Gröchenig’s problem

What Lean checks

Complete proposed infinite counterexample to the upper Beurling-density comparison in Problem 4.12, with absolutely summable coefficients. Two internal mathematical reviews and exact block-algebra diagnostics passed. The separate lower-density inequality remains undecided here. No Lean formalization is claimed.

Statement and scope

Karlheinz Gröchenig, Infinite totally positive matrices: some open problems, Acta Scientiarum Mathematicarum (2026), DOI 10.1007/s44146-026-00264-3, Problem 4.12 asks whether

D±(Z(f))≤D±(c)D^\pm(Z(f))\le D^\pm(c)

for every Pólya frequency function gg and every

f(x)=∑k∈ℤckg(x−k)∈V∞(g),c∈ℓ∞(ℤ).f(x)=\sum_{k\in\mathbb Z}c_k g(x-k)\in V^\infty(g),\qquad c\in\ell^\infty(\mathbb Z).

Here Z(f)={x∈ℝ:f(x)=0}Z(f)=\{x\in\mathbb R:f(x)=0\}. The source defines coefficient sign changes by

M(c)={j∈ℤ:there is r∈ℕ with cjcj+r<0 and cj+l=0(1≤l<r)}.M(c)=\{j\in\mathbb Z:\text{there is }r\in\mathbb N\text{ with }c_jc_{j+r}<0 \text{ and }c_{j+l}=0\ (1\le l\lt r)\}.

Thus zeros in the coefficient sequence are skipped when deciding whether its next nonzero term has the opposite sign. Its upper density is

D+(c)=limsupn→∞supa∈ℤ#(M(c)∩[a,a+n])n.D^+(c)=\limsup_{n\to\infty}\sup_{a\in\mathbb Z} \frac{\#(M(c)\cap[a,a+n])}{n}.

For a locally finite subset E⊂ℝE\subset\mathbb R, write

D+(E)=limsupR→∞supa∈ℝ#(E∩[a,a+R])R.D^+(E)=\limsup_{R\to\infty}\sup_{a\in\mathbb R} \frac{\#(E\cap[a,a+R])}{R}.

Restricting to positive integer lengths does not change this density. Using integer starting points also suffices for the lower bounds below. If a supremum is infinite, the same formulas have their extended-real meaning.

We construct a real sequence c∈ℓ1(ℤ)⊂ℓ∞(ℤ)c\in\ell^1(\mathbb Z)\subset\ell^\infty(\mathbb Z) for the Gaussian

g(x)=e−πx2g(x)=e^{-\pi x^2}

such that the corresponding nonzero function is entire and

D+(c)=12,D+(Z(f))≥1.D^+(c)=\frac12, \qquad D^+(Z(f))\ge1.

The Gaussian is explicitly listed as a Pólya frequency function in the source following Definition 2.5. This refutes the upper-density inequality and therefore answers the universal assertion in Problem 4.12 negatively. It does not settle the separate lower-density inequality.

1. Finite Gaussian blocks

For every integer N≥1N\ge1, define

PN(t)=∏j=1N(t−e4πj)=∑k=0NpN,ktk,AN,k=pN,ke4πk2.P_N(t)=\prod_{j=1}^{N}(t-e^{4\pi j}) =\sum_{k=0}^{N}p_{N,k}t^k, \qquad A_{N,k}=p_{N,k}e^{4\pi k^2}.

All roots of PNP_N are positive, so

(−1)N−kpN,k>0(0≤k≤N).(-1)^{N-k}p_{N,k}>0\qquad(0\le k\le N).

In particular every coefficient is nonzero, and consecutive coefficients have opposite signs. Put

BN=∑k=0N|AN,k|>0,aN,k=2−NBNAN,k,hN(x)=∑k=0NaN,kg(x−2k).B_N=\sum_{k=0}^{N}|A_{N,k}|>0, \qquad a_{N,k}=\frac{2^{-N}}{B_N}A_{N,k}, \qquad h_N(x)=\sum_{k=0}^{N}a_{N,k}g(x-2k).

The elementary identity

e−π(x−2k)2=e−πx2e4πkxe−4πk2e^{-\pi(x-2k)^2}=e^{-\pi x^2}e^{4\pi kx}e^{-4\pi k^2}

gives the exact factorization

hN(x)=2−NBNe−πx2PN(e4πx).h_N(x)=\frac{2^{-N}}{B_N}e^{-\pi x^2}P_N(e^{4\pi x}). \tag{1}

Consequently hNh_N has simple real zeros at 1,2,…,N1,2,\ldots,N and changes sign between the successive test points

tN,j=j+12,0≤j≤N.t_{N,j}=j+\frac12,\qquad 0\le j\le N.

More precisely,

signhN(tN,j)=(−1)N−j.\mathop{\mathrm{sign}}h_N(t_{N,j})=(-1)^{N-j}. \tag{2}

Define the positive number

δN=min0≤j≤N|hN(tN,j)|>0.\delta_N=\min_{0\le j\le N}|h_N(t_{N,j})|>0. \tag{3}

Also,

∑k=0N|aN,k|=2−N,|hN(x)|≤2−N(x∈ℝ),\sum_{k=0}^{N}|a_{N,k}|=2^{-N}, \qquad |h_N(x)|\le2^{-N}\quad(x\in\mathbb R), \tag{4}

and hN(x)→0h_N(x)\to0 as x→±∞x\to\pm\infty, since it is a finite sum of Gaussian translates.

