# Full support and positive density in the zero-temperature SK model

*Will Cook · An account of the quantitative proof of OpenAI (2026a) · 3 November 2026*

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

<!-- passage:abstract -->
**Abstract.** For the pure, zero-field Sherrington–Kirkpatrick model at
zero temperature, the Stieltjes measure of every integrable minimiser of
the Parisi functional has full support on $[0,1)$ and a smooth
density that is strictly positive on $(0,1)$. We reconstruct OpenAI's
quantitative proof, which uses one spatial-moment inequality twice.
On a positive constant interval, the inequality makes the derivative of
the gradient second moment strictly convex, contradicting the endpoint
conditions imposed by minimality, and together with the bounded
gradient it also excludes terminal gaps. Once the support is full, moment
identities give smoothness, and a local constant replacement transfers
the inequality to the minimiser and forces its density to be positive.
H.-B. Chen previously proved full support and smooth nonnegative density.
We isolate the fixed-time spatial moment and the positive margin that
survives the local replacement, and we import its shape comparisons and
full polynomial identity. Along the way we derive the scalar value
identity and explain its classical optimisation consequence. For the quantum approximate
optimization algorithm (QAOA), we identify the rooted energy of a finite
Gaussian observable and the separate circuit theorem that approximates it,
and a further second-moment theorem then gives a measured near-optimal
configuration on the same random instance, with the circuit word fixed
before the number of spins grows. The construction gives no bound on
the depth needed for a given accuracy.
<!-- /passage:abstract -->

<a id="model-and-result"></a>

## 1. The model and the result

<!-- passage:model -->
For independent standard Gaussian couplings $J_{ij}$, the pure,
zero-field [Sherrington–Kirkpatrick (SK) Hamiltonian][SK] is
$$
H_n(\sigma)=n^{-1/2}\sum_{i<j}J_{ij}\sigma_i\sigma_j,
\qquad \sigma\in\{-1,1\}^n.
$$
With the maximisation convention, its limiting ground-state energy per
spin is
$$
P_*:=\lim_{n\to\infty}n^{-1}\mathbb E\max_\sigma H_n(\sigma).
$$
The Parisi formula expresses $P_*$ as a minimum over nondecreasing
functions. We study the minimiser itself. Its increments charge every
interval of overlap scales below one, and they have a strictly positive
density at each interior scale. One quantitative bound gives both facts:
it first excludes constant intervals and then, after an approximation,
gives pointwise positivity.

The overlap of two configurations is
$R(\sigma,\tau)=n^{-1}\sum_i\sigma_i\tau_i$, and it determines the
energy covariance, $\mathbb E[H_n(\sigma)H_n(\tau)]=nR^2/2-1/2$.
Hence the Gaussian law of the energies is encoded by the overlaps.
We use the covariance convention $\xi(t)=t^2/2$, so $\xi''=1$.
Adding one independent centred Gaussian of variance $1/2$ to every
configuration's energy gives covariance $n\xi(R)$ and leaves the
expected maximum unchanged, which is the normalisation of the
ground-state Parisi formula of [Auffinger and Chen (2017), Theorem 1][AC].
The model has no higher-order interactions and no external field.

Let $\mathcal U$ consist of the nonnegative, nondecreasing,
right-continuous functions $\gamma:[0,1)\to[0,\infty)$ satisfying
$\int_0^1\gamma(t)\,dt<\infty$, and define the locally finite
Stieltjes measure by
$$
\mu_\gamma([0,t])=\gamma(t),\qquad 0\leq t<1.
$$
Thus $\mu_\gamma((s,t])=\gamma(t)-\gamma(s)$, so the measure charges
each interval with the increment of $\gamma$ across it. It may have a point mass (an atom) of size
$\gamma(0)$ at zero, and its total mass need not be finite or equal to
one. Support statements refer to the relative topology of $[0,1)$.
A support gap is a relatively open interval of positive length and zero
mass, and the cumulative coefficient is constant on it.

The Parisi functional $\mathcal P(\gamma)$ is defined through an
auxiliary value function $\Phi_\gamma(t,x)$, in which $x\in\mathbb R$
is a spatial variable and the overlap parameter $t$ plays the role of
time. Write $\Phi=\Phi_\gamma$ and use subscripts for partial
derivatives. The backward equation, terminal condition and functional are
$$
\Phi_t=-\tfrac12\bigl(\Phi_{xx}+\gamma(t)\Phi_x^2\bigr),
\qquad \Phi(1,x)=|x|,
\qquad
\mathcal P(\gamma)=\Phi_\gamma(0,0)-\tfrac12\int_0^1t\gamma(t)\,dt.
$$
The terminal value $|x|$ is the zero-temperature value of one spin
in a field $x$. For an integrable coefficient, the solution of the PDE
is defined by approximation from bounded monotone coefficients.
[OpenAI (2026a), §1 and Lemmas 2.1–2.2][FS] give the exact control
definition and the stability theorem.

To turn minimality of $\mathcal P$ into a condition on $\gamma$, we use
a random process whose gradient moment measures the first-order cost of
changing $\gamma$. Write $u=\Phi_x$ for the spatial gradient and
$a=\Phi_{xx}$ for its spatial derivative. Let $B$ be standard Brownian
motion, and define the diffusion $X$ and its self-consistency function
$q$ by
$$
dX_t=\gamma(t)u(t,X_t)\,dt+dB_t,\qquad X_0=0,
\qquad q(t)=\mathbb E[u(t,X_t)^2].
$$
The discrepancy $q(t)-t$ enters the first variation of $\mathcal P$,
and minimality will force $q(t)=t$ on the support of $\mu_\gamma$.
The SDE has a unique strong solution. Unless an endpoint limit is
justified explicitly, all stochastic calculations below take place on
compact time intervals ending before one.
<!-- /passage:model -->

<!-- passage:main-result -->
The support of a minimiser is infinite by
[Auffinger, Chen and Zeng (2020), Theorem 1][ACZ]. Infinite support
still allows gaps: the measure $\sum_{k\geq1}2^{-k}\delta_{1-2^{-k}}$
has infinitely many atoms and many gaps. Full support and a smooth
nonnegative density were then proved in [H.-B. Chen (2026), Theorem 1.3][Chen]. We follow OpenAI's
quantitative proof of the support conclusion and of its further
strict-positivity result. The feature used twice is a lower bound
that holds throughout every positive constant interval of an admissible
coefficient, whether or not the coefficient minimises.

**Theorem 1 (full support; quantitative proof following
[OpenAI (2026a), Theorem 1.1][FS]).**
If $\gamma\in\mathcal U$ minimises $\mathcal P$, then
$$
\operatorname{supp}\mu_\gamma=[0,1),\qquad
\gamma(t_2)>\gamma(t_1)\quad(0\leq t_1<t_2<1).
$$
For the solution of the displayed SDE and every $0\leq t<1$,
$$
q(t)=t,\qquad \mathbb E[a(t,X_t)^2]=1.
$$

Corollary 4 in §4 adds a smooth density, strictly positive at every
interior point.

The existence of a minimiser is a separate input, from
[Auffinger and Chen (2017), Theorem 1 and the proof of Lemma 3][AC].
Integrability allows $\gamma$ to grow near one. Zero field gives
$u(0,0)=0$, and the curvature calculation uses $\xi''=1$, so the
theorem is stated only for the pure SK covariance at zero field.
Nonzero field and general mixed-spin covariances lie outside its
hypotheses.

Full support has a concrete meaning for the energy landscape.
For each fixed $\rho\in[0,1)$ and $\delta,\eta>0$, with probability
tending to one there are configurations $\sigma,\tau$ with
$|R(\sigma,\tau)-\rho|<\delta$ whose energies are both at least
$n(P_*-\eta)$. This follows from
[Auffinger, Chen and Zeng (2020), Proposition 1, equation (5)][ACZ]
in the Gaussian normalisation above.

We isolate the spatial moment that the support and density arguments
share, explain why it needs an integrated estimate, and identify what
survives the local-replacement limit. The conclusions and the
quantitative method are those of the cited sources.

