Independent review of the Malek subpermanent transform argument

What Lean checks

Complete written all-degree negative-transform proof with all equality cases, two independent internal reviews and exact scalar identity checks. The permanent endpoint is prior work; the named subsidiary is documented in the Minc site's linked survey. No proof-assistant work is claimed.

Review date: 20 November 2026 (America/Chicago). This receipt checks the mathematical argument communicated by the root researcher, including the new capacity-cost estimate, its constants, the transformed upper bound, the small cases, and extension to matrices with zero entries. A separate source audit is required for the historical status and any priority claim.

Statement and normalization

Let n>=2, let A be a nonnegative doubly stochastic matrix, and write

J=1n𝟏𝟏𝖳,X=A−J,τ(A)=nJ−An−1=J−Xn−1.J=\frac1n\mathbf1\mathbf1^{\mathsf T},\qquad X=A-J,\qquad \tau(A)=\frac{nJ-A}{n-1}=J-\frac{X}{n-1}.

The claimed comparison is

σk(A)≥σk(τ(A)),1≤k≤n,\sigma_k(A)\ge\sigma_k(\tau(A)),\qquad 1\le k\le n,

where sigma_k is the sum of all permanents of square submatrices of order k. The degree-one case is an identity. The full permanent case k=n is the previously proved Marcus–Minc theorem. The new argument covers every proper degree 2<=k<n, rather than a bounded collection of orders.

For these proper degrees set

Sn,k=k!(nk)2n−k,αn,k=k(k−1)n(n−1),Δ=αn,k∥X∥F2.S_{n,k}=k!\binom nk^2 n^{-k},\qquad \alpha_{n,k}=\frac{k(k-1)}{n(n-1)},\qquad \Delta=\alpha_{n,k}\|X\|_F^2.

The key new lower bound, checked below, is

σk(A)Sn,k≥exp(5Δ4n2)(n≥5,2≤k<n).\frac{\sigma_k(A)}{S_{n,k}} \ge \exp\!\left(\frac{5\Delta}{4n^2}\right) \qquad(n\ge5,\ 2\le k<n). \tag{1}

Published inputs inspected

The following full primary texts were read for this review.

  • Y. Lavi, The Kopotun Subpermanent Transform Inequality, hexagon:2611.00201v1, dated 11 October 2026. Sections 3.1–3.3 give the polynomial normalization, optimized slices, and root-movement estimate used here. Lemma 2.2 and equation (3.8) give the coefficient consequence. The proposed argument replaces the capacity-cost estimate of Section 4.1 and its subsequent constants; it does not claim the already published Kopotun comparison as new.
  • Y. Lavi, The Marcus–Minc Transform Inequality, arXiv:2609.29262v1, 24 September 2026. Lemma 3.1 supplies the centered subpermanent estimate, including its contraction hypothesis and positivity of sigma_r(I-J). Section 5 gives the rational bound for the full-degree upper constant. Theorem 1.1 supplies the full permanent case and its equality exceptions.

Both inputs were checked at the indicated arguments, not inferred solely from their abstracts.

Polynomial and slice hypotheses

First assume every entry of A is positive. Define

P(z)=ek(Az)(nk)(z1+⋯+znn)n−k.P(z)=\frac{e_k(Az)}{\binom nk} \left(\frac{z_1+\cdots+z_n}{n}\right)^{n-k}.

This is stable, homogeneous of degree n, and has positive coefficients in every degree-n monomial. It satisfies P(1)=1, grad P(1)=1, and has capacity one. Its central coefficient is

[z1⋯zn]P=n!nnσk(A)Sn,k.[z_1\cdots z_n]P=\frac{n!}{n^n}\frac{\sigma_k(A)}{S_{n,k}}.

For a fixed column j, fix all other positive coordinates x_h. Let t_* minimize P(x_1,...,t,...,x_n)/t. Positivity and stability give n strictly negative roots and a unique positive minimizer. The associated probability vectors p=p(x) and p^0=p(1), the variances d=||p-1/n||_2^2, d_0=||p^0-1/n||_2^2, and the cost

F=logP(x1,…,t*,…,xn)t*∏h≠jxhF=\log\frac{P(x_1,\ldots,t_*,\ldots,x_n)}{t_*\prod_{h\ne j}x_h}

satisfy

F≥0,d0=αn,k∥X:j∥22,0≤d0≤1−1n.F\ge0,\qquad d_0=\alpha_{n,k}\|X_{:j}\|_2^2, \qquad 0\le d_0\le1-\frac1n.

If R=log(max_(h≠j) x_h/min_(h≠j) x_h), the inspected root-movement argument gives

d0≤d+pmaxsinh(R/2),pmax≤1n+(1−1/n)d.\sqrt{d_0}\le\sqrt d+\sqrt{p_{\max}}\sinh(R/2), \qquad p_{\max}\le\frac1n+\sqrt{(1-1/n)d}. \tag{2}

The root pairing in that source is legitimate even with repeated roots: the e_k roots are ordered within that factor, while the n-k uniform-factor roots are paired separately. The optimized scale t_* cancels from the range of the logarithmic likelihood ratios. Thus (2) does not assume that the slice optimizer stays at one.

