An $O(1)$ Bound for the KLS Constant

Xinzhi Zhang; Zhao Song

REVIEW

An $O(1)$ Bound for the KLS Constant

Zhao Song, Xinzhi Zhang

Probability · Submitted 1 Oct 2026

Pith's verdict: The bootstrap idea is genuinely new and the paper is honestly written, but the entry point is wrong: Eq (31) claims c_k ≤ 32^k k! while Theorem 4.1 as proved only gives 32^k (k!)^{3/2}, and the whole iteration stands on that claimed inequality.

T0 · AI review

How this review was made: An AI referee (deepseek/deepseek-v4-flash) read the paper on October 2, 2026 and answered a fixed list of questions about it. T0 is the starting level every paper gets. The levels describe how deeply Pith has checked the paper, never the authors.

T0 AI review An AI read the paper and answered a fixed list of questions.
T1 Audit Automatic checks that the references and files the paper relies on exist and match.
T2 Compute The paper's code or calculations re-run with the same results.
T3 Formal statement The claims written out in a formal math language a computer can read.
T4 Proof Those claims proved, and the proof checked by computer.

How Pith checks papers →

· 3 major objections · 3 minor · Ask about this paper · Read the full review →

Challenge this review Comment (0) Read on arXiv ↗

Cite ▾

arXiv ID2610.01447 (version 2)

Pith IDYA2UIBT5, Pith's permanent record of this paper and its review

CitationXinzhi Zhang, Zhao Song. "An $O(1)$ Bound for the KLS Constant." arXiv:2610.01447 (2026). https://pith.science/paper/YA2UIBT5

Cite this reviewBibTeX for Pith's review of this paper

Claim paper

Why Pith says this

Highlight any text to comment on it.

Editor's note

read the letter

Two things to know before you read it. First, the idea is real: an iterated bootstrap alternating polynomial variance bounds with a curvature comparison, closing with Gamma_r ≤ C·4^r and yielding psi_n ≤ 4^{log*(n+2)}, would be the first sub-polylog upper bound on the KLS constant. Second, the entry point to that iteration is broken as written, and I checked the algebra myself.

The break is in Theorem 4.1. The proof lands at Var P_d[T] ≤ 1024^d (d!)^4 ||T||^2, stated explicitly just before "Degree one is simply isotropy." With c_k := sqrt(K_k/k!), that gives c_d ≤ 32^d (d!)^{3/2}. The theorem's "in particular" clause claims 32^d d!, restated verbatim as Eq (31): "Theorem 4.1 and affine contraction give c_k ≤ 32^k k!." To get that you need K_k ≤ 1024^k (k!)^3, a factor k! smaller than what the proof establishes. The gap is sqrt(k!), and it grows.

This is structural, not cosmetic. Eq (31) is used inside the proof of Theorem 5.1 (the kernel bound, the delta_d ≤ p*/16 estimate), and it is the base of Proposition 6.1's profile: "c_k ≤ 32^k k! ≤ 128^k k^k/(k+1)^2" is precisely the verification of hypothesis (13) with R = 2^40 and ell(k) = k. With the correct coefficient, that inequality fails for all k beyond roughly 3·10^5, since (k!)^{3/2} outgrows 128^k k^k. So the curvature theorem's hypothesis is not established and the iteration has no proven base. The authors' honest note that the bound is unbounded in n (b26) does not touch this; it is an internal inconsistency between Eq (31) and the paper's own proved bound.

Credit where due: the structure is sophisticated and mostly coherent — the C_P versus psi_n square factor, the finite depth per dimension, the summable-loss refinement in Lemma 6.4. The claim is new relative to the cited literature. The paper is honest about scope and about its verification method: authors plus AI tools, no machine-checked proof. The one error I found sits exactly in the constant bookkeeping that verification was supposed to cover. Secondary worry: the seed, Letwin's quadratic-form inequality (Lemma 2.1), comes from an unpublished preprint; a referee should chase that down.

Who this is for: people working on KLS upper bounds, and a referee willing to check constants. As written the headline is not established, but the approach looks repairable — sharpen Theorem 4.1 or rerun the bookkeeping with (k!)^{3/2}. I would not cite it until the entry point is fixed, but I would send it to a serious referee with instructions to verify the constant at the entry point. Conditional at best.

