Malek's subpermanent conjecture

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.

Research date: 20 November 2026 (America/Chicago).

Status: complete proposed proof. Two independent internal reviews found no mathematical gap; the bounded source audit found no prior full endpoint theorem. The proof is unrefereed. This is the named all-degree extension attributed to Malek in Cheon and Wanless's discussion of Minc Conjecture 3. It is a subsidiary in the Minc source chain, not a second resolution of the already proved permanent case. No Lean verification is part of this work.

Statement

Let Ωn\Omega_n be the nonnegative doubly stochastic matrices of order nn, let Jn=𝟏𝟏T/nJ_n=\mathbf1\mathbf1^T/n, and let σk(A)\sigma_k(A) be the sum of the permanents of all kk-by-kk submatrices, with σ0=1\sigma_0=1. Define

𝒯n(A)=nJn−An−1=Jn−A−Jnn−1.\mathcal T_n(A)=\frac{nJ_n-A}{n-1} =J_n-\frac{A-J_n}{n-1}.

Theorem. For every integer n≥2n\ge2, every A∈ΩnA\in\Omega_n, and every 0≤k≤n0\le k\le n,

σk(A)≥σk(𝒯n(A)).\sigma_k(A)\ge \sigma_k(\mathcal T_n(A)). \tag{1}

For k=0,1k=0,1, equality always holds. For k≥2k\ge2, the equality cases are:

  • n=2,k=2n=2,\ k=2: every doubly stochastic matrix;
  • n=3,k=2n=3,\ k=2: only J3J_3;
  • n=3,k=3n=3,\ k=3: J3J_3 and (𝟏𝟏T−P)/2(\mathbf1\mathbf1^T-P)/2, where PP is a permutation matrix;
  • n≥4n\ge4: only JnJ_n.

The case k=nk=n, including its equality cases, is Lavi's previously proved Marcus–Minc theorem [2]. The new argument covers every proper degree simultaneously. It does not assert monotonicity along the entire segment from AA to its transform; that stronger statement is false.

Inputs and normalization

Put

dn=n!nn,Sn,k=(nk)2k!nk,αn,k=k(k−1)n(n−1),X=A−Jn,r2=∥X∥F2,d_n=\frac{n!}{n^n},\qquad S_{n,k}=\binom nk^2\frac{k!}{n^k},\qquad \alpha_{n,k}=\frac{k(k-1)}{n(n-1)},\qquad X=A-J_n,\qquad r^2=\|X\|_F^2,

and write qk(A)=σk(A)/Sn,kq_k(A)=\sigma_k(A)/S_{n,k}. The proof establishes, for n≥5n\ge5 and 2≤k<n2\le k<n,

qk(A)≥exp(54n2αn,kr2).q_k(A)\ge \exp\!\left(\frac{5}{4n^2}\alpha_{n,k}r^2\right). \tag{2}

It then compares this with a strictly smaller linear upper coefficient for qk(𝒯n(A))q_k(\mathcal T_n(A)).

We use two published analytic inputs, with their roles kept separate:

  1. Gurvits's coefficient–capacity inequality: if pp is homogeneous of degree mm in mm variables, real stable, and has nonnegative coefficients, then [z1⋯zm]p≥dmCap(p)[z_1\cdots z_m]p\ge d_m\mathop{\mathrm{Cap}}(p), where Cap(p)=infz>0p(z)/∏izi\mathop{\mathrm{Cap}}(p)=\inf_{z>0}p(z)/\prod_i z_i [4].
  2. Lavi's centered-subpermanent estimate [2, Lemma 3.1; 3, Lemma 5.1]: if YY has zero row and column sums and ∥Y∥2→2≤1\|Y\|_{2\to2}\le1, and Q=I−JnQ=I-J_n, then |σj(Y)|≤σj(Q)∥Y∥F2/(n−1)|\sigma_j(Y)|\le \sigma_j(Q)\|Y\|_F^2/(n-1) for 2≤j≤n2\le j\le n.

The stable polynomial, optimized-slice notation, root movement, and scalar coefficient step below are those of [3, Sections 2–3]. They are reproduced to make the improvement in their capacity estimate explicit.