The proof has two stages. A gap with two support endpoints faces two
demands. Minimality forces $q$ to agree with time at both endpoints,
with endpoint slopes at most one. On a positive constant interval, a
quantitative estimate makes $q'$ strictly convex, so its integral over
the gap is smaller than the gap's length, which contradicts the required
change in $q$. Once every gap is excluded, the identity $q'=1$ supplies
the equation from which the same estimate forces $\gamma'>0$.
For a bounded gap, a qualitative crossing criterion already suffices.
The density argument needs more: an estimate for a local constant
replacement, which need not minimise, at a time where $q''$ need not
vanish, with a positive lower bound that survives the limit.

Section 2 turns minimality into contact conditions and proves Theorem 1
from the curvature bound of Lemma 2. In §3 we derive the moment
polynomial behind that bound, together with the fixed-time estimate of
Proposition 3, which §4 applies to prove Corollary 4. Section 5 follows
the support identities to the value formula, the classical algorithm and
a finite Gaussian observable whose rooted energy a QAOA circuit
approximates, and then identifies the additional concentration input
for a measured configuration.

These optimisation consequences use full support only. Strict positivity of the density is a further analytic
conclusion, and readers interested in optimisation may accept Theorem 1
and go directly to §5. Sections 2–4 assume Itô calculus and martingale
convergence. The circuit construction is imported as a separate theorem.
Omitted proofs are linked to *From Parisi full support to deterministic
QAOA circuits: A long reasoning record*, the <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf">[companion edition of 3 November 2026](../long/reasoning-record.md)</span>,
and citations of the original sources stay beside the imported inputs.
<!-- /passage:main-result -->

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

## 2. Contact conditions and the exclusion of gaps

<!-- passage:contact -->
For an admissible coefficient, $\Phi$ is even, convex and one-Lipschitz
in space. Its spatial derivatives of positive order are bounded and
continuous on each strip $[0,T]\times\mathbb R$, $T<1$, and the PDE
holds for almost every time. The compact-strip approximation theorem of
[OpenAI (2026a), Lemmas 2.1–2.2][FS] justifies the Itô calculations
below even when $\gamma$ has jumps. This control method has antecedents
in [Jagannath and Tobasco (2016)][JT] at positive temperature and in
[Auffinger and Chen (2017), Theorem 2 and Corollary 2][AC] at zero
temperature (see also [Chen, Handschy and Lerman (2018), Proposition 2
and Theorem 5][CHL]).

We fix a minimiser. Itô's formula cancels the drift of $u(t,X_t)$:
$$
du(t,X_t)=a(t,X_t)\,dB_t,
\qquad q(0)=0,\qquad q'(t)=\mathbb E[a(t,X_t)^2].
$$
In particular, $q'$ is continuous across jumps of the coefficient.
Minimality means that every allowed one-sided change has nonnegative
first-order cost, and the one-sided variation formula computes that
cost. When $\gamma$ is replaced by $\gamma+\varepsilon\eta$ and
$\varepsilon\downarrow0$, the functional changes at the rate
$$
D_+\mathcal P(\gamma)[\eta]
=\tfrac12\int_0^1\eta(t)(q(t)-t)\,dt,                 \tag{1}
$$
for every integrable direction $\eta$ for which
$\gamma+\varepsilon\eta\in\mathcal U$ for all sufficiently small
$\varepsilon\geq0$ (see [OpenAI (2026a), Lemma 6.1][FS]). Earlier
zero-temperature variation and support identities appear in
[Chen, Handschy and Lerman (2018), Propositions 3–4][CHL].

Adding a unit atom at $s$ raises the cumulative function at every
$t\geq s$, so its direction is $\eta=\mathbf1_{[s,1)}$. Its
first-order cost is half of a tail integral, and minimality makes the
integral nonnegative:
$$
G(s):=\int_s^1(q(t)-t)\,dt\geq0.
$$
Thus $G(s)/2$ is the first-order cost of the added atom.

Mass can also be removed. Let $\nu=\mu_\gamma|_K$ for a compact
$K\subset[0,1)$. For $0\leq\varepsilon\leq1$, the measure
$\mu_\gamma-\varepsilon\nu=(1-\varepsilon\mathbf1_K)\mu_\gamma$
is nonnegative, and its cumulative function is
$\gamma-\varepsilon\nu([0,\cdot])\in\mathcal U$. Minimality gives
$$
0\leq D_+\mathcal P(\gamma)[-\nu([0,\cdot])]
=-\tfrac12\int_K G(s)\,d\mu_\gamma(s).
$$
The bounds $|q-t|\leq1$ and $\mu_\gamma(K)<\infty$ justify the
change in the order of integration. Since $G\geq0$, it vanishes almost
everywhere for $\mu_\gamma$, and by continuity it vanishes on
$S=\operatorname{supp}\mu_\gamma$. The variation removes a fraction of
all the mass in $K$, atoms and diffuse mass alike, and assumes no
density or probability constraint.

At a support point $s>0$, the function $G$ has a minimum. Hence
$$
G'(s)=s-q(s)=0,\qquad G''(s)=1-q'(s)\geq0,
\quad\text{so}\quad q(s)=s,\quad q'(s)\leq1.       \tag{2}
$$
If $0\in S$, contact gives $G(0)=0$, and zero field gives
$q(0)=0$ independently. Consequently
$G(h)=\tfrac12(1-q'(0))h^2+o(h^2)$, and nonnegativity gives the
same slope bound at zero. These are the contact conditions of
[OpenAI (2026a), Lemma 6.2][FS].
<!-- /passage:contact -->

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

<!-- passage:gap-proof -->
**Lemma 2 (constant-interval curvature; [OpenAI (2026a),
Proposition 5.1][FS]).** Let $\gamma\in\mathcal U$, with its
associated diffusion and moment $q$. If $\gamma=c$ on an open
interval $I\subset(0,1)$, called a plateau, then for $t\in I$
$$
q'''(t)\geq\frac{c^2}{1000}\mathbb E[a(t,X_t)^4]>0
\quad\text{if }c>0.                               \tag{3}
$$
If $c=0$, then $q''(t)=\mathbb E[\Phi_{xxx}(t,X_t)^2]>0$.

The lemma holds for every admissible coefficient, minimising or not.
Section 3 derives its moment polynomial and explains the integrated
estimate imported from the source.

**Proof of Theorem 1, assuming Lemma 2.** We treat the intervals of the
complement of $S$ by type. An interval with two support endpoints admits
two contact conditions. Initial intervals without support at zero have
coefficient zero. For a terminal gap, the bound $q\leq1$ replaces the
missing right contact condition.

Suppose first that $(\ell,r)$ is a support gap with endpoints in $S$
and $r<1$. The coefficient is a positive constant there, because
support at $\ell$ and the absence of mass immediately to its right
force $\mu_\gamma([0,\ell])>0$. Strict convexity of $q'$ and (2) give
$$
q'(t)<\frac{r-t}{r-\ell}q'(\ell)
       +\frac{t-\ell}{r-\ell}q'(r)\leq1
\qquad(\ell<t<r).
$$
Integrating contradicts $q(r)-q(\ell)=r-\ell$, so the geometric
obstruction is an area deficit under $q'$.

An initial gap $[0,r)$ with $0\notin S$ and $r\in S$ carries no
mass, hence $\gamma=0$ there. Its right endpoint gives $q(r)=r$
and $q'(r)\leq1$, and zero field gives $q(0)=0$. The zero-plateau
identity gives $q''>0$, hence $q'(t)<q'(r)\leq1$, and the same
integral contradiction follows without assuming $G(0)=0$. A gap
$(0,r)$ with $0\in S$ is already covered by the two-contact case.

A terminal gap $(r,1)$ with nonempty support has $r\in S$ and a
positive plateau height. Suppose that $q(t)\geq t$ for some $r<t<1$.
Strict convexity of $q'$ and $q'(r)\leq1$ give
$$
t-r\leq\int_r^t q'(s)\,ds
<\frac{t-r}{2}\bigl(q'(r)+q'(t)\bigr),
$$
so $q'(t)>1$. Convexity then gives $q'(s)>q'(t)>1$ for $s>t$, and
consequently $q(s)>t+(s-t)q'(t)>1$ for $s<1$ sufficiently close
to one, contrary to $q\leq1$. Thus $q(t)<t$ on $(r,1)$,
and $G(r)<0$ contradicts minimality. This case uses contact only at
the left endpoint.

If $S$ is empty, then $\gamma=0$. Lemma 2 makes $q$ strictly
convex, and $q(0)=0$ and $q\leq1$ give $q(t)<(t/b)q(b)\leq t/b$ for
$0<t<b<1$. Letting $b\uparrow1$ gives $q(t)\leq t$. Strict convexity
applied once more with $t<b<1$ now gives $q(t)<(t/b)q(b)\leq t$, so
$G(0)<0$.