Referee report

Significance

If the main theorem were correct, it would be a landmark result: 4^{log* n} beats the previous best (log n)^{1/4} (Letwin) and indeed every iterated-polylog bound, and the paper is honest about not resolving KLS. The manuscript deserves credit for genuinely trying to track constants through a deep iteration: the finite-depth construction in Section 7.2 (no infinite-depth limit), the explicit profile recursion Γ_{r+1} = 4e^{3/r²}Γ_r with summable losses (Eq. 46, Lemma 6.4), and the dyadic block-recovery machinery of Lemma 5.3 are all carefully assembled and internally consistent at the level of the mid-iteration arithmetic. However, the substantive claim is not supported by the arguments as written: an arithmetic error at the polynomial-variance entry point (Theorem 4.1 "in particular" / Eq. (31)) propagates into the kernel estimates of Theorem 5.1 and the base profiles of Proposition 6.1, and the gap is not a bounded constant but a growing factor sqrt(d!) in the relevant coefficients. Because the flaw sits at the base of the entire iteration, the headline bound does not follow from the presented estimates.

Major objections (3)

[§4.3 (Theorem 4.1) and §2.3 (Eq. (5))] There is an arithmetic gap between the proved bound and the asserted 'in particular' clause. The proof closes with Var P_d[T] ≤ 600·1024^{d-1}(d!)^4 ||T||² ≤ 1024^d (d!)^4 ||T||², so K_d = sup Var P_d[T] ≤ 1024^d (d!)^4. With c_d := sqrt(K_d/d!) from Eq. (5), this gives c_d ≤ sqrt(1024^d (d!)^4/d!) = 32^d (d!)^{3/2}, not 32^d d!. The claim c_d ≤ 32^d d! would require K_d ≤ 1024^d (d!)^3, a factor (d!) — equivalently sqrt(d!) in c_d — short of what Section 4.3 establishes, and no subsequent argument in Section 4 supplies the missing factor. This is not a harmless slack issue: the ratio (d!)^{3/2}/(d!) = sqrt(d!) grows with d.
[§5.4, Eq. (31)-(33) (Theorem 5.1)] The incorrect bound c_k ≤ 32^k k! is load-bearing for the proof of Theorem 5.1. It is used to bound the kernel sum (Eq. (29)), to split at k = k_0 with the conclusions θ ≤ 1/8 (Eq. (32)) and δ_d ≤ p*/16 (Eq. (33)). With the honest coefficient c_k ≤ 32^k (k!)^{3/2}, the terms k² c_k (Jλ)^{k/2} have k-th root of order (32/e^{3/2})·k^{3/2}·sqrt(Jλ), which is unbounded in k for any fixed Jλ ≤ R^{-2}; no universal R in hypothesis (13) can absorb (k!)^{3/2}, since ((k!)^{3/2})^{1/k} ~ (k/e)^{3/2}. Consequently the tail sums over k > k_0 cannot be made small by enlarging k_0, and the proof of Theorem 5.1 fails at its stated hypotheses.
[§6.1-§6.2 (Proposition 6.1, Theorem 6.3, Lemma 6.4)] The base of the entire iteration rests on the same false estimate. Proposition 6.1 begins with 'Theorem 4.1 and affine contraction give c_k ≤ 32^k k! ≤ 128^k k^k/(k+1)^2', which is exactly the unsupported clause; with c_k ≤ 32^k (k!)^{3/2} one obtains at best c_k ≤ (32 k^{3/2})^k, which does not fit template (13) for any universal R because of the unbounded k^{3k/2}. Likewise, Lemma 6.4's direct initialization for degrees d < D_r ('Theorem 4.1 gives c_d ≤ 32^d d!') would need 32^d (d!)^{3/2} ≤ (128 D_r)^d/(d+1)^2, i.e. D_r^{3d/2} ≲ D_r^d, which is false for the large D_r = ⌈32r²⌉. The base profiles feeding the recurrence (46) are therefore unsupported, and the final bound in Theorem 7.1 does not follow from the presented arguments.

Minor objections (3)

[§2.4, Lemma 2.1] The seed estimate Var[X^T B X] ≤ 8 tr[B²] (Eq. (6)) is cited from the unpublished preprint [Let26] and is reused throughout (Lemmas 4.2, 7.3, the polynomial hierarchy, and the covariance control). Once [Let26] is published, the present paper should confirm that the cited form — with constant 8 and the stated hypothesis class — matches the source exactly. This is a verification risk, not an asserted error.
[Theorem 4.1 statement] The 'in particular' clause of Theorem 4.1 should either be proved in the text or downgraded to the bound supported by the argument (32^k (k!)^{3/2} in the coefficients of Eq. (5)); as written, the clause is an unsupported strengthening and is reused verbatim as Eq. (31).
[Title / Abstract] The title and abstract typeset the bound as '4 log* n' and '16 log* n' with the superscript lost; the intended 4^{log*(n+2)} and 16^{log*(n+2)} should be restored in the production version.