1. A stronger capacity cost for proper degrees

Initially suppose every entry of AA is positive. Let 2≤k<n2\le k<n, m=n−km=n-k, and define

P(z)=ek(Az)(nk)(z1+⋯+znn)m.P(z)=\frac{e_k(Az)}{\binom nk} \left(\frac{z_1+\cdots+z_n}{n}\right)^m. \tag{3}

The elementary symmetric polynomial eke_k is real stable. Composition with the positive linear forms in AzAz, and multiplication by the uniform linear factors, preserve stability. Thus PP is real stable, homogeneous of degree nn, and has positive coefficients. Direct differentiation gives

P(𝟏)=1,∂jP(𝟏)=1,∇2logP(𝟏)=−Jn−αn,kXTX.P(\mathbf1)=1,\qquad \partial_jP(\mathbf1)=1,\qquad \nabla^2\log P(\mathbf1)=-J_n-\alpha_{n,k}X^TX. \tag{4}

Its central coefficient is

[z1⋯zn]P=(n−k)!(nk)nn−kσk(A)=dnqk(A).[z_1\cdots z_n]P =\frac{(n-k)!}{\binom nk\,n^{n-k}}\sigma_k(A) =d_nq_k(A). \tag{5}

For a positive vector zz, write Ej(z)=ej(z)/(nj)E_j(z)=e_j(z)/\binom nj. Since AA is doubly stochastic, AzAz is majorized by zz. Elementary symmetric polynomials are Schur-concave on the positive orthant, hence Ek(Az)≥Ek(z)E_k(Az)\ge E_k(z). One can see this Schur-concavity directly: averaging a pair of coordinates with fixed sum increases their product, and the coefficient of that product in eke_k is nonnegative.

Newton's inequalities imply that Ej+1(z)/Ej(z)E_{j+1}(z)/E_j(z) is nonincreasing in jj. Therefore

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

Combining these facts yields the capacity-cost estimate

P(z)∏izi≥En−1(z)E1(z)∏izi=(∑izin)(∑izi−1n).\frac{P(z)}{\prod_i z_i} \ge \frac{E_{n-1}(z)E_1(z)}{\prod_i z_i} = \left(\frac{\sum_i z_i}{n}\right) \left(\frac{\sum_i z_i^{-1}}{n}\right). \tag{6}

This replaces the weaker estimate that keeps only the n−kn-k uniform factors. It uses k<nk<n essentially.

Fix a coordinate jj, hold the other coordinates x=(xh)h≠jx=(x_h)_{h\ne j} positive, and set zj=tz_j=t. The polynomial hx(t)=P(z)h_x(t)=P(z) has degree nn, positive coefficients and negative real roots. The function hx(t)/th_x(t)/t has a unique minimizer t*>0t_*>0. Write

hx(t)=hx(0)∏i=1n(1+tλi(x)),pi=t*t*+λi(x),d=∑ipi2−1n,h_x(t)=h_x(0)\prod_{i=1}^n\left(1+\frac t{\lambda_i(x)}\right), \qquad p_i=\frac{t_*}{t_*+\lambda_i(x)},\qquad d=\sum_i p_i^2-\frac1n,

F=loghx(t*)t*∏h≠jxh,R=logmaxh≠jxhminh≠jxh.F=\log\frac{h_x(t_*)}{t_*\prod_{h\ne j}x_h}, \qquad R=\log\frac{\max_{h\ne j}x_h}{\min_{h\ne j}x_h}. \tag{7}

The minimizing equation is ∑ipi=1\sum_i p_i=1. At x=𝟏x=\mathbf1, equation (4) gives t*=1t_*=1; write the corresponding probability vector as p0p^0. Then

d0:=∥p0−𝟏n∥22=αn,k∥X:j∥22,0≤d0≤1−1n.d_0:=\left\|p^0-\frac{\mathbf1}{n}\right\|_2^2 =\alpha_{n,k}\|X_{:j}\|_2^2,\qquad 0\le d_0\le1-\frac1n. \tag{8}

To apply (6), choose two of the fixed coordinates with ratio eRe^R. Their sums in zz and z−1z^{-1} have product 4cosh2(R/2)4\cosh^2(R/2). If S,TS,T are the corresponding sums of the other n−2n-2 coordinates and their reciprocals, then ST≥(n−2)2ST\ge(n-2)^2. Cauchy–Schwarz on the cross terms gives

(∑izi)(∑izi−1)≥(n−2+2cosh(R/2))2.\left(\sum_i z_i\right)\left(\sum_i z_i^{-1}\right) \ge\left(n-2+2\cosh(R/2)\right)^2.