The terminal exclusion was already noted in
[Auffinger, Chen and Zeng (2020), Remark 7][ACZ]. Our argument uses
only the curvature needed for the other gaps, and the full endpoint
discussion is in
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="qualitative-terminal-gap">[long §5, “A shorter qualitative exclusion”](../long/reasoning-record.md#qualitative-terminal-gap)</span>. A stronger boundary estimate
$1-q(t)\geq C\sqrt{1-t}$, for arbitrary coefficients constant near
one, is proved in [OpenAI (2026a), Lemma 6.3][FS] and in
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="terminal-gap">[long §5, terminal boundary estimate](../long/reasoning-record.md#terminal-gap)</span>,
and [Chen (2026), Proposition 4.5, equations (4.27)–(4.38)][Chen]
gives the square-root asymptotic under the terminal-gap hypothesis.
Antecedents of the zero-coefficient heat argument and the late-step
variation appear in [Chen, Handschy and Lerman (2018), Theorem 6][CHL].

These cases exhaust the complement of the relatively closed set $S$.
Consequently $S=[0,1)$, contact gives $q(t)=t$, and differentiation
gives $\mathbb E a^2=1$, including the right-hand value at zero.
Moreover,
$\gamma(t_2)-\gamma(t_1)=\mu_\gamma((t_1,t_2])>0$.

This proves Theorem 1 from the stated analytic inputs.
<!-- /passage:gap-proof -->

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

## 3. The integrated estimate

<!-- passage:curvature -->
Both arguments use one spatial moment. Recall $a=\Phi_{xx}$, write
$b=\Phi_{xxx}$ and $d=\Phi_{xxxx}$, and for $0<t<1$ set
$$
\mathcal C_\gamma(t)
=\mathbb E\bigl[d^2-12\gamma(t)ab^2+6\gamma(t)^2a^4\bigr],
$$
with the spatial factors evaluated at $(t,X_t)$.
The definition requires spatial regularity and no time derivative of
$\gamma$, so the estimate we need can be stated at a single time.

**Proposition 3 (fixed-time spatial-moment estimate).** Let
$\gamma\in\mathcal U$, let $0<t<1$ be a continuity point of $\gamma$,
and suppose that $c=\gamma(t)>0$. For its associated diffusion,
$$
\mathcal C_\gamma(t)\geq\frac{c^2}{1000}
                       \mathbb E[a(t,X_t)^4].
$$
Neither minimality nor a constant neighbourhood of $t$ is required.
This is the consequence of [OpenAI (2026a), Proposition 5.1 and
Lemmas 2.1–2.2][FS] stated in the companion's
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="fixed-time-coercivity">[Proposition 4.1, §4.5](../long/reasoning-record.md#fixed-time-coercivity)</span>.
We first explain the estimate on a positive plateau and then obtain
the fixed-time form by inserting a short plateau.

For a smooth coefficient, the two applications are connected by the
identity
$$
q'''+2\gamma'\mathbb E a^3=\mathcal C_\gamma.
$$
On a plateau $\gamma'=0$, while at a smooth minimiser full support
gives $q=t$ and hence $q'''=0$. Section 4 proves the required
smoothness before this second use.

On a plateau with $\gamma=c>0$, put $r=c\Phi_x=c u$,
$v=r_x=c a$, $w=r_{xx}$ and $z=r_{xxx}$. This rescaling makes
the derivative identities polynomial. The PDE becomes
$r_t=-\tfrac12r_{xx}-rr_x$, and the diffusion has spatial generator
$\tfrac12\partial_{xx}+r\partial_x$. Differentiating in space and
applying Itô's formula gives the drift and diffusion pairs
$(0,v)$, $(-v^2,w)$ and $(-3vw,z)$ for $r,v,w$, respectively.
All expectations in this section evaluate spatial functions at
$(t,X_t)$. Starting from $c^2q'=\mathbb E v^2$, Itô's formula gives
$$
\begin{aligned}
c^2q''&=\mathbb E[w^2-2v^3],\\
c^2q'''&=\mathbb E[z^2-12vw^2+6v^4]
       =c^2\mathcal C_\gamma(t).
\end{aligned}                                      \tag{4}
$$
For example, the drifts of $w^2$ and $v^3$ are
$z^2-6vw^2$ and $3vw^2-3v^4$, respectively.

The negative mixed term is the obstacle, and convexity, which gives
$v\geq0$, does not control it. Take, for instance, the smooth auxiliary
terminal value $\log(2\cosh x)$ and the coefficient $\gamma\equiv1$.
The solution is $\Phi(t,x)=\log(2\cosh x)+(1-t)/2$, so $r=\tanh x$,
and at $r^2=1/2$ the polynomial in (4) equals $-19/8$.
This example is a smoothed finite-step problem, and its coefficient is
no minimiser of the absolute-value terminal problem. The calculation
is in
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="curvature-moments">[long §4.1](../long/reasoning-record.md#curvature-moments)</span>,
and that section also finds a negative pointwise value
with the actual $|x|$ terminal datum for a small positive constant
coefficient
(<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="pointwise-curvature-obstruction">[long §4.1, pointwise obstruction](../long/reasoning-record.md#pointwise-curvature-obstruction)</span>).

The proof therefore subtracts a weighted divergence whose expectation
under the law of $X_t$ is zero, and what remains has a positive
fourth-moment remainder. For this decomposition to have the required
signs, the backward value function and the forward density must satisfy
joint shape restrictions. Those comparisons and the exact polynomial
identity are the substantial inputs from [OpenAI (2026a), §§3–4][FS].
We explain their roles first for finite-step coefficients and then pass
the integrated inequality to an integrable coefficient by approximation.

We begin with a positive nondecreasing finite-step coefficient
$\gamma_{\rm step}$, that is, a function constant on each interval of a
finite partition, and we replace the terminal value $|x|$ by the smooth
approximation $M^{-1}\log(2\cosh Mx)$, with
$M\geq\|\gamma_{\rm step}\|_\infty$.
On a step of height $c$, set $h=e^{c\Phi}$. Then
$h_t=-h_{xx}/2$ and $r=h_x/h$, and the same exponential removes the
drift from the forward density equation. Indeed, writing $f=hg$ gives
$$
\begin{aligned}
f_t&=\tfrac12(hg)_{xx}-(h_xg)_x\\
   &=-\tfrac12h_{xx}g+\tfrac12hg_{xx}
    =h_tg+\tfrac12hg_{xx}.
\end{aligned}
$$
Comparison with $f_t=h_tg+hg_t$ yields $g_t=g_{xx}/2$. One backward
heat factor and one forward heat factor thus determine the law against
which the moment polynomial is integrated.

For $s=-\partial_x\log g$, the source proves joint shape restrictions
on $x>0$: $r_x>0$ and $s_x\geq1/t$, both $r$ and $s$ are concave,
and $H:=w/(2r)+v\leq0$ is nondecreasing. They also give
$$
j:=s_x-(s/r)v=r\,\partial_x(s/r)\geq0.
$$
This last inequality couples the backward gradient to the forward
density, and separate monotonicity and concavity do not supply it.
On the first step $s=x/t$, so concavity of $r$ gives $j\geq0$, and a
parabolic comparison propagates the sign within each step.
At a forward jump with $k=c_+/c_-\geq1$, continuity of $\Phi$ and $f$
gives
$$
r_+=kr_-,\qquad s_+=s_-+(k-1)r_-,\qquad
j_+=(s_x)_+-\frac{s_+}{r_+}(r_x)_+=j_-.
$$
The joint sign therefore survives jumps of the coefficient as well.
The quotients $w/r$ and $s/r$ extend smoothly through zero because
$r,s,w$ are odd and $r_x(0)>0$. Both the boundary controls of the
comparison and these extensions are analytic inputs from
[OpenAI (2026a), Lemmas 3.2–3.4][FS], reconstructed in <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="shape-comparisons">[long §4.3](../long/reasoning-record.md#shape-comparisons)</span>.
This ratio sign and its preservation at jumps also follow from
[Zhou (2026), Appendix E.3, (E.37) and (E.41)][zhou-gardner], after
rescaling the smoothed finite-step problem. Section 4.3 of the
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="shape-comparisons">[long record](../long/reasoning-record.md#shape-comparisons)</span> gives the
change of variables and the resulting inequality.



