Vishnyakova's conjecture

What Lean checks

Lean proves the original sampled-sequence and weighted transform-class theorems for every positive degree and weight. The proof uses all finite ordered Toeplitz minors and proves both centered disks and attained exact radii. All 17 modules passed strict sources, the complete axiom audit, and stock-kernel replay.

21 November 2026. Ordinary proof, independently reviewed internally. The source and bounded priority audit are in SOURCE_AUDIT.md. No external peer review or exhaustive priority guarantee is claimed. No numerical root computation is used in the proof.

Statement and source

Anna Vishnyakova's Polynomials interpolated totally positive sequences, arXiv:2607.06207v1, Open Problem 6.2, printed page 24, conjectures the following. Among real polynomials of degree nn whose values at nonnegative integers form a totally positive sequence, the largest possible modulus of a complex zero is nn.

Her later paper Polynomially Deformed Normalized Pochhammer Sequences Having Generating Functions With Only Real Non-positive Zeros, arXiv:2608.03723v1, Open Problem 7.2, printed page 27, states the full weighted form below. Both assertions are treated as one conjecture family.

Fix β>0\beta>0. Write (β)k=β(β+1)⋯(β+k−1)(\beta)_k=\beta(\beta+1)\cdots(\beta+k-1) and zk_=z(z−1)⋯(z−k+1)z^{\underline{k}}=z(z-1)\cdots(z-k+1), with empty products equal to one. Define the linear map

Lβ(∑k=0nqkzk_(β)k)(w)=∑k=0nqkwk.L_\beta\!\left(\sum_{k=0}^n q_k \frac{z^{\underline{k}}}{(\beta)_k}\right)(w) =\sum_{k=0}^nq_kw^k.

Let Un,βU_{n,\beta} consist of the degree-nn real polynomials PP such that all zeros of Lβ(P)L_\beta(P) lie in [−1,0][-1,0]. Multiplying PP by a nonzero constant does not affect any zero under discussion. The weighted conjecture says

supP∈Un,βmax{|z|:P(z)=0}=n−1+β.\sup_{P\in U_{n,\beta}} \max\{|z|:P(z)=0\}=n-1+\beta.

For β=1\beta=1, the equivalence with totally positive sampled sequences is proved in Section 1 of the first paper. The weighted class is defined by the transformed polynomial, without claiming that every noninteger-weight generating function is itself a totally positive generating function.

We prove the stronger disk theorem.

An earlier 7 October 2026 public report already proved this disk bound for β=1\beta=1 through degree five, and for equal parameters in every degree. Its all-degree equal-parameter argument includes the generating function, three-term recurrence, and tridiagonal representation used below in weighted form. The report explicitly leaves independent parameters in every degree open. The new step here is the exclusion for complex upper-half-plane parameters and the consequent polarization argument, which handles all independent real parameters at all degrees.

Theorem. If β>0\beta>0, n≥1n\ge 1, and P∈Un,βP\in U_{n,\beta}, then every zero of PP lies in

|z+β2|≤n−1+β2.\left|z+\frac\beta2\right|\le n-1+\frac\beta2. \tag{1}

Consequently the weighted conjecture holds, including its sharp constant. In particular, Vishnyakova's original radius is exactly nn for every degree nn.

1. A symmetric multiaffine family

Normalize Lβ(P)L_\beta(P) to be monic and factor it as

Q(w)=∏i=1n(w+ai),0≤ai≤1.Q(w)=\prod_{i=1}^n(w+a_i),\qquad 0\le a_i\le1.

Then PP is a positive or negative constant times

Fn,β(z;𝐚)=∑k=0nen−k(a1,…,an)zk_(β)k.F_{n,\beta}(z;\mathbf a) =\sum_{k=0}^n e_{n-k}(a_1,\ldots,a_n) \frac{z^{\underline{k}}}{(\beta)_k}. \tag{2}

For each fixed zz, this is symmetric and multiaffine in the aia_i. Its leading coefficient as a polynomial in zz is always 1/(β)n1/(\beta)_n, independently of every parameter. We first allow the parameters to be complex, and later real without any interval restriction.