Consequently, for every positive xx,

F≥2log[1+2n(cosh(R/2)−1)].F\ge 2\log\left[1+\frac2n\left(\cosh(R/2)-1\right)\right]. \tag{9}

In particular, F≥0F\ge0.

2. The root-movement estimate

We recall the root comparison in [3, Lemma 3.1 and equations (3.10)–(3.14)]. Set

ai=aij,bi=∑h≠jaihxh,bi0=1−ai,a_i=a_{ij},\qquad b_i=\sum_{h\ne j}a_{ih}x_h,\qquad b_i^0=1-a_i,

and let u=minh≠jxhu=\min_{h\ne j}x_h, v=maxh≠jxhv=\max_{h\ne j}x_h. Then ub0≤b≤vb0ub^0\le b\le vb^0 coordinatewise.

The ordered negative roots of ek(b+at)e_k(b+at) are nonincreasing in every coordinate of bb, and scale linearly when bb is scaled. For completeness, varying one coordinate of bb gives a pencil H(t)+sL(t)H(t)+sL(t). Real stability places HH and LL in proper position; their zeros interlace and the ordered zeros of this pencil move to the left as ss increases. The monotonicity extends across multiple roots by continuity. The additional uniform factors have root −∑h≠jxh-\sum_{h\ne j}x_h, with the same bounds. After merging the ordered root lists,

uλi(𝟏)≤λi(x)≤vλi(𝟏),u≤t*≤v.u\lambda_i(\mathbf1)\le\lambda_i(x)\le v\lambda_i(\mathbf1), \qquad u\le t_*\le v. \tag{10}

The second assertion follows from the strictly increasing equation ∑it/(t+λi(x))=1\sum_i t/(t+\lambda_i(x))=1. Therefore the likelihood ratios pi0/pip_i^0/p_i lie in an interval whose largest-to-smallest ratio is at most v/u=eRv/u=e^R.

Here is the probability calculation that converts this to a Euclidean bound. If two positive probability vectors have likelihood ratios Li=pi0/piL_i=p_i^0/p_i between ℓ\ell and qℓq\ell, with q≤eRq\le e^R, then ∑ipiLi=1\sum_i p_iL_i=1 and

∑ipi(Li−1)2≤(qℓ−1)(1−ℓ)≤(q−1)24q≤sinh2(R/2).\sum_i p_i(L_i-1)^2 \le(q\ell-1)(1-\ell) \le\frac{(q-1)^2}{4q} \le\sinh^2(R/2).

Writing pmax=maxipip_{\max}=\max_i p_i, this gives

∥p0−p∥22≤pmaxsinh2(R/2),d0≤d+pmaxsinh(R/2).\|p^0-p\|_2^2\le p_{\max}\sinh^2(R/2), \qquad \sqrt{d_0}\le\sqrt d+\sqrt{p_{\max}}\sinh(R/2). \tag{11}

Also, since the deviations of the coordinates of pp from 1/n1/n sum to zero,

pmax≤1n+n−1nd.p_{\max}\le\frac1n+\sqrt{\frac{n-1}{n}\,d}. \tag{12}

3. A uniform slice bound

We prove that, for n≥5n\ge5,

F+d2≥54nd0.F+\frac d2\ge\frac{5}{4n}d_0. \tag{13}

If d0=0d_0=0, this follows from F,d≥0F,d\ge0. Otherwise suppose, for a contradiction, that the reverse weak inequality holds. Set

s=dd0,w=54−n2s2.s=\sqrt{\frac d{d_0}},\qquad w=\frac54-\frac n2s^2.

Then w≥0w\ge0, s2≤5/(2n)≤1/2s^2\le5/(2n)\le1/2, and

0≤F≤d0wn≤14,npmax≤1+(n−1)s.0\le F\le\frac{d_0w}{n}\le\frac14,\qquad np_{\max}\le1+(n-1)s. \tag{14}

For 0≤F≤1/40\le F\le1/4, the exponential series gives

eF/2−1≤F/21−F/4≤815F.e^{F/2}-1 \le\frac{F/2}{1-F/4} \le\frac8{15}F. \tag{15}

Indeed, r!≥2r−1r!\ge2^{r-1} for r≥1r\ge1 bounds the series by the displayed geometric series.