The factorisation of the density gives $f_x/f=r-s$. For a suitable
expression $F$ in the shape quantities, define
$$
\mathcal W(F)=r\partial_xF+(v+r^2-rs)F.
$$
Then $f\mathcal W(F)=\partial_x(rfF)$. The smooth extensions
avoid an interior boundary term at zero, and the source's differentiated
Gaussian tail estimates give $rfF\to0$ at both spatial infinities
for the polynomial correction used below. Hence
$\mathbb E_f\mathcal W(F)=0$, where $\mathbb E_f$ denotes integration
against $f(x)\,dx$.

What remains is algebra under these analytic restrictions. The
correction polynomial is chosen in two steps: first arrange a square
and other nonnegative terms containing the highest derivative $z$, then
cancel their terms linear in $z$ with the weighted divergence. Only
lower derivatives and the shape quantities then remain in the residual.
[OpenAI (2026a), Lemma 4.1][FS] supplies an explicit polynomial $F_2$
and a decomposition
$$
z^2-12vw^2+6v^4
=\mathcal W(F_2)+\kappa v^4+\mathcal R,
\qquad
\mathcal R\geq0,\qquad
\kappa=\frac{1661507}{1521000000}>\frac1{1000}.
$$
The coefficients of $F_2$ have both signs. Since the correction
integrates to zero, the signs of $F_2$ and $\mathcal W(F_2)$ play no
role, whereas the shape inequalities make every term of $\mathcal R$
nonnegative. Two of its actual terms are
$$
\frac{337}{10000}\,r^2j\,(-w/r-7v)^2,
\qquad
(2rH_x)v\left(\frac{2331}{10000}r^2+\frac{87}{2500}j\right).
$$
The first needs the joint comparison $j\geq0$, and the second also
uses $H_x\geq0$ on the positive half-line. Reflection preserves
these signs on the negative half-line. The remaining terms use
$H\leq0$, forward concavity and the other shape bounds, so the
positivity of the decomposition depends on the shapes the PDE supplies.
Integration removes $\mathcal W(F_2)$ and keeps the fourth-moment
margin, giving
$$
\int_{\mathbb R}(z^2-12vw^2+6v^4)f\,dx
\geq\frac1{1000}\int_{\mathbb R}v^4f\,dx.           \tag{5}
$$
We use (5) as an imported analytic estimate. The full polynomial and
its decomposition are in [Lemma 4.1 and Table 1][certificate], and
Proposition 4.2 integrates them using the signs and zero boundary
flux supplied by Lemma 3.4. The definitions and the complete identity
are reproduced in <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="curvature-algebra">[long §4.4, “The polynomial correction with zero expectation”](../long/reasoning-record.md#curvature-algebra)</span>.

To obtain Lemma 2 for an integrable coefficient, choose positive
monotone step coefficients $\gamma_n$ that keep the plateau near a
fixed interior time, with $\|\gamma_n-\gamma\|_1\to0$, and terminal
parameters $M_n\geq\|\gamma_n\|_\infty$ with $M_n\to\infty$.
Spatial derivatives through order four then converge uniformly on a
fixed strip below one, and the diffusions driven by the same Brownian
motion converge
(see [OpenAI (2026a), Lemmas 2.1–2.2 and Proposition 5.1][FS]). Only
$v,w,z$ remain in (5), so the auxiliary logarithmic derivative $s$
needs no convergence theorem. Dividing by $c^2$ gives (3), and
strictness follows from $\Phi_{xx}>0$ on the plateau (see [OpenAI (2026a), Lemma 2.3][FS]).
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="plateau-limit">[Long §4.5](../long/reasoning-record.md#plateau-limit)</span>
specifies the plateau-preserving approximants and their common strip
bounds.

The zero plateau has the separate heat identity stated in Lemma 2. Its
strictness uses a positive diffusion density and the fact that
$\Phi_{xxx}$ cannot vanish identically, since an even quadratic profile
would contradict the one-Lipschitz bound and $|x|+O(1)$ growth.

**Passage to a fixed time.** Now let $\gamma\in\mathcal U$ be continuous
at $t\in(0,1)$, with $c=\gamma(t)>0$. For
$0<\delta<\min(t,1-t)$ define
$$
\gamma_\delta(s)=
\begin{cases}
c,&t-\delta\leq s<t+\delta,\\
\gamma(s),&\text{otherwise}.
\end{cases}
$$
The inequalities $\gamma(t-\delta)\leq c\leq\gamma(t+\delta)$
make both joins nondecreasing, so $\gamma_\delta\in\mathcal U$, although
it need not minimise $\mathcal P$. Moreover,
$$
\|\gamma_\delta-\gamma\|_1
\leq2\delta\,[\gamma(t+\delta)-\gamma(t-\delta)]
\longrightarrow0.
$$
We use a subscript $\delta$ for its value-function derivatives and
diffusion. Lemma 2 and (4), already extended to integrable
coefficients, give at the fixed time $t$
$$
\mathcal C_{\gamma_\delta}(t)
=\mathbb E[d_\delta^2-12c\,a_\delta b_\delta^2+6c^2a_\delta^4]
\geq\frac{c^2}{1000}\mathbb E[a_\delta^4].           \tag{6}
$$

We now pass to the limit of shrinking plateaus and choose $t<T<T_1<1$. Monotonicity gives the
common coefficient bound
$$
\sup_{s\leq T_1}\gamma_\delta(s)
\leq\frac{\|\gamma_\delta\|_1}{1-T_1}.
$$
This is the first uniform estimate, and the bounds on spatial
derivatives require the smoothing theorem. Lemma 2.1 of
[OpenAI (2026a)][FS] gives uniform convergence in space and uniform
bounds for $a_\delta,b_\delta,d_\delta$ on $[0,T]$, with the next
spatial derivatives bounded as well. By Lemma 2.2, the diffusions
coupled through the same Brownian motion satisfy
$\sup_{s\leq T}|X_s^\delta-X_s|\to0$. The derivative bounds control
evaluation along these paths, so the integrands in (6) converge and
are uniformly bounded, and dominated convergence passes their
expectations to the limit. These estimates depend only on the common
$L^1$ bound on the coefficient and the fixed distance from time one,
whatever the width of the shrinking plateau. The result is
$\mathcal C_\gamma(t)\ge\gamma(t)^2\mathbb E a^4/1000$,
which proves Proposition 3. Neither a time derivative of the
coefficient nor minimality has entered the limit.

For a bounded gap, a qualitative crossing criterion already suffices:
if $q''=0$ implies $q'''>0$, then $q'$ cannot attain an interior
maximum, so the endpoint slope bounds give $q'<1$ inside the gap and
the same strict area deficit. Proposition 3 supplies more for the next
argument. It applies at the chosen comparison time without assuming
$q''=0$, and its fourth-moment lower bound survives the local
replacement, where bare strict positivity could be lost in the limit.
<!-- /passage:curvature -->

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

## 4. Strict positivity of the density

<!-- passage:density-consequence -->
Full support alone does not force a pointwise positive density.
For instance, $\gamma(t)=\int_0^t(s-\tfrac12)^2\,ds$ has full Stieltjes
support but $\gamma'(\tfrac12)=0$: every interval of positive length
has a positive integral, although the integrand vanishes at one point.
The quantitative remainder in (5) rules this out for a minimiser.
The identity $\mathbb E a^2=1$ first expresses $\gamma$ as a quotient
of continuous moments and then gives smoothness. Only afterwards do we
apply the fixed-time spatial estimate to obtain $\gamma'>0$.

**Corollary 4 (smooth positive density; [OpenAI (2026a),
Corollary 6.4][FS]).** Under the hypotheses of Theorem 1,
$\gamma(0)=0$ and $\gamma\in C^\infty([0,1))$, with derivatives
at zero taken from the right. Its Stieltjes measure is
$\mu_\gamma(dt)=\gamma'(t)\,dt$, and
$$
2\gamma'(t)\mathbb E[a(t,X_t)^3]
\geq\frac{\gamma(t)^2}{1000}\mathbb E[a(t,X_t)^4]>0
\qquad(0<t<1).                                    \tag{7}
$$
Only nonnegativity of the density at zero is asserted.