On the diagonal define the monic polynomial

pn(z;t)=(β)nFn,β(z;t,…,t)=(β)n∑k=0n(nk)tn−kzk_(β)k.p_n(z;t)=(\beta)_nF_{n,\beta}(z;t,\ldots,t) =(\beta)_n\sum_{k=0}^n\binom nk t^{n-k} \frac{z^{\underline{k}}}{(\beta)_k}. \tag{3}

These polynomials obey

p0=1,p1=z+βt,p_0=1,\qquad p_1=z+\beta t,

pn+1(z;t)=[z−n+(2n+β)t]pn(z;t)+t(1−t)n(n+β−1)pn−1(z;t).p_{n+1}(z;t) =[z-n+(2n+\beta)t]p_n(z;t) +t(1-t)n(n+\beta-1)p_{n-1}(z;t). \tag{4}

For completeness, their exponential generating series is the formal identity

∑n≥0pn(z;t)n!un=(1+(1−t)u)z(1−tu)z+β.\sum_{n\ge0}\frac{p_n(z;t)}{n!}u^n =\frac{(1+(1-t)u)^z}{(1-tu)^{z+\beta}}. \tag{5}

It follows by inserting (3), putting n=k+jn=k+j, summing (β+k)j(tu)j/j!(\beta+k)_j(tu)^j/j!, and using the generalized binomial expansion in kk. If the right side is GG, then

[1+(1−2t)u−t(1−t)u2]G′=[z+βt+βt(1−t)u]G.[1+(1-2t)u-t(1-t)u^2]G' =[z+\beta t+\beta t(1-t)u]G.

Comparison of coefficients gives (4). Thus no orthogonality assumption or restriction that tt be real is required.

2. A reciprocal-segment lemma

Lemma. Suppose Imu<0\mathop{\mathrm{Im}} u<0, choose either square root of 1−u21-u^2, and let −1≤κ≤1-1\le \kappa\le 1. If

w=u+iκ1−u2w=u+i\kappa\sqrt{1-u^2}

has positive imaginary part, then |w|<1|w|<1.

Proof. The endpoints

r=u+i1−u2,s=u−i1−u2r=u+i\sqrt{1-u^2},\qquad s=u-i\sqrt{1-u^2}

satisfy rs=1rs=1. Neither endpoint is real: otherwise both would be real and so would uu. Their imaginary parts have opposite signs. Relabel them so that Imr>0\mathop{\mathrm{Im}} r>0, hence s=1/rs=1/r. Since

2Imu=Imr(1−1|r|2)<0,2\mathop{\mathrm{Im}}u =\mathop{\mathrm{Im}}r\left(1-\frac1{|r|^2}\right)<0,

we have |r|<1|r|<1. The line segment joining rr to 1/r1/r meets the real axis at

x0=2Rer1+|r|2.x_0=\frac{2\mathop{\mathrm{Re}}r}{1+|r|^2}.

Here |x0|<1|x_0|<1, because rr is nonreal. The part of this segment above the real axis joins the two points rr and x0x_0 in the open unit disk, and therefore lies in that disk by convexity. Every allowed ww lies on the segment. This proves the claim. The choice of square root only exchanges the endpoints. QED.

3. A diagonal zero-exclusion theorem

Let H={t:Imt>0}H=\{t:\mathop{\mathrm{Im}}t>0\} and put

Ωn,β={z:Imz>0,|z+β2|>n−1+β2}.\Omega_{n,\beta} =\left\{z:\mathop{\mathrm{Im}}z>0, \left|z+\frac\beta2\right|>n-1+\frac\beta2\right\}. \tag{6}

We show

pn(z;t)≠0(z∈Ωn,β,t∈H).p_n(z;t)\ne0\qquad(z\in\Omega_{n,\beta},\ t\in H). \tag{7}

Let D=diag(0,1,…,n−1)D=\mathop{\mathrm{diag}}(0,1,\ldots,n-1) and let the real symmetric matrix KK have zero diagonal and off-diagonal entries

Kj−1,j=Kj,j−1=j(j+β−1),1≤j≤n−1.K_{j-1,j}=K_{j,j-1}=\sqrt{j(j+\beta-1)}, \qquad 1\le j\le n-1.

