://github.com/wcook04/plectis-erdos/blob/a0878310af794d7a0f76cc11f4de3a20c795dc60/paper/reasoning-parts/erdos269/core.tex
[method-243]: https://github.com/wcook04/plectis-erdos/blob/a0878310af794d7a0f76cc11f4de3a20c795dc60/paper/reasoning-parts/erdos243/core.tex#L3358-L3366
[method-skill]: https://github.com/wcook04/plectis-erdos/blob/a0878310af794d7a0f76cc11f4de3a20c795dc60/skills/public-mathematical-writing/SKILL.md
[method-audit]: https://github.com/wcook04/plectis-erdos/blob/a0878310af794d7a0f76cc11f4de3a20c795dc60/scripts/reasoning_record_audit.py

[ac17]: https://arxiv.org/abs/1606.05335v2
[ac-properties]: https://arxiv.org/abs/1303.3573v1
[chen26]: https://arxiv.org/abs/2607.18032v1
[eams20]: https://arxiv.org/abs/2001.00904v1
[fs-density]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/order-parameter-regularity.tex
[fs-algorithm]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/algorithm-application.tex
[fs-variation]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/variation.tex
[scalar-contact-integral]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Variation/ContactIntegral.lean#L65-L125
[scalar-hump]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Coercivity/GapBoundary.lean#L84-L153
[scalar-convexity]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Coercivity/CurvatureMoment.lean#L75-L133
[scalar-zero-hump]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Coercivity/ZeroGapHump.lean#L8-L26
[scalar-contact-assembly]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Variation/ContactAssembly.lean#L21-L72
[q-fock]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L53-L149
[q-root-energy]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L150-L280
[q-gaussian]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L282-L365
[q-rooted-classical]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/03-classical-optimization.tex#L123-L255
[q-bounded-classical]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/03-classical-optimization.tex#L256-L338
[q-colors]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/03-classical-optimization.tex#L384-L608
[q-weights]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L366-L503
[q-unitary]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L504-L701
[q-substitution]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L784-L950
[q-gate-calculus]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L951-L1317
[q-synthesis-statement]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/04-classical-synthesis.tex#L38-L112
[q-masked-controls]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/04-classical-synthesis.tex#L113-L206
[q-helper-statement]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/04-classical-synthesis.tex#L207-L353
[q-helper-quadratics]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/04-classical-synthesis.tex#L354-L488
[q-additive-response]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L1318-L1459
[q-helper-responses]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/04-classical-synthesis.tex#L489-L960
[q-compilation]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/04-classical-synthesis.tex#L961-L1119
[q-seed-statement]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/05-seed-simulation.tex#L1-L49
[q-seed-packets]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/05-seed-simulation.tex#L50-L271
[q-seed-selective]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/05-seed-simulation.tex#L272-L504
[q-seed-echo]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/05-seed-simulation.tex#L505-L734
[q-seed-gram]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/05-seed-simulation.tex#L735-L873
[q-seed-visibility]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/05-seed-simulation.tex#L874-L1006
[q-transfer]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L1-L52
[q-ordinary-updates]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/02-foundations.tex#L702-L783
[fggz]: https://arxiv.org/abs/1910.08187v4
[bfmvz]: https://arxiv.org/abs/2110.14206v3
[binkowski]: https://arxiv.org/abs/2302.04968v1
[spin-boson]: https://arxiv.org/abs/2505.07929v2
[decoupling]: https://arxiv.org/abs/2609.40309v1
[goh]: https://arxiv.org/abs/2404.06087v6
[q-maxcut]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/08-regular-maxcut.tex

[support-model]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKSupport/Model.lean#L12-L88

[value-model]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Model.lean#L33-L118

[scalar-spin-limit]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/BlockCoeff.lean#L237-L263

[scalar-upper]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/WeakGuerra.lean#L117-L142

[scalar-column]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/GaussianColumn.lean#L14-L187

[scalar-adaptive]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/AdaptiveColumns.lean#L13-L134

[scalar-finite-adaptive]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/QuadraticSK.lean#L42-L88

[scalar-empirical-directions]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/EmpiricalVectors.lean#L106-L180

[scalar-transpose]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/GaussianTranspose.lean#L112-L140

[scalar-gram-schmidt]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/GramSchmidtLimit.lean#L33-L176

[scalar-empirical-limit]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/EmpiricalEnergyLimit.lean#L21-L122

[scalar-empirical-bound]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/EmpiricalCalculus.lean#L62-L141

[scalar-predictable]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/ConstantFatou.lean#L14-L134

[scalar-literal-lower]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/LiteralLower.lean#L14-L27

[q-formula-definition]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/build/sections/03-classical-optimization.tex#L339-L383

[scalar-main-conclusions]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/GroundState/MainConclusions.lean#L13-L27

[support-fixed-time]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKSupport/Density/SampleJets.lean#L135-L143

[zhou-gardner]: https://arxiv.org/abs/2608.06523v1