**Proof.** All spatial factors in the expectations below are evaluated
at $(t,X_t)$. Since $da=-\gamma a^2\,dt+b\,dB$, Itô's formula applied
to $a^2$ gives, for almost every $t$,
$0=\frac{d}{dt}\mathbb E a^2=\mathbb E b^2-2\gamma\mathbb E a^3$,
where the left side vanishes because full support gives
$\mathbb E a^2=1$. Solving for the coefficient yields
$$
\gamma(t)=\frac{\mathbb E b^2}{2\mathbb E a^3}.      \tag{8}
$$
At first (8) holds almost everywhere. Its denominator is at least
two, because $a\geq0$ and
$\mathbb E a^3\geq(\mathbb E a^2)^{3/2}=1$.
The moments are continuous, so right continuity of $\gamma$ extends
(8) to every time and proves that $\gamma$ is continuous. Evenness
gives $b(0,0)=0$, hence $\gamma(0)=0$.

For smoothness, fix $T<1$ and consider the polynomial moments
$$
M_P(t)=\mathbb E P(\Phi_x(t,X_t),\ldots,\partial_x^m\Phi(t,X_t)),
$$
where $P$ is a polynomial in any finite number $m$ of variables.
Spatial differentiation of the PDE and Itô's product rule give
polynomials $Q_P,R_P$, involving finitely many further spatial
derivatives, such that
$$
M_P(t)-M_P(0)=\int_0^t
 [M_{Q_P}(s)+\gamma(s)M_{R_P}(s)]\,ds.
$$
The identity involves $\gamma$ and none of its derivatives. Bounded
spatial derivatives on the fixed strip make every moment continuous.
The integral identity then makes every moment $C^1$, and (8) makes
$\gamma$ $C^1$. Repeating the argument proves smoothness. This is the
quotient-and-moment argument of [Chen (2026), Proposition 4.6
and Appendix B.6][Chen], which uses the polynomial-moment method of
[Auffinger and Chen (2015), Lemma 2 and the proof of Theorem 2(ii)][AC-properties].
[OpenAI (2026a), proof of Corollary 6.4][FS]
and <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="smooth-density-consequence">[long §9.1](../long/reasoning-record.md#smooth-density-consequence)</span>
give the explicit polynomials and the passage from initially nonsmooth
coefficients.

The two differentiated moments are now
$$
\frac{d}{dt}\mathbb E b^2=\mathbb E[d^2-6\gamma ab^2],
\qquad
\frac{d}{dt}\mathbb E a^3=\mathbb E[3ab^2-3\gamma a^4].
$$
For a smooth time coefficient, differentiating
$q''=\mathbb E b^2-2\gamma\mathbb E a^3$ therefore gives
$$
q^{(3)}+2\gamma'\mathbb E a^3
=\mathcal C_\gamma(t).                              \tag{9}
$$
This identity displays the two uses of the same spatial moment.
On a plateau, $\gamma'=0$ and $(v,w,z)=c(a,b,d)$ recover (4).
For the now smooth minimiser $q=t$, so $q^{(3)}=0$ and the left
side is $2\gamma'\mathbb E a^3$, while Proposition 3 bounds the right
side from below.

We now apply Proposition 3 to the smooth minimiser. Strict increase and
$\gamma(0)=0$ give $\gamma(t)>0$ for $t>0$, so (9) gives (7).
Strictness follows from $\gamma(t)>0$ and
$\mathbb E a^4\geq(\mathbb E a^2)^2=1$.
In particular, Cauchy–Schwarz gives
$$
\mathbb E a^3\leq\sqrt{\mathbb E a^2\,\mathbb E a^4}
=\sqrt{\mathbb E a^4}\leq\mathbb E a^4,
$$
so the same inequality implies
$\gamma'(t)\geq\gamma(t)^2/2000$ for $0<t<1$.

Proposition 3 passes spatial moments to the limit, and its proof does
not identify $q_\delta'''$ with the third derivative of the limiting
moment. Indeed, the minimiser has $q(t)=t$ and $q'''=0$, and the
coefficient derivative in (9) is exactly the term that the
constant-interval formula omits. <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="smooth-density-consequence">[Long §§9.1–9.2](../long/reasoning-record.md#smooth-density-consequence)</span> retain the smoothness argument and the
fixed-time transfer, including the common approximation bounds.

**Earlier and related estimates.** The backward shape comparisons,
the transformed density and the positive-temperature gap-crossing
argument appear in [Lopatto (2026), v3, Propositions 3.6, 7.1 and 9.1][Lopatto].
[H.-B. Chen (2026), Proposition 4.2][Chen] proves the zero-temperature
crossing implication $q''=0\Rightarrow q'''>0$ under the minimiser
convention of its §3.1. [Lopatto's v4, Proposition 5.1 and
Corollary 5.6][Lopatto-v4] give qualitative crossing for arbitrary
positive-temperature probability measures. In unit-noise coordinates,
the cumulative coefficient there is $\gamma_\beta=\beta\alpha_\nu$,
where $\alpha_\nu(t)=\nu([0,t])$, and the terminal value is
$\beta^{-1}\log\cosh(\beta x)$. Every bounded $\gamma\in\mathcal U$
has this form: choose $\beta>0$ with $\beta\geq\sup\gamma$
and put the remaining probability mass at one. The coefficient classes
alone therefore do not distinguish the estimates. What distinguishes
them is the terminal datum and the quantitative control through its
zero-temperature limit.

[Lopatto's external-field paper, Theorem 2.3 and §8][Lopatto-field]
also uses a fixed-time spatial inequality for support and positive
density. At a continuity point, monotonicity supplies its one-sided
coefficient bounds, and on a smooth full-support interval of the
minimiser its scalar conditions become equalities. Its Appendix B gives
the positive-temperature, zero-field strict-density result. The bound
used here instead applies at a nonstationary comparison time and keeps
an explicit fourth-moment margin through the local replacement with
terminal datum $|x|$.
The rescaling and the exact comparison of hypotheses are in
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="finite-temperature-dictionary">[long §2.3, finite-temperature dictionary](../long/reasoning-record.md#finite-temperature-dictionary)</span>.
<!-- /passage:density-consequence -->

<a id="value-and-qaoa"></a>

## 5. Consequences for optimisation

<!-- passage:scalar-value -->
### 5.1 A martingale formula for the ground-state value

The support identities give a scalar representation of $P_*$. Set
$M_t=u(t,X_t)$, the bounded gradient martingale, which satisfies
$dM_t=a(t,X_t)\,dB_t$. Its quadratic variations distinguish the two
Hessian moments:
$$
d\langle M\rangle_t=a(t,X_t)^2\,dt,
\qquad d\langle M,B\rangle_t=a(t,X_t)\,dt.
$$
Full support gives $\mathbb E a(t,X_t)^2=1$, so each increment
$M_t-M_s$ has unconditional variance $t-s$. This normalises the
discrete increments below. The first moment of $a$ gives instead the
covariance with the driving noise. Let $U$ be the almost-sure and
$L^2$ limit of $M_t$. The value identity is
$$
P_*=\mathbb E[UB_1]
   =\int_0^1\mathbb E[a(t,X_t)]\,dt.                \tag{10}
$$
Boundedness gives $|U|\leq1$, while $\mathbb E M_t^2\to1$ gives
$\mathbb E U^2=1$, so $U^2=1$ almost surely.

The cancellation behind (10) runs as follows. Terminal convergence of
the convex value functions gives $UX_1=|X_1|$ (the equality is
automatic on $X_1=0$), and Itô's formula gives
$\mathcal P(\gamma)=\mathbb E|X_1|-\int_0^1t\gamma(t)\,dt$.
Multiplying $X_1=B_1+\int_0^1\gamma(t)M_t\,dt$ by $U$ cancels
that integral, because $\mathbb E[UM_t]=t$, and the Parisi formula
then gives (10). [OpenAI (2026a), Corollary 7.1][FS] and the long
record, <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="scalar-value-identities">[§6, “Recovering the value from a martingale”](../long/reasoning-record.md#scalar-value-identities)</span>, prove the endpoint limits and the
stochastic-integral covariance. The argument uses only full
support. Moreover, convexity and
Cauchy–Schwarz give $0\leq\mathbb E a\leq1$, so
$$
0\leq P_*-\int_0^T\mathbb E[a(t,X_t)]\,dt\leq1-T.
$$

We can therefore fix the terminal cutoff before discretising the
diffusion.