Equation (9) implies cosh(R/2)≤1+n2(eF/2−1)\cosh(R/2)\le1+\frac n2(e^{F/2}-1). Applying (14) and (15), and using d0≤1d_0\le1 and w≤5/4w\le5/4, we obtain

sinh2(R/2)≤n(eF/2−1)+n24(eF/2−1)2≤815d0w+16225d02w2≤2845d0w.\begin{aligned} \sinh^2(R/2) &\le n(e^{F/2}-1)+\frac{n^2}{4}(e^{F/2}-1)^2\\ &\le \frac8{15}d_0w+\frac{16}{225}d_0^2w^2\\ &\le\frac{28}{45}d_0w. \end{aligned} \tag{16}

Dividing (11) by d0\sqrt{d_0} now gives

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{17}

This is impossible. For fixed 0≤s<10\le s<1, the expression

(1+(n−1)s)(54n−s22)=54s+5(1−s)4n−s2(1−s)2−ns32(1+(n-1)s)\left(\frac5{4n}-\frac{s^2}{2}\right) =\frac54s+\frac{5(1-s)}{4n} -\frac{s^2(1-s)}2-\frac{ns^3}2

decreases with nn. At n=5n=5, the difference between (1−s)2(1-s)^2 and the expression under the square root in (17) is

56s3+59s2−118s+3845>0.\frac{56s^3+59s^2-118s+38}{45}>0. \tag{18}

To verify this strictly and without numerical approximation, convexity of s3s^3 on s≥0s\ge0, with its tangent at 5/95/9, gives

56s3+59s2−118s+38≥59s2−178627s+13702729.56s^3+59s^2-118s+38 \ge59s^2-\frac{1786}{27}s+\frac{13702}{729}.

The quadratic has positive leading coefficient and discriminant −43876/729-43876/729. It is strictly positive for every real ss. Since s<1s<1, equation (18) makes the right side of (17) strictly less than one. This proves (13).

4. From slices to the subpermanent lower bound

The scalar step of Gurvits's proof has a useful refinement, reproduced in [3, Lemma 2.2]. For the optimized slice in (7),

hx′(0)≥(n−1n)n−1hx(t*)t*ed/2.h_x'(0)\ge \left(\frac{n-1}{n}\right)^{n-1} \frac{h_x(t_*)}{t_*}e^{d/2}. \tag{19}

It also follows directly. In terms of the probability vector pp,

hx′(0)hx(t*)/t*=∑ipi∏ℓ≠i(1−pℓ)≥∏i(1−pi)1−pi≥(n−1n)n−1ed/2.\frac{h_x'(0)}{h_x(t_*)/t_*} =\sum_i p_i\prod_{\ell\ne i}(1-p_\ell) \ge\prod_i(1-p_i)^{1-p_i} \ge\left(\frac{n-1}{n}\right)^{n-1}e^{d/2}.

The first inequality is weighted AM–GM. For the second, expand (1−t)log(1−t)=−t+∑ℓ≥2tℓ/[ℓ(ℓ−1)](1-t)\log(1-t)=-t+\sum_{\ell\ge2}t^\ell/[\ell(\ell-1)]. The linear terms cancel; the quadratic term contributes d/2d/2; convexity makes every higher contribution relative to the uniform probability vector nonnegative.

Let Pj=∂jP|zj=0P_j=\partial_jP|_{z_j=0}. Equations (13) and (19) give

Cap(Pj)≥(n−1n)n−1exp(54nd0).\mathop{\mathrm{Cap}}(P_j) \ge \left(\frac{n-1}{n}\right)^{n-1} \exp\!\left(\frac{5}{4n}d_0\right).

The polynomial PjP_j is stable, homogeneous of degree n−1n-1 in n−1n-1 variables, and has nonnegative coefficients. Applying the coefficient–capacity inequality and using dn−1((n−1)/n)n−1=dnd_{n-1}((n-1)/n)^{n-1}=d_n, equations (5) and (8) imply

qk(A)≥exp(54nαn,k∥X:j∥22)for every j.q_k(A)\ge \exp\!\left(\frac{5}{4n}\alpha_{n,k}\|X_{:j}\|_2^2\right) \quad\text{for every }j.

Some column has squared norm at least r2/nr^2/n. This proves (2) for positive AA. Replacing an arbitrary A∈ΩnA\in\Omega_n by (1−ε)A+εJn(1-\varepsilon)A+\varepsilon J_n and taking ε↓0\varepsilon\downarrow0 proves (2) on the whole polytope, including its boundary.