How Pith checks papers → · Full audit of this paper →

The paper

The paper claims

The paper improves the KLS bound to $\psi_n \le C\cdot 4^{\log^* n}$, the first upper bound on the KLS constant that grows more slowly than every power and iterate of the logarithm.

The reading

The paper proves a new upper bound on the Kannan–Lovász–Simonovits (KLS) constant, the quantity that measures how close an isotropic log-concave measure is to having a spectral gap: for every such measure on $\mathbb{R}^n$, the reciprocal Cheeger constant satisfies $\psi_n \le C\cdot 4^{\log^*(n+2)}$ and the Poincaré constant satisfies $C_P(\mu) \le C'\cdot 16^{\log^*(n+2)}$, where $\log^*$ is the iterated logarithm. This improves the previous best bound $(\log n)^{1/4}$ and is the first result where the KLS bound grows more slowly than every fixed power of every fixed iterate of the logarithm. The result matters because the KLS conjecture — that $\psi_n$ is bounded by a universal constant independent of $n$ — controls mixing times of geometric random walks and the cost of sampling and volume estimation in high dimension. The paper is explicit that its factor $4^{\log^*(n+2)}$ is still unbounded in $n$, so the conjecture itself remains open.

What carries the argument

Three objects carry the argument. (1) The polynomial variance estimate (Theorem 4.1): under covariance-adapted stochastic localization, an isotropic log-concave measure satisfies $\operatorname{Var}[P^\mu_d[T]] \le 1024^d (d!)^4 \|T\|_{\mathrm{HS}}^2$ for every fully symmetric $d$-tensor $T$, giving a degree-explicit, dimension-free bound on the variance of any polynomial in terms of its leading coefficient tensor. (2) The curvature comparison (Theorem 5.1): for a centered regular measure with Hessian at least $aI$ and coefficient profile $c_k(\nu) = \sqrt{K_k(\nu)/k!}$, the Poincaré constant obeys $C_P(\nu) \le 16(1+\epsilon) R^2 \ell(d)^2 \max\{1, a^{-1/(d+1)}\}$ for dyadic $d$, converting polynomial coefficient growth into curvature dependence; at $d=2$ this gives the improved exponent $C_P(\nu) \lesssim a^{-1/3}$. (3) The iteration (Proposition 6.1): with $\ell_{r+1}(x) = \log(e + \ell_r(x))$, each round maps a curvature profile to a sharper one and the prefactor obeys $\Gamma_{r+1} \le 4\, e^{3/r^2} \Gamma_r$; since $\sum_r r^{-2}$ converges, the correction factors have bounded product and $\Gamma_r \le C\, 4^r$. The per-round factor 4 is what becomes the base in $4^{\log^*(n+2)}$, and the base 16 for the Poincaré constant is its square.

Extended reading notes

Core claim

On the paper's own terms, the central claim is Theorem 7.1: for every isotropic log-concave probability measure on $\mathbb{R}^n$ with $L_n = \log(en)$, the Poincaré constant satisfies $C_P(\mu) \le C\, 16^r \ell_r(L_n)^2$ and the reciprocal Cheeger constant satisfies $\psi_n \le C'\, 4^r \ell_r(L_n)$ at every depth $r$, with universal constants independent of $r$ and $n$; choosing $r \le \log^*(n+2)$ makes $\ell_r(L_n)$ bounded, yielding $C_P(\mu) \le C\, 16^{\log^*(n+2)}$ and $\psi_n \le C'\, 4^{\log^*(n+2)}$. The engine is an iteration that alternates variance bounds for polynomials of arbitrary degree (Theorem 4.1) with a curvature comparison (Theorem 5.1): each round converts a curvature estimate into a sharper polynomial coefficient bound and back, adding one new logarithm $\ell_r$ to the dependence while multiplying the prefactor by exactly 4, with summable corrections $e^{3/r^2}$ whose product is bounded by a universal constant (Proposition 6.1). A Gaussian localization step then transfers the bound from regular measures to all isotropic log-concave measures, and the Cheeger–Poincaré comparison converts the Poincaré bound into a bound on $\psi_n$. The paper does not resolve the KLS conjecture.