Choose either square root ss of t(1−t)t(1-t) and define

Jn(t)=D−t(2D+βI)+isK.J_n(t)=D-t(2D+\beta I)+isK. \tag{8}

The determinant recurrence for this tridiagonal matrix and (4) give

det(zI−Jn(t))=pn(z;t).\det(zI-J_n(t))=p_n(z;t). \tag{9}

Suppose, toward a contradiction, that z∈Ωn,βz\in \Omega_{n,\beta} is an eigenvalue for some t∈Ht\in H. Choose an eigenvector vv with v*v=1v^* v=1, and write

d=v*Dv∈[0,n−1],k=v*Kv∈ℝ.d=v^*Dv\in[0,n-1],\qquad k=v^*Kv\in\mathbb R.

The second quantity is real because KK is Hermitian, even though both the eigenvector and Jn(t)J_n(t) may be complex. Cauchy's elementary inequality gives

|k|≤2∑j=1n−1j(j+β−1)|vj||vj−1|≤∑j=1n−1(j|vj|2+(j+β−1)|vj−1|2)=2d+β−(n+β−1)|vn−1|2≤2d+β.\begin{aligned} |k| &\le2\sum_{j=1}^{n-1}\sqrt{j(j+\beta-1)} |v_j|\,|v_{j-1}|\\ &\le\sum_{j=1}^{n-1} \bigl(j|v_j|^2+(j+\beta-1)|v_{j-1}|^2\bigr)\\ &=2d+\beta-(n+\beta-1)|v_{n-1}|^2 \le2d+\beta. \end{aligned} \tag{10}

For n=1n=1 the empty sum is zero and the same identity holds. Put

c=d+β2>0,κ=k2d+β∈[−1,1],u=1−2t.c=d+\frac\beta2>0,\quad \kappa=\frac{k}{2d+\beta}\in[-1,1],\quad u=1-2t.

Taking v*Jn(t)v=zv^* J_n(t) v=z, and choosing 1−u2=2s\sqrt{1-u^2}=2s, gives

z+β2=c(u+iκ1−u2).z+\frac\beta2 =c\left(u+i\kappa\sqrt{1-u^2}\right). \tag{11}

Here Imu<0\mathop{\mathrm{Im}} u<0. Since Imz>0\mathop{\mathrm{Im}} z>0, the reciprocal-segment lemma applies to the bracket in (11), so

|z+β2|<c≤n−1+β2.\left|z+\frac\beta2\right|<c \le n-1+\frac\beta2.

This contradicts (6) and establishes (7).

4. Polarization and real specialization

Fix z∈Ωn,βz\in \Omega_{n,\beta}. By (7), the diagonal polynomial t↦Fn,β(z;t,…,t)t\longmapsto F_{n,\beta}(z;t,\ldots,t) has no zero in the open upper half-plane. The Grace–Walsh–Szegő coincidence theorem for the convex circular domain HH therefore implies

Fn,β(z;a1,…,an)≠0whenever every ai∈H.F_{n,\beta}(z;a_1,\ldots,a_n)\ne0 \quad\text{whenever every }a_i\in H. \tag{12}

The hypotheses are exactly symmetry and multiaffinity in the parameters, established in Section 1. The upper half-plane is convex, so no additional degree condition is needed. A primary reference is Borcea–Brändén, The Lee–Yang and Pólya–Schur Programs. II, Theorem 1.1 on printed page 4; their Theorem 2.1 gives a more general form.