5. The transformed upper bound

For a matrix YY with zero margins, selecting the entries supplied by YY in a subpermanent gives

qk(Jn+tY)=1+∑j=2k(k)jnj(n)j2σj(Y)tj.q_k(J_n+tY) =1+\sum_{j=2}^k \frac{(k)_j n^j}{(n)_j^2}\,\sigma_j(Y)t^j. \tag{20}

Here (a)j=a(a−1)⋯(a−j+1)(a)_j=a(a-1)\cdots(a-j+1). A doubly stochastic matrix is a convex combination of permutation matrices and hence has Euclidean operator norm at most one. Its restriction to 𝟏⊥\mathbf1^\perp is X=A−JnX=A-J_n, so ∥X∥2→2≤1\|X\|_{2\to2}\le1.

Apply the centered estimate from [2,3] term by term in (20), with t=−1/(n−1)t=-1/(n-1). Since σj(Q)≥0\sigma_j(Q)\ge0, the result is

qk(𝒯n(A))≤1+bn,kαn,kr2,bn,k:=qk(Jn+Q/(n−1))−1(n−1)αn,k.q_k(\mathcal T_n(A)) \le1+b_{n,k}\alpha_{n,k}r^2, \qquad b_{n,k}:= \frac{q_k(J_n+Q/(n-1))-1}{(n-1)\alpha_{n,k}}. \tag{21}

The positivity of σj(Q)\sigma_j(Q) also follows from the tensor identity σj(Q)=⟨uj,Q⊗juj⟩≥0\sigma_j(Q)=\langle u_j,Q^{\otimes j}u_j\rangle\ge0, with uju_j the sum of the normalized square-free symmetric basis tensors.

For fixed nn, bn,kb_{n,k} is nondecreasing in k≥2k\ge2. In the centered expansion, division by αn,k\alpha_{n,k} leaves the jj-th coefficient proportional to (k−2)j−2(k-2)_{j-2}, with every other factor nonnegative. Terms first appearing at a larger kk are nonnegative as well.

For exact evaluation, selecting identity entries gives

qk(Jn+tQ)=∑j=0k(k)j(n)j(nt)j(1−t)k−jj!.q_k(J_n+tQ)= \sum_{j=0}^k \frac{(k)_j}{(n)_j} \frac{(nt)^j(1-t)^{k-j}}{j!}. \tag{22}

The following rational comparisons handle the initial orders:

Range Upper coefficient used Exact gap from 5/(4n2)5/(4n^2)
n=5,2≤k<5n=5,\ 2\le k<5 b5,4=3685/73728b_{5,4}=3685/73728 7/3686407/368640
n=6,2≤k<6n=6,\ 2\le k<6 b6,5=4767/156250b_{6,5}=4767/156250 47401/1125000047401/11250000
n=7,2≤k<7n=7,\ 2\le k<7 b7,7=4421129/201553920b_{7,7}=4421129/201553920 35307079/987614208035307079/9876142080

These are checks of three scalar constants in an all-order proof, not finite matrix tests.

For n≥8n\ge8, use bn,k≤bn,n=βnb_{n,k}\le b_{n,n}=\beta_n and [2, Section 5]'s bound

βn≤n(n−1)(2(n−2)2−n)<54n2.\beta_n\le \frac{n}{(n-1)(2(n-2)^2-n)} <\frac5{4n^2}. \tag{23}

The first bound can be checked directly from (22) at k=nk=n. For 0<t<10<t<1,

qn(Jn+tQ)≤(1−t)nexp(nt1−t).q_n(J_n+tQ)\le (1-t)^n\exp\!\left(\frac{nt}{1-t}\right).

Its logarithm is at most

n(log(1−t)+t1−t)=n∑ℓ≥2ℓ−1ℓtℓ≤nt22(1−t)2.n\left(\log(1-t)+\frac{t}{1-t}\right) =n\sum_{\ell\ge2}\frac{\ell-1}{\ell}t^\ell \le\frac{nt^2}{2(1-t)^2}.

At t=1/(n−1)t=1/(n-1), put x=n/[2(n−2)2]<1x=n/[2(n-2)^2]<1. The estimate ex−1≤x/(1−x)e^x-1\le x/(1-x) gives the first inequality in (23).

The second inequality is equivalent to 6n3−55n2+85n−40>06n^3-55n^2+85n-40>0. On writing n=8+mn=8+m, this polynomial is 6m3+89m2+357m+1926m^3+89m^2+357m+192, positive for every m≥0m\ge0.