The gradient-martingale and value methods have antecedents in
[Montanari, preprint v2, Lemmas 3.2–3.4][Montanari], under its
positive-temperature support assumption, and in the zero-temperature
calculations of [El Alaoui, Montanari and Sellke, preprint v1,
Lemmas 6.5 and 6.13][EAMS]. [Sellke, preprint v4,
Proposition 3.10 and Lemma 3.11, p. 17][Sellke] gives the closely
related finite-cutoff cancellation by applying Itô's formula to $\Phi$
and $x\Phi_x$.

### 5.2 The classical algorithm

Full support and existence give an attained, strictly increasing
minimiser in the monotone zero-temperature class, which is Assumption 2
of [El Alaoui, Montanari and Sellke, preprint v1][EAMS]. Their
Corollary 2.2 then gives, for every fixed $\varepsilon>0$, a randomised
algorithm that returns a configuration within $\varepsilon n$ of the
finite SK maximum with probability tending to one over couplings and
algorithmic randomness. It costs $C(\varepsilon)n^2$ real-arithmetic
operations when the prescribed scalar functions and normalisers are
available. This count includes no separate bound for numerical
preprocessing or finite precision.
[OpenAI (2026a), Corollary 7.3][FS] and <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="classical-algorithm-consequence">[long §9.4](../long/reasoning-record.md#classical-algorithm-consequence)</span>
work out the Gaussian convention and the operation count. Chen already
notes the classical consequence in [§1, p. 4][Chen].

The incremental approximate message-passing iterates of El Alaoui,
Montanari and Sellke use the coupling matrix, and their scalar
description uses the martingale multiplier and state drift
$$
u_{\rm EAMS}(t,x)=a(t,x),\qquad
v_{\rm EAMS}(t,x)=\gamma(t)u(t,x).
$$
Their noise multiplier is thus the Hessian, whereas our $u$ is the
gradient. In our notation, the bounded martingale $M_t$ and $\mathbb E a^2=1$ give
the required variance normalisation. With the fixed-strip regularity
and discretisation estimates, [their Theorem 2][EAMS] achieves energy
arbitrarily close to the truncated first-moment integral in (10).

Strict positivity of the density is unnecessary for this algorithm.

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

<!-- passage:qaoa-boundary -->
### 5.3 A fixed QAOA circuit before the thermodynamic limit

The remaining consequence uses a separate circuit approximation theorem,
stated as Proposition 5 below. We give its finite Gaussian input and
identify the energy it approximates. The companion develops its operator
and control construction, and the support and density proofs above
need none of those additional prerequisites.

Let $\widehat Z_i,\widehat X_i$ be Pauli operators, and define the cost
and mixer by
$$
C_{n,J}=\frac1{\sqrt n}\sum_{i<j}J_{ij}\widehat Z_i\widehat Z_j,
\qquad B_n=\sum_i\widehat X_i.
$$
The eigenvalues of the cost are $H_n(\sigma)$. QAOA alternates cost
evolution and mixing, starting from the uniform state
$|+\rangle^{\otimes n}$ ([Farhi, Goldstone and Gutmann (2014)][FGG]).
For a fixed sequence
$W=(\theta_1,\beta_1,\ldots,\theta_p,\beta_p)$, write
$$
|\psi_{n,J}(W)\rangle
=e^{-i\beta_p B_n}e^{-i\theta_p C_{n,J}}\cdots
 e^{-i\beta_1 B_n}e^{-i\theta_1 C_{n,J}}|+\rangle^{\otimes n},
$$
$$
v(W)=\lim_{n\to\infty}\frac1n\mathbb E_J
\langle\psi_{n,J}(W)|C_{n,J}|\psi_{n,J}(W)\rangle.
$$
[Farhi, Goldstone, Gutmann and Zhou (2022)][FGGZ] established this
fixed-angle limit. [OpenAI (2026b), Theorem 1.1][Q] proves that for
every $\varepsilon>0$ there is one finite $W$ with
$v(W)\geq P_*-\varepsilon$, whose depth and angles depend on
$\varepsilon$ but not on $n$ or $J$. The instance still determines
$C_{n,J}$. Here depth counts global cost–mixer layers, whose
elementary gate counts depend on the implementation.

This order of limits differs from the fixed-instance depth limit in
[Farhi, Goldstone and Gutmann, §VI][FGG], which permits
instance-dependent choices. [Basso, Farhi, Marwaha, Villalonga and
Zhou (2022), §7, equation (7.2)][BFMVZ] conjectured the common-circuit
conclusion.

**The finite Gaussian input.** Fix the minimiser and its diffusion,
a cutoff $0<T<1$, and a mesh $t_j=j\Delta$ with $\Delta=T/N$.
The mesh count $N$ is distinct from the number of spins $n$.
From independent standard normals $Z_1,\ldots,Z_{N+1}$, form
$$
\begin{aligned}
Y_0&=0,\\
Y_j&=Y_{j-1}+\sqrt\Delta Z_j+
       \Delta\gamma(t_{j-1})u(t_{j-1},Y_{j-1}),\\
h_j&=\frac{a(t_{j-1},Y_{j-1})}
 {\sqrt{\mathbb E a(t_{j-1},Y_{j-1})^2}},\qquad g_j=h_jZ_j
 \quad(1\leq j\leq N).
\end{aligned}
$$
[OpenAI (2026a), Proposition 7.2][FS] proves that the denominators
are positive for all sufficiently large $N$. Since $h_j$ depends
only on $Z_1,\ldots,Z_{j-1}$, the $g_j$ are centred and
orthonormal in $L^2$, although they need not be independent Gaussians.
With $\Phi_{\rm G}$ the standard normal distribution function, set
$$
\begin{aligned}
F_N&=\operatorname{sgn}\{u(T,Y_N)+2\Phi_{\rm G}(Z_{N+1})-1\},\\
\mathbb E[F_N\mid Z_1,\ldots,Z_N]&=u(T,Y_N).
\end{aligned}
$$
The second equality holds because $\Phi_{\rm G}(Z_{N+1})$ is
uniform on $[0,1]$, so this random rounding preserves every correlation
with a function of the preceding coordinates. The circuit construction
needs one finite quantity, the adjacent pairing
$$
\begin{aligned}
S_{T,N}&:=\sum_{j=1}^N
 \mathbb E[F_Ng_j]\,\mathbb E[F_NZ_{j+1}],\\
\lim_{N\to\infty}S_{T,N}
&=I_T:=\int_0^T\mathbb E[a(t,X_t)]\,dt.
\end{aligned}                                      \tag{11}
$$

The same proposition proves this limit. Its proof starts from the
Euler approximation and the gradient martingale, which give
$$
\left\|u(T,Y_N)-\sum_{j=1}^N\sqrt\Delta\,
 a(t_{j-1},Y_{j-1})Z_j\right\|_2\longrightarrow0.
$$
Conditional-mean preservation transfers these correlations to $F_N$
for the first $N$ coordinates. The rounding error itself stays of order
one, since its squared $L^2$ error satisfies
$\mathbb E(F_N-u(T,Y_N))^2\to1-T$. Bessel's inequality, applied
separately to $(g_j)$ and $(Z_j)$, controls the errors of both
correlation vectors in $\ell^2$, and their bounded norms control the
growing pairing by Cauchy–Schwarz. The terms with $j<N$ give a Riemann
sum, and the last term, from the rounding coordinate, vanishes. All
these limits keep $T$ fixed. Only afterwards does $T\uparrow1$ give
$P_*$. The construction and the estimates are in
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="finite-gaussian-approximation">[long §7, “A finite Gaussian spin and its adjacent correlations”](../long/reasoning-record.md#finite-gaussian-approximation)</span>.