Finally, writing M_j=inf_(x>0)(F+d/2), the scalar capacity step and the remaining Gurvits induction give

σk(A)Sn,k≥exp(maxjMj).\frac{\sigma_k(A)}{S_{n,k}}\ge\exp(\max_j M_j). \tag{3}

This takes a lower bound before the infimum; it does not confuse evaluation at x=1 with a lower estimate for M_j.

The stronger capacity cost

For positive z, put E_r(z)=e_r(z)/binom(n,r). Since Az is majorized by z and e_k is Schur-concave,

Ek(Az)≥Ek(z).E_k(Az)\ge E_k(z).

Newton's inequalities imply that E_(r+1)(z)/E_r(z)<=E_1(z). Multiplying these ratios gives

Ek(z)E1(z)n−k−1≥En−1(z).E_k(z)E_1(z)^{n-k-1}\ge E_{n-1}(z).

Here k<n is essential; at k=n-1 the last comparison is an equality. Using E_(n-1)(z)/prod z=E_1(1/z), we obtain

P(z)∏izi≥E1(z)E1(1/z).\frac{P(z)}{\prod_i z_i}\ge E_1(z)E_1(1/z). \tag{4}

To bound the last product with two coordinates in ratio e^R, scale those two to e^(R/2) and e^(-R/2). Put C=2 cosh(R/2) and let S,T be the sums of the other coordinates and their reciprocals. Then S+T>=2(n-2) and ST>=(n-2)^2. Consequently

E1(z)E1(1/z)=(C+S)(C+T)n2≥[1+2n(cosh(R/2)−1)]2.E_1(z)E_1(1/z) =\frac{(C+S)(C+T)}{n^2} \ge\left[1+\frac2n(\cosh(R/2)-1)\right]^2.

This remains true when one of the other coordinates is the optimized t_*. Equations (4) and the definition of F prove

F≥2log[1+2n(cosh(R/2)−1)].F\ge2\log\left[1+\frac2n(\cosh(R/2)-1)\right]. \tag{5}

Constant audit for the lower bound

If d_0=0, the required pointwise bound follows from F,d>=0. Otherwise suppose, for contradiction,

F+d2≤5d04n.F+\frac d2\le\frac{5d_0}{4n}.

Write s=sqrt(d/d_0) and w=5/4-ns^2/2. Then

0≤s≤5/(2n)<1,0≤w≤5/4,0≤F≤d0wn≤140\le s\le\sqrt{5/(2n)}<1,\qquad 0\le w\le5/4,\qquad 0\le F\le\frac{d_0w}{n}\le\frac14

for n>=5. The estimate exp(F/2)-1<=8F/15 holds on this whole interval. One rational verification uses exp(x)-1<=x+x^2/[2(1-x/3)] for 0<=x<=1/8; the latter bound is at most (49/46)x<(16/15)x.

Equation (5) therefore gives

cosh(R/2)−1≤415d0w,\cosh(R/2)-1\le\frac4{15}d_0w,

and, using d_0w<=5/4,

sinh2(R/2)≤d0w(815+16d0w225)≤2845d0w.\sinh^2(R/2) \le d_0w\left(\frac8{15}+\frac{16d_0w}{225}\right) \le\frac{28}{45}d_0w. \tag{6}

Also (2) and d_0<=1-1/n give np_max<=1+(n-1)s. Substituting in (2), dividing by sqrt(d_0), and using (6) yields

1≤s+2845(1+(n−1)s)(54n−s22).1\le s+\sqrt{\frac{28}{45}(1+(n-1)s) \left(\frac5{4n}-\frac{s^2}{2}\right)}. \tag{7}

For fixed 0<=s<1, the expression before the factor 28/45 under the square root decreases with n. At n=5, its separation from (1-s)^2 is

56s3+59s2−118s+3845>0.\frac{56s^3+59s^2-118s+38}{45}>0.

The positivity has the exact certificate

56s3+59s2−118s+38=59s2−178627s+13702729+56729(9s−5)2(9s+10).\begin{aligned} 56s^3+59s^2-118s+38 ={}&59s^2-\frac{1786}{27}s+\frac{13702}{729}\\ &+\frac{56}{729}(9s-5)^2(9s+10). \end{aligned}

For s>=0, the last term is nonnegative. The quadratic has positive leading coefficient and discriminant -43876/729<0. This makes the right side of (7) strictly less than one, a contradiction.

Thus M_j>=5d_0/(4n). Because the sum over columns of d_0 is Delta, some column has d_0>=Delta/n. Equation (3) proves (1).

Upper bound and its constants

The centered estimate applies to X=A-J because X has zero margins and ||X||_(2→2)<=1. Put Q=I-J and rho=1/(n-1). The exact centered expansion has coefficients

cn,k,r=(k−r)!(n−rk−r)2n−(k−r).c_{n,k,r}=(k-r)!\binom{n-r}{k-r}^2n^{-(k-r)}.