Thus bn,k<5/(4n2)b_{n,k}<5/(4n^2) for all n≥5,2≤k<nn\ge5,\ 2\le k<n. Combining (2) and (21) gives

qk(A)−qk(𝒯n(A))≥(54n2−bn,k)αn,kr2.q_k(A)-q_k(\mathcal T_n(A)) \ge \left(\frac5{4n^2}-b_{n,k}\right)\alpha_{n,k}r^2. \tag{24}

This proves the inequality and strictness away from JnJ_n for all those parameters.

6. Initial orders and equality

The degrees k=0,1k=0,1 are constant on Ωn\Omega_n. The full degree k=nk=n is [2, Theorem 1.1], with the equality cases listed above.

For proper degree two, the centered expansion is exact and

σ2(A)−σ2(𝒯n(A))=(1−1(n−1)2)σ2(X),σ2(X)=12∥X∥F2.\sigma_2(A)-\sigma_2(\mathcal T_n(A)) =\left(1-\frac1{(n-1)^2}\right)\sigma_2(X), \qquad \sigma_2(X)=\frac12\|X\|_F^2. \tag{25}

This is strictly positive for n≥3n\ge3 unless A=JnA=J_n.

The only remaining pair is n=4,k=3n=4,\ k=3. For every zero-margin matrix,

σ3(X)=23∑i,jxij3.\sigma_3(X)=\frac23\sum_{i,j}x_{ij}^3.

The identity follows by inclusion–exclusion on the repeated row and column indices in the ordered triple sum. Since xij≥−1/nx_{ij}\ge-1/n,

σ3(X)≥−43nσ2(X).\sigma_3(X)\ge-\frac4{3n}\sigma_2(X).

Using

σ3(Jn+tX)=σ3(Jn)+(n−2)2nt2σ2(X)+t3σ3(X)\sigma_3(J_n+tX) =\sigma_3(J_n)+\frac{(n-2)^2}{n}t^2\sigma_2(X) +t^3\sigma_3(X)

at n=4n=4, t=1t=1 and t=−1/3t=-1/3, gives

σ3(A)−σ3(𝒯4(A))≥4481σ2(X).\sigma_3(A)-\sigma_3(\mathcal T_4(A)) \ge\frac{44}{81}\sigma_2(X). \tag{26}

It is strict away from J4J_4. Together with the prior full-degree theorem, this completes every parameter and every equality case.

Scope, dependency, and review status

This proves one named historical subsidiary conjecture if the argument withstands review. The full-permanent endpoint, the centered estimate, the stable polynomial, and the root-movement estimate are prior work. The proposed new step is the reciprocal-mean capacity cost (6), its slice consequence (13), and their use to close the negative transform for every proper degree.

It does not solve Lih–Wang's half-segment inequality, the false Holens–Djokovic conjecture, Gyires's inequality, or the full boundary-monotonicity classification. It supplies no new claim for the already resolved numbered Conjecture 3.

The narrow order-five rational margin is explicit. The proof of its lower coefficient is an all-order analytic inequality, and the scalar checks accompany the proof for reproducibility. They do not replace the proof, and internal review is not external refereeing. See the separate source audit for priority and eligibility limits.

References

  1. G.-S. Cheon and I. M. Wanless, An update on Minc's survey of open problems involving permanents, Linear Algebra and its Applications 403 (2005), 314–342, DOI 10.1016/j.laa.2005.02.030. The Conjecture 3 discussion, equation (9), attributes the all-degree endpoint to Malek. Author PDF.
  2. Y. Lavi, The Marcus–Minc Transform Inequality, 24 September 2026, arXiv:2609.29262v1. Theorem 1.1, Lemma 3.1, and Section 5 are used. Numbered Minc page.
  3. Y. Lavi, The Kopotun Subpermanent Transform Inequality, manuscript dated 11 October 2026, Hexagon 2611.00201v1. Sections 2–3 supply the polynomial and optimized-root method. Its proved transform is (nJn+A)/(n+1)(nJ_n+A)/(n+1), distinct from (1).
  4. L. Gurvits, Van der Waerden/Schrijver–Valiant like conjectures and stable (aka hyperbolic) homogeneous polynomials: one theorem for all, Electronic Journal of Combinatorics 15 (2008), R66, DOI 10.37236/790.
  5. M. Malek, A note on a permanental conjecture of M. Marcus and H. Minc, Linear and Multilinear Algebra 25 (1989), 71–73, DOI 10.1080/03081088908817929. The exact conjecture is transcribed from [1]; full-text access limits for this original paper are stated in the source audit.