Geometric spectrum conjecture for the finite free coupling matrix

From papers

Let n3n\geq 3, let hRnh\in\mathbb{R}^n be the normalized Hermite root vector, and define

En:=Ωn(,h)αα=hRn×n.E_n:=\left.\frac{\partial\,\Omega_{\boxplus_n}(\cdot,h)}{\partial\alpha}\right|_{\alpha=h}\in\mathbb{R}^{n\times n}.

Let W=1={uRn:u1=0}W=\mathbf{1}^\perp=\{u\in\mathbb{R}^n:u^\top\mathbf{1}=0\} be the mean-zero subspace. Geometric spectrum conjecture. The singular values of EnE_n restricted to WW are

spec(EnW)={2k/2:k=1,,n1},\operatorname{spec}(E_n|_W)=\{2^{-k/2}:k=1,\ldots,n-1\},

independently of nn, and the singular value on span(1)\operatorname{span}(\mathbf{1}) is 11. This conjecture would identify the exact linearized spectral structure at the Hermite diagonal and, conditionally, yield uniform local stability and finite free central-limit convergence rates; the stated evidence is computational, so the claim remains open.

Progress summary

Open

The conjecture remains open: a 2026 preprint reports strong numerical evidence but no proof or counterexample.

The conjecture asserts that the linearized finite free convolution map at the Hermite root configuration has singular values {2k/2:k=1,,n1}\{2^{-k/2}: k=1,\ldots,n-1\} on the mean-zero subspace WW, and singular value 11 in the translation direction. A proof would imply the stated local stability and central-limit consequences.

April 2026 numerical audit

The preprint FlowBoost Reveals Phase Transitions and Spectral Structure in Finite Free Information Inequalities states the conjecture and reports high-precision agreement for n{10,20,30,40,50,60,70,80,90}n\in\{10,20,30,40,50,60,70,80,90\}. It also reports numerical evidence that EnE_n is symmetric, but gives no proof, counterexample, refutation, or withdrawal; the consequences remain explicitly conditional.

Current status (as of August 2026): The conjecture remains unproved and undisproved; only numerical evidence for the proposed spectrum and symmetry is recorded.

Sources
Sources & referencesView supporting material

Primary source

Baran Hashemi, “Spectral Structure in Finite Free Information Inequalities and p-Stam Phase Transitions”, arXiv:2604.11922 (2026).

Solutions 1

Proof

The conjecture holds for every degree, and the coupling matrix is in fact real symmetric with an explicit orthogonal diagonalization.

Let HmH_m be the monic probabilists' Hermite polynomial,

Hm(x)=ex2/2xm,Hm(x)=mHm1(x),H_m(x)=e^{-\partial_x^2/2}x^m, \qquad H_m'(x)=mH_{m-1}(x),

and let h1<<hnh_1<\cdots<h_n be the roots of HnH_n. Common rescaling of both root vectors leaves the Jacobian unchanged, so this is equivalent to the normalized Hermite convention in the question.

For degree-nn polynomials, the defining finite free convolution coefficient formula gives

fnHn=j0(1)j2jj!f(2j)=ex2/2f.f\boxplus_n H_n = \sum_{j\ge0}\frac{(-1)^j}{2^j j!}f^{(2j)} = e^{-\partial_x^2/2}f.

Therefore

HnnHn=ex2xn=2n/2Hn(x/2),H_n\boxplus_n H_n = e^{-\partial_x^2}x^n = 2^{n/2}H_n(x/\sqrt2),

whose roots are 2hi\sqrt2\,h_i.

For 0m<n0\le m<n, perturb the first polynomial by

fε(x)=Hn(x)+εHm(x).f_\varepsilon(x)=H_n(x)+\varepsilon H_m(x).

Its input root-velocity vector is

vi(m)=Hm(hi)Hn(hi).v_i^{(m)} = -\frac{H_m(h_i)}{H_n'(h_i)}.

The corresponding output polynomial is

fεnHn=2n/2Hn(x/2)+ε2m/2Hm(x/2).f_\varepsilon\boxplus_n H_n = 2^{n/2}H_n(x/\sqrt2) + \varepsilon\,2^{m/2}H_m(x/\sqrt2).

Differentiating its roots at ε=0\varepsilon=0 gives

Env(m)=2(n1m)/2v(m).E_n v^{(m)} = 2^{-(n-1-m)/2}v^{(m)}.

To identify singular values rather than merely eigenvalues, apply nn-point Gaussian quadrature. At the roots hih_i, its weights are

wi=n!Hn(hi)2,w_i=\frac{n!}{H_n'(h_i)^2},

and the rule is exact through degree 2n12n-1. Since deg(HmH)2n2\deg(H_mH_\ell)\le2n-2,

i=1nHm(hi)H(hi)Hn(hi)2=1n!RHm(x)H(x)ex2/22πdx=m!n!1m=.\sum_{i=1}^{n} \frac{H_m(h_i)H_\ell(h_i)}{H_n'(h_i)^2} = \frac1{n!} \int_{\mathbb R}H_m(x)H_\ell(x) \frac{e^{-x^2/2}}{\sqrt{2\pi}}\,dx = \frac{m!}{n!}\mathbf1_{m=\ell}.

Thus the v(m)v^{(m)} form an orthogonal basis. Set

Qi,m=n!m!Hm(hi)Hn(hi),0m<n.Q_{i,m} = \sqrt{\frac{n!}{m!}}\, \frac{H_m(h_i)}{H_n'(h_i)}, \qquad 0\le m<n.

Then QTQ=IQ^{\mathsf T}Q=I, and

En=Qdiag(2(n1)/2,2(n2)/2,,21/2,1)QT.\boxed{ E_n = Q\,\operatorname{diag} \left( 2^{-(n-1)/2}, 2^{-(n-2)/2}, \ldots, 2^{-1/2}, 1 \right)Q^{\mathsf T}. }

In particular, EnE_n is symmetric positive definite, proving also the separate symmetry assertion of Remark 4.2. Because Hn=nHn1H_n'=nH_{n-1}, the m=n1m=n-1 eigenvector is constant, with eigenvalue 11. All other eigenvectors belong to W=1W=\mathbf1^\perp. Consequently

specsing(EnW)={2k/2:k=1,,n1}.\boxed{ \operatorname{spec}_{\mathrm{sing}}(E_n|_W) = \{2^{-k/2}:k=1,\ldots,n-1\}. }

This proves Conjecture 4.1 and removes the corresponding conditional assumptions from its stated spectral and local-stability consequences.

Source: B. Hashemi, “Spectral Structure in Finite Free Information Inequalities and pp-Stam Phase Transitions,” arXiv:2604.11922v2, equations (3)–(4), Conjecture 4.1, and Remark 4.2.

0 endorsements
Shivam Patel ·