Taking the odd terms in absolute value gives

σk(τ(A))Sn,k≤1+bn,kΔ,bn,k=1(n−1)αn,kSn,k∑r=2kcn,k,rρrσr(Q).\frac{\sigma_k(\tau(A))}{S_{n,k}}\le1+b_{n,k}\Delta, \qquad b_{n,k}=\frac1{(n-1)\alpha_{n,k}S_{n,k}} \sum_{r=2}^k c_{n,k,r}\rho^r\sigma_r(Q). \tag{8}

All summands are nonnegative. The identity

cn,k,rSn,k=nr(k)r(n)r2\frac{c_{n,k,r}}{S_{n,k}}= \frac{n^r(k)_r}{(n)_r^2}

shows that, after division by alpha_(n,k), the dependence on k is through (k-2)_(r-2). Hence b_(n,k) is nondecreasing in k. There is no assumption on the signs of the centered coefficients of X.

Independent exact evaluation gives

constant usedvalue5/(4n2)−valueb5,43685/737287/368640b6,54767/15625047401/11250000b7,74421129/20155392035307079/9876142080\begin{array}{c|c|c} \text{constant used}&\text{value}&5/(4n^2)-\text{value}\\\hline b_{5,4}&3685/73728&7/368640\\ b_{6,5}&4767/156250&47401/11250000\\ b_{7,7}&4421129/201553920&35307079/9876142080 \end{array}

For n>=8, the published full-degree estimate implies

bn,k≤bn,n<n(n−1)(2(n−2)2−n)<54n2.b_{n,k}\le b_{n,n} <\frac{n}{(n-1)(2(n-2)^2-n)}<\frac5{4n^2}.

After multiplying positive denominators, the last strict inequality is equivalent to

6n3−55n2+85n−40>0.6n^3-55n^2+85n-40>0.

At n=8+u this is 6u^3+89u^2+357u+192, which is positive for every u>=0. Thus (8) has a strictly smaller coefficient than (1) for every proper degree with n>=5.

Small cases, boundary, and equality

For k=2, the exact identity sigma_2(X)=||X||_F^2/2 gives

σ2(A)−σ2(τ(A))=(1−1(n−1)2)σ2(X).\sigma_2(A)-\sigma_2(\tau(A)) =\left(1-\frac1{(n-1)^2}\right)\sigma_2(X).

This is strict for n>=3 unless A=J. The only remaining proper degree below order five is (n,k)=(4,3). The centered expansion and sigma_3(X)=(2/3)sum x_ij^3 give

σ3(A)−σ3(τ(A))=89σ2(X)+2827σ3(X)≥4481σ2(X).\sigma_3(A)-\sigma_3(\tau(A)) =\frac89\sigma_2(X)+\frac{28}{27}\sigma_3(X) \ge\frac{44}{81}\sigma_2(X).

The last inequality uses only x_ij>=-1/4, which implies x_ij^3>=-x_ij^2/4. It is again strict away from J.

For arbitrary nonnegative A, apply (1) first to A_epsilon=(1-epsilon)A+epsilon J. The lower estimate passes to the limit using only continuity of sigma_k and the Frobenius norm. The centered upper estimate already holds on the boundary. Strictness then follows from the positive difference between the constants in (1) and (8), not from taking a limit of strict inequalities.

Accordingly every proper-degree comparison 2<=k<n is strict unless A=J. The cases k=n retain the established Marcus–Minc equality classification: all doubly stochastic matrices in order two; J_3 and the six complements of permutation matrices in order three; only J_n when n>=4. Degree one is always an equality.

Review conclusion

No mathematical gap was found in the argument described above. Its new lower bound and strict upper-constant separation cover all proper degrees and all support boundaries. The original Malek source, the exact historical scope of that conjecture, and whether a prior full proof exists must be audited separately before claiming a newly resolved named conjecture.

Final draft read

The complete draft at /workspace/scratch/0c03194a47dc/proof.md was read after this derivation was checked. Its equations (1)–(26), normalization, initial cases, support-boundary argument, and equality list agree with the reviewed proof. In particular, its identity-entry expansion (22) has the correct factor (k)_j/(n)_j, and its simpler geometric-series proof of exp(F/2)-1<=8F/15 is valid on 0<=F<=1/4.

The draft's additional bound u<=t_*<=v is also valid: substituting t=u and t=v into the increasing minimizing equation brackets one. For clarity, the following identity makes its subsequent likelihood-ratio claim explicit. If c_i=lambda_i(x)/lambda_i(1), then

pi0pi=11+λi(1)+cit*λi(1)1+λi(1)∈[ut*,vt*].\frac{p_i^0}{p_i} =\frac1{1+\lambda_i(1)} +\frac{c_i}{t_*}\frac{\lambda_i(1)}{1+\lambda_i(1)} \in\left[\frac{u}{t_*},\frac{v}{t_*}\right].

Thus the ratio of the largest to smallest likelihood ratio is at most v/u. This supplies the intermediate step in the prose following equation (10); no stronger root comparison is needed.