Now fix arbitrary real parameters a1,…,ana_1,\ldots,a_n. For every ϵ>0\epsilon>0, the polynomial

z↦Fn,β(z;a1+iϵ,…,an+iϵ)z\longmapsto F_{n,\beta}(z;a_1+i\epsilon,\ldots,a_n+i\epsilon)

has no zero on Ωn,β\Omega_{n,\beta}, by (12). These polynomials converge locally uniformly there to Fn,β(z;𝐚)F_{n,\beta}(z;\mathbf a). The domain in (6) is connected. Hurwitz's theorem shows that the limit is either zero-free there or identically zero. Its fixed, nonzero leading coefficient 1/(β)n1/(\beta)_n excludes the latter possibility. Thus it has no zero on Ωn,β\Omega_{n,\beta}. Its coefficients are real, so conjugation gives the same conclusion in the lower half-plane.

We have proved a statement stronger than the nonreal part needed below: for every real parameter vector aa, every nonreal zero of (2) belongs to the disk in (1).

5. Real zeros and sharpness

Restore 0≤ai≤10\le a_i\le 1. For real z>n−1z>n-1, every falling factorial in (2) is positive, the coefficients are nonnegative, and the coefficient of the term of degree nn is one. Hence Fn,β(z;𝐚)>0F_{n,\beta}(z;\mathbf a)>0.

We use the reflection identity

Fn,β(−z−β;𝐚)=(−1)nFn,β(z;𝟏−𝐚).F_{n,\beta}(-z-\beta;\mathbf a) =(-1)^nF_{n,\beta}(z;\mathbf1-\mathbf a). \tag{13}

Here is a direct derivation. For any polynomial Q(w)=∑jqjwjQ(w)=\sum_jq_jw^j, the finite Chu–Vandermonde identity is

∑k=0j(jk)zk_(β)k=(z+β)j(β)j.\sum_{k=0}^j\binom jk \frac{z^{\underline{k}}}{(\beta)_k} =\frac{(z+\beta)_j}{(\beta)_j}.

Expanding Q(−w−1)Q(-w-1) contributes an overall factor (−1)j(-1)^j for its term of degree jj. Applying Lβ−1L_\beta^{-1} therefore gives

Lβ−1[Q(−w−1)](z)=(Lβ−1Q)(−z−β).L_\beta^{-1}[Q(-w-1)](z) =(L_\beta^{-1}Q)(-z-\beta).

For Q(w)=∏i(w+ai)Q(w)=\prod_i(w+a_i), its reflected polynomial is (−1)n∏i(w+1−ai)(-1)^n\prod_i(w+1-a_i), yielding (13). It is also formula (12) in Vishnyakova's later paper.

Since every 1−ai1-a_i is still in [0,1][0,1], (13) excludes real zeros below −β−n+1-\beta-n+1. Thus all real zeros lie in [−β−n+1,n−1][-\beta-n+1,n-1], which is the real diameter of (1). Section 4 already bounds every nonreal zero, proving the disk theorem.

Finally choose all ai=1a_i=1. The same finite Vandermonde identity gives

Fn,β(z;1,…,1)=(z+β)n(β)n.F_{n,\beta}(z;1,\ldots,1) =\frac{(z+\beta)_n}{(\beta)_n}.

It has the zero z=−(β+n−1)z=-(\beta+n-1) and belongs to Un,βU_{n,\beta} because its transform is (w+1)n(w+1)^n. On the other hand, (1) implies |z|≤n−1+β|z|\le n-1+\beta for every root of every polynomial in that class. The supremum is therefore attained and equals n−1+βn-1+\beta. QED.

Scope

This proves the entire radius conjecture, at all degrees and all positive weights. The original sampled-total-positivity statement is its β=1\beta=1 instance. They are one resolution, not separate counts. The source's different conjectures about a uniformly nearby zero, gaps between real zeros, and intervals of shifts for even powers are not claimed here.