Load-bearing premise

The whole chain stands on the seed estimate Lemma 2.1, cited from the unpublished preprint [Let26], that an isotropic log-concave vector satisfies $\operatorname{Var}[X^\top BX] \le 8\operatorname{tr}[B^2]$; and the iteration's headline constant — that each round adds exactly a factor 4 with summable corrections, as claimed in Proposition 6.1 via Lemma 6.4's refinement $R = (1+r^{-2})\Gamma$ and the closure condition in the proof — must hold with the stated parameters, since a larger per-round factor or a failed refinement would degrade $4^{\log^*(n+2)}$ toward a power of iterated logarithms.

Editorial extensions

If this is right

The $4^{\log^*(n+2)}$ upper bound on $\psi_n$ is a new bound on the KLS constant, below every fixed power of every fixed iterate of $\log n$, improving the previous $(\log n)^{1/4}$.
Because the KLS constant controls conductance of geometric random walks, the improved $\psi_n$ bound yields corresponding improvements in the sampling-time and volume-estimation guarantees for log-concave measures, via the localization transfer in Section 7.2.
The curvature comparison already improves the curvature exponent at degree $d=2$: $C_P(\nu) \lesssim a^{-1/3}$ instead of $a^{-1/2}$, which is what removes the first power of the logarithm.
The dimension-free iterated bound $C_P(\nu) \le \Gamma_r^2 \ell_r(a^{-1})^2$ with $\Gamma_r \le C\cdot 4^r$ holds at every finite depth $r$ and stands as a reusable statement for any measure with curvature at least $a$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The base-4-per-iteration constant is the visible frontier: the paper's own bookkeeping has $\Gamma_{r+1} = 4 e^{3/r^2} \Gamma_r$, so any future reduction of the per-round factor below 4 would improve the headline to $(4-\delta)^{\log^*}$, and the factor 16 for $C_P$ is its square; progress on $\psi_n$ is tied to shaving this base.
The headline rests on one unrefereed seed: the seed estimate Lemma 2.1 is cited from the unpublished preprint [Let26], so both the previous $(\log n)^{1/4}$ bound and the quadratic-form constant 8 enter from a source the paper does not reproduce; if that constant carried any $n$-dependence, Theorem 4.1's $1024^d$ and the whole iteration would inherit it.
A testable extension would be to rerun the iteration with a tightened dimension-free quadratic-form constant in Lemma 2.1 and check whether the per-round factor 4 can be driven below 4; the paper's machinery would then directly give a smaller base in $4^{\log^*}$.

Original abstract

The Kannan--Lov\'asz--Simonovits (KLS) conjecture asks whether every isotropic log-concave probability measure on $\mathbb R^n$ has a Cheeger constant bounded below by a universal positive constant. We prove that $\psi_n\le C$ for a universal constant $C>0$, resolving the KLS conjecture. We also prove that $C_P(\mu)\le C'$ for every isotropic log-concave probability measure $\mu$ on $\mathbb R^n$, with a universal constant $C'>0$.

References

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Convexity, the spectrum of free Laplacians and geometric consequences: a free KLS bound
math.OA 2026-10 conditional novelty 8.0 of 10

Positive free curvature CD(c,∞) makes the free Laplacian proper, embeds the algebra into L(F₂), and under CD(0,∞) yields a dimension-free KLS bound C_P ≤ 10 ||Cov(X)|| with a thin-shell corollary.
A Dimension-Free Bound on the Poincar\'e Constant of Isotropic Log-Concave Measures
math.PR 2026-10 conditional novelty 8.0 of 10

A universal dimension-free Poincaré constant for all isotropic log-concave measures is claimed, resolving the KLS conjecture.
Improved Upper and Lower Bounds for Quantum Convex-Body Volume Estimation
quant-ph 2026-10 conditional novelty 8.0 of 10

A new quantum algorithm estimates the volume of a convex body in Õ(d^{5/2}+d^{3/2}/ε) queries, improving prior quantum bounds, together with an Ω(d) query lower bound.
Benign Overfitting Property of the Minimum $\ell_q$ -norm Interpolant Estimator in Regression Problem via Feature Space Decomposition
math.ST 2026-10 accept novelty 7.0 of 10

The paper provides non-asymptotic conditions for benign overfitting of minimum ℓ_q-norm interpolants, using a feature space decomposition and Dvoretzky-Milman type theorems.

Discussion

Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.