**Proposition 5 (circuit approximation; a consequence of
[OpenAI (2026b), §§2–5][Q]).** For each fixed $T,N$ in the preceding
construction with finite positive normalisers, and every $\eta>0$,
there is a finite ordinary QAOA word $W$, with deterministic angles
independent of $n$ and $J$, such that
$$
|v(W)-S_{T,N}|<\eta.
$$

**The adjacent pairing as a rooted energy.** We use the rooted Gaussian
representation of [OpenAI (2026b), §2 and Lemma 3.1][Q], which builds
on the spin–boson description of [Boulebnane et al. (2026)][spin-boson].
Its classical sector consists of square-integrable functions built
recursively from a local Gaussian seed and child-generated Gaussian
coordinates, with inner product given by expectation.
For any finite list of fixed centred real child directions, their
coordinates at the receiving root have a joint centred Gaussian law,
and for two such directions $f,g$,
$$
\mathbb E[\mathsf G(f)\mathsf G(g)]=\langle f,g\rangle.
$$
These directions are fixed elements of the child space, even when their
values depend on the child's Gaussian coordinates. Moreover, the seed of
the receiving root is independent of these coordinates, and compatible
embeddings preserve them as the recursive depth grows.
The root shift $\mathsf L$ inserts a child direction into the first
Gaussian chaos (the linear Gaussian part) with constant receiving-root
state, and its adjoint projects onto that part and returns the child
direction. A
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="depth-one-calibration">[one-pair calculation at the start of long §8](../long/reasoning-record.md#depth-one-calibration)</span>
compares the energy of this representation with the finite SK circuit
before the general construction in §8.1.

We start with a standard normal receiving-root seed $g_0$ and put
$Z_{j+1}=\mathsf G(g_j)$ recursively. The orthonormal child directions
give independent standard Gaussian lifts, which reproduce the scalar
law. Since the finite spin $F_N$ depends only on those lifts, projection
gives the exact rooted-energy identity
$$
\begin{aligned}
\mathsf L^*F_N&=\sum_{j=0}^N\mathbb E[F_NZ_{j+1}]g_j,\\
\mathcal E(F_N):=\operatorname{Re}\langle F_N,\mathsf L F_N\rangle
&=\sum_{j=0}^N\mathbb E[F_Ng_j]\,\mathbb E[F_NZ_{j+1}]
 =S_{T,N}.
\end{aligned}                                     \tag{12}
$$
Independence of the receiving-root seed removes the $j=0$ term.
This identity concerns the finite Gaussian formula in a limiting
rooted space and makes no statement about a finite-$n$ circuit. An
unshifted pairing can have the same scalar limit without identifying
this energy.
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="rooted-energy">[Long §8.1](../long/reasoning-record.md#rooted-energy)</span>
proves the construction and the projection.

The shift carries the lift from a child direction $g_j$ to the receiving
coordinate $Z_{j+1}$, and a one-step example separates it from the
unshifted pairing. For the nonminimising heat coefficient $\gamma=0$,
take $T=4/5$ and $N=1$. Then $g_1=Z_1$ and
$F_1=\operatorname{sgn}(2Z_1+Z_2)$, and Gaussian regression gives
$$
S_{T,1}=\frac4{5\pi},\qquad
\mathbb E[F_1g_1]\,\mathbb E[F_1Z_1]=\frac8{5\pi}.
$$
The example tests only the finite representation, since the limit in
(11) depends on the minimiser. The calculation is in
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="one-step-example">[long §10.4, “An exact one-step calculation”](../long/reasoning-record.md#one-step-example)</span>.

**The construction behind Proposition 5.**
[OpenAI (2026b), §§3–5][Q] smooths the spin and approximates every
transmitted function by a bounded smooth map that preserves oddness.
Smoothing the final sign alone would be insufficient. The class
exclusions of the construction prevent return through a traversed
branch and reserve helper classes on which the output is zero, with
positive probabilities fixed during synthesis. This finite formula is
then approximated in energy using temporary Gaussian controls,
with the same sitewise seeds throughout the word. Removing those
controls produces an ordinary word, whose limiting energy is $v(W)$ by
the fixed-word transfer. The approximation errors are absolute and can
be chosen successively for each fixed $S_{T,N}$, before it is chosen
close to $P_*$. Energy is the preserved quantity, and the conclusion
about a measured configuration below uses a further concentration
theorem. The complete error bound is in
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="finite-formula-energy-error">[long §8.6, finite-formula energy approximation](../long/reasoning-record.md#finite-formula-energy-error)</span>,
and the constructions and their hypotheses are in
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="qaoa-construction">[long §§8.1–8.6](../long/reasoning-record.md#qaoa-construction)</span>.

Near-optimality follows from the decomposition
$$
P_*-v(W)\leq
 (P_*-I_T)+|I_T-S_{T,N}|+|S_{T,N}-v(W)|.            \tag{13}
$$
Choose $T$ so that $1-T<\varepsilon/3$, then $N$ so that the
second term is below $\varepsilon/3$, and finally apply Proposition 5
with $\eta<\varepsilon/3$. These finite choices give
$v(W)>P_*-\varepsilon$ before $n$ tends to infinity. Appendix A of the
source supplies an alternative positive-temperature input to the same
synthesis.

The construction gives no bound on the depth or on the cost-angle
magnitudes in terms of the accuracy, and no efficient procedure for
finding the angles. A separate estimate for the expected energy of a
specified word appears in <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="finite-size-comparison">[long §11.5](../long/reasoning-record.md#finite-size-comparison)</span>.
Its threshold depends on the depth and cost angles of the word, so
without bounds on those parameters it gives no threshold in terms of
optimisation accuracy alone. It also gives no rate for the
concentration of measurements.

**Corollary 6 (a measured configuration on the same instance).**
For every $\varepsilon>0$ there is a finite ordinary QAOA word
$W_\varepsilon$, chosen independently of $n$ and $J$, such that one
computational-basis measurement $\Sigma$ satisfies
$$
\Pr_{J,\Sigma}\!\left\{
\max_\sigma H_{n,J}(\sigma)-H_{n,J}(\Sigma)\leq\varepsilon n
\right\}\longrightarrow1
\qquad(n\to\infty).                               \tag{14}
$$

**Proof.** For a fixed word, put
$Y_n=H_{n,J}(\Sigma)/n$ and
$K_n=\max_\sigma H_{n,J}(\sigma)/n$.
[Farhi, Goldstone, Gutmann and Zhou, §4, equations (13)–(22)][FGGZ]
prove
$$
\mathbb E_{J,\Sigma}Y_n\to v(W),\qquad
\mathbb E_{J,\Sigma}Y_n^2\to v(W)^2.
$$
The second moment is $\mathbb E_J\langle(C_{n,J}/n)^2\rangle$,
so it includes measurement variance as well as disorder fluctuations.
Thus $Y_n\to v(W)$ in joint $L^2$. Gaussian concentration of the
maximum and its defining mean limit give $K_n\to P_*$ in probability.
Choose $W$ with $P_*-v(W)<\varepsilon/3$. With probability tending
to one, both $K_n-P_*<\varepsilon/3$ and
$v(W)-Y_n<\varepsilon/3$, which proves (14). The two energies use
the same $J$, and the argument needs no independence between them.
The concentration argument and its further probability consequences
are in the companion's
<span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="fixed-word-probability">[§11.2, “What is averaged, and a fixed-word probability consequence”](../long/reasoning-record.md#fixed-word-probability)</span>.
<!-- /passage:qaoa-boundary -->

<a id="evidence-scope"></a>

## Formal verification

<!-- passage:evidence-scope -->
The proofs above are ordinary mathematical arguments that import three
analytic inputs: the integrated estimate (5), the shape comparisons
behind it and the compact-strip approximation. OpenAI's formalisation
in [Lean][Lean4] with [mathlib][mathlib] contains two formal statements
that bear on this paper. The
[formal statement of full support][support-challenge] asserts the
support and moment conclusions of Theorem 1, with an admissible
minimiser as its hypothesis, and constructs the associated diffusion.
Lean checks OpenAI's [proof of the corresponding theorem][support-lean].

A separate
[formal statement of the scalar consequences][value-challenge] covers
the value identity and the finite Gaussian approximation, with
minimiser and diffusion premises for those clauses, and its third
clause identifies the ground-state and Parisi values from a minimiser,
constructing the diffusion internally. Its [proof][value-lean] has not
been checked, so those results rest here on the paper proofs cited in
§5. The two formal statements use separate models of the order
parameter, which have not been related to each other. Neither covers
Corollary 4 or the convergence of QAOA. The long record compares each
formal statement with the printed one in <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="formal-interfaces">[§12, “The formal statements and their scope”](../long/reasoning-record.md#formal-interfaces)</span>, and gives versions and checking outcomes in <span class="pdf-companion" data-pdf-target="../long/reasoning-record.pdf" data-pdf-anchor="verification">[§13, “Source checks and finite calculation”](../long/reasoning-record.md#verification)</span>.
<!-- /passage:evidence-scope -->

## References

- Auffinger, A. and Chen, W.-K. (2015). “On properties of Parisi
  measures.” *Probability Theory and Related Fields* **161**, 817–850.
  [Preprint v1][AC-properties]: Lemma 2 and proof of Theorem 2(ii).
- Auffinger, A. and Chen, W.-K. (2017). “Parisi formula for the ground
  state energy in the mixed $p$-spin model.” *Annals of Probability*
  **45**, 4617–4631. [Preprint v2][AC]: Theorem 1 and proof of Lemma 3.
- Auffinger, A., Chen, W.-K. and Zeng, Q. (2020). “The SK model is
  infinite step replica symmetry breaking at zero temperature.”
  *Communications on Pure and Applied Mathematics* **73**, 921–943.
  [arXiv:1703.06872v2][ACZ]: Theorem 1, Proposition 1 and Remark 7.
- Basso, J., Farhi, E., Marwaha, K., Villalonga, B. and Zhou, L. (2022).
  “The Quantum Approximate Optimization Algorithm at High Depth for
  MaxCut on Large-Girth Regular Graphs and the Sherrington–Kirkpatrick
  Model.” *TQC 2022*, 7:1–7:21. [Full preprint v3][BFMVZ]: transfer
  theorem and §7, equation (7.2); the full preprint has separate numbering.
- Boulebnane, S., Khan, A., Liu, M., Larson, J., Herman, D., Shaydulin, R.
  and Pistoia, M. (2026). *Spin–Boson Mapping of the Quantum Approximate
  Optimization Algorithm*. [Preprint v2][spin-boson]: Theorem 2
  (informal) and equations (10)–(11).
- Chen, H.-B. (2026). *FRSB in the SK spin glass: convergence to
  full-interval support at zero temperature*. [arXiv:2607.18032v1][Chen],
  20 July 2026. §1, p. 4; Theorem 1.3; Propositions 4.2, 4.5 and 4.6;
  Appendices B.6 and C.
- Chen, W.-K., Handschy, M. and Lerman, G. (2018). “On the energy
  landscape of the mixed even $p$-spin model.” *Probability Theory and
  Related Fields* **171**, 53–95. [Preprint v4][CHL]: Proposition 2,
  Theorem 5, Lemma 3 and Propositions 3–4.
- de Moura, L. and Ullrich, S. (2021). “The Lean 4 Theorem Prover and
  Programming Language (System Description).” *CADE 28*.
  [System description][Lean4].
- El Alaoui, A., Montanari, A. and Sellke, M. (2021). “Optimization of
  mean-field spin glasses.” *Annals of Probability* **49**, 2922–2960.
  [arXiv:2001.00904v1][EAMS]: Theorem 2, p. 6; Remark 2.1 and Theorem 3, p. 7;
  Assumption 2 and Corollary 2.2, p. 8; §§5–6.
- Farhi, E., Goldstone, J. and Gutmann, S. (2014). *A Quantum Approximate
  Optimization Algorithm*. [Preprint v1][FGG]. Algorithm definition;
  §VI concerns the fixed-instance depth limit.
- Farhi, E., Goldstone, J., Gutmann, S. and Zhou, L. (2022). “The Quantum
  Approximate Optimization Algorithm and the Sherrington–Kirkpatrick
  Model at Infinite Size.” *Quantum* **6**, 759. [arXiv:1910.08187v4][FGGZ].
- Jagannath, A. and Tobasco, I. (2016). “A dynamic programming approach
  to the Parisi functional.” *Proceedings of the AMS* **144**, 3135–3150.
  [Preprint v4][JT]. Positive-temperature control methods.
- Lopatto, P. (2026). *Full replica symmetry breaking in the
  Sherrington–Kirkpatrick model*. [arXiv:2607.11756v3][Lopatto], revised
  16 July 2026. Propositions 3.6, 7.1 and 9.1; related Cole–Hopf, density
  and crossing arguments.
- Lopatto, P. (2026). *Full replica symmetry breaking in the
  Sherrington–Kirkpatrick model*. [arXiv:2607.11756v4][Lopatto-v4], revised
  5 October 2026. Proposition 5.1 and Corollary 5.6; qualitative
  crossing for arbitrary probability measures.
- Lopatto, P. (2026). *Parisi measure of the Sherrington–Kirkpatrick
  model with external field*. [arXiv:2610.00127v1][Lopatto-field].
  Theorem 2.3 and Appendix B; related positive-temperature two-use argument.
- The mathlib Community (2020). “The Lean Mathematical Library.”
  *Certified Programs and Proofs (CPP '20)*. [Library description][mathlib].
- Montanari, A. (2025). “Optimization of the Sherrington–Kirkpatrick
  Hamiltonian.” *SIAM Journal on Computing* **54**, FOCS19-1–FOCS19-38.
  [Preprint v2, 2019][Montanari]: §§2–3.
- OpenAI (2026a). *Full support of the zero-temperature Sherrington–Kirkpatrick
  order parameter*. 27 September 2026. [Manuscript][FS], fixed revision.
  Theorem 1.1; §§2–6; Corollaries 6.4 and 7.1; Proposition 7.2.
- OpenAI (2026b). *QAOA attains the SK ground-state energy in the
  thermodynamic-first limit*. 25 September 2026. [Manuscript][Q], fixed
  revision. Theorem 1.1; §§2–6 and Appendix A.
- Sellke, M. (2021). *Optimizing Mean Field Spin Glasses with External Field*.
  [Preprint v4][Sellke], 6 October 2021. Proposition 3.10 and
  Lemma 3.11, p. 17.
- Sherrington, D. and Kirkpatrick, S. (1975). “Solvable Model of a Spin-Glass.”
  *Physical Review Letters* **35**, 1792–1796. [Original model][SK].

- Zhou, Y. (2026). “On the Gardner Transition in the Ising Pure p-Spin Glass II”.
  [arXiv:2608.06523v1][zhou-gardner], Appendix E.3, (E.37), (E.41).

[Lean4]: https://doi.org/10.1007/978-3-030-79876-5_37
[mathlib]: https://doi.org/10.1145/3372885.3373824
[FS]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/main.pdf
[Q]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026/QAOA-attains-the-SK-ground-state-energy-in-the-thermodynamic-first-limit-September-25-2026.pdf
[AC]: https://arxiv.org/abs/1606.05335v2
[ACZ]: https://arxiv.org/abs/1703.06872v2
[Chen]: https://arxiv.org/abs/2607.18032v1
[Lopatto]: https://arxiv.org/abs/2607.11756v3
[support-lean]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKSupport/Main.lean#L35
[support-challenge]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/ComparatorChallenges/SKFullSupport.lean#L81-L92
[support-config]: https://raw.githubusercontent.com/openai/math/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/ComparatorChallenges/SKFullSupport.json
[value-lean]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/OAI/Probability/SKValue/Main.lean#L16
[value-challenge]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/ComparatorChallenges/SKValue.lean#L163
[value-config]: https://raw.githubusercontent.com/openai/math/adc7f1241b42e322a6451854ab7e4b4c146bf78a/lean/ComparatorChallenges/SKValue.json

[SK]: https://doi.org/10.1103/PhysRevLett.35.1792
[JT]: https://arxiv.org/abs/1502.04398v4
[CHL]: https://arxiv.org/abs/1609.04368v4
[AC-properties]: https://arxiv.org/abs/1303.3573v1
[Montanari]: https://arxiv.org/abs/1812.10897v2
[EAMS]: https://arxiv.org/abs/2001.00904v1
[FGG]: https://arxiv.org/abs/1411.4028v1
[FGGZ]: https://arxiv.org/abs/1910.08187v4
[BFMVZ]: https://arxiv.org/abs/2110.14206v3

[certificate]: https://github.com/openai/math/blob/adc7f1241b42e322a6451854ab7e4b4c146bf78a/preprints/Full-support-of-the-zero-temperature-Sherrington-Kirkpatrick-order-parameter-September-27-2026/build/certificate.tex#L117-L199

[Lopatto-v4]: https://arxiv.org/abs/2607.11756v4
[Lopatto-field]: https://arxiv.org/abs/2610.00127v1

[Sellke]: https://arxiv.org/abs/2105.03506v4

[spin-boson]: https://arxiv.org/abs/2505.07929v2

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