2. Recursive separation of the blocks

Choose even integers LNL_N recursively. Let L1=0L_1=0. After L1,…,LN−1L_1,\ldots,L_{N-1} have been chosen, choose LNL_N so large that all the following finite families of strict inequalities hold:

LN>LN−1+2(N−1)+N2+2;L_N>L_{N-1}+2(N-1)+N^2+2; \tag{5}

|∑m=1N−1hm(LN+tN,j−Lm)|<δN4(0≤j≤N);\left|\sum_{m=1}^{N-1}h_m(L_N+t_{N,j}-L_m)\right| <\frac{\delta_N}{4} \qquad(0\le j\le N); \tag{6}

|hN(Lm+tm,j−LN)|<δm2−N−2(1≤m<N,0≤j≤m).|h_N(L_m+t_{m,j}-L_N)| <\delta_m2^{-N-2} \qquad(1\le m\lt N,\ 0\le j\le m). \tag{7}

Such an even integer exists. In (6), every summand tends to zero as LN→+∞L_N\to+\infty. In (7), the argument of the fixed function hNh_N tends to −∞-\infty. There are only finitely many inequalities at each step, and every right side is positive. One may, for definiteness, choose the least even integer satisfying (5)–(7).

This is an exact recursive definition by explicit elementary functions. No numerical approximation is required to establish existence or any subsequent conclusion.

3. The infinite shift sum

The integer intervals [LN,LN+2N][L_N,L_N+2N] are pairwise disjoint by (5). Define

cLN+2k=aN,k(N≥1,0≤k≤N),c_{L_N+2k}=a_{N,k}\qquad(N\ge1,\ 0\le k\le N),

and cl=0c_l=0 at every other integer. The disjointness makes this unambiguous. The support is contained in 2ℤ2\mathbb Z, and (4) gives

∑l∈ℤ|cl|=∑N=1∞2−N=1.\sum_{l\in\mathbb Z}|c_l|= \sum_{N=1}^{\infty}2^{-N}=1. \tag{8}

Set

f(z)=∑l∈ℤcle−π(z−l)2=∑N=1∞hN(z−LN).f(z)=\sum_{l\in\mathbb Z}c_l e^{-\pi(z-l)^2} =\sum_{N=1}^{\infty}h_N(z-L_N). \tag{9}

Both equalities are legitimate by absolute local uniform convergence. Indeed, for z=x+iyz=x+iy,

|e−π(z−l)2|=eπy2e−π(x−l)2≤eπy2.|e^{-\pi(z-l)^2}|=e^{\pi y^2}e^{-\pi(x-l)^2} \le e^{\pi y^2}.

On every compact subset of ℂ\mathbb C, this is bounded independently of ll, and (8) supplies a summable majorant. The sum is therefore entire. On the real line, |f(x)|≤1|f(x)|\le1, and its representation has c∈ℓ∞(ℤ)c\in\ell^\infty(\mathbb Z), so f∈V∞(g)f\in V^\infty(g) exactly as required in Problem 4.12.

4. Every block retains its sign changes

Fix m≥1m\ge1 and 0≤j≤m0\le j\le m, and evaluate (9) at

xm,j=Lm+tm,j.x_{m,j}=L_m+t_{m,j}.

At the step when LmL_m was chosen, (6) made the contribution from all earlier blocks less than δm/4\delta_m/4 in absolute value; for m=1m=1 that contribution is zero. Every later block N>mN>m satisfies (7), so its total contribution is bounded by

∑N=m+1∞|hN(xm,j−LN)|≤δm∑N=m+1∞2−N−2=δm2−m−2≤δm8.\sum_{N=m+1}^{\infty}|h_N(x_{m,j}-L_N)| \le\delta_m\sum_{N=m+1}^{\infty}2^{-N-2} =\delta_m2^{-m-2} \le\frac{\delta_m}{8}.

Thus

|f(xm,j)−hm(tm,j)|<3δm8<|hm(tm,j)|.|f(x_{m,j})-h_m(t_{m,j})|<\frac{3\delta_m}{8} <|h_m(t_{m,j})|. \tag{10}

The values of ff at xm,0,…,xm,mx_{m,0},\ldots,x_{m,m} consequently have the same alternating signs as in (2). In particular ff is not identically zero. Its real zeros are locally finite because it is a nonzero entire function.

By the intermediate value theorem, every one of the mm disjoint open intervals

(Lm+j+1/2,Lm+j+3/2),j=0,…,m−1,(L_m+j+1/2,\ L_m+j+3/2),\qquad j=0,\ldots,m-1,

contains a zero. Therefore

#(Z(f)∩[Lm+1/2,Lm+m+1/2])≥m.\#\bigl(Z(f)\cap[L_m+1/2,L_m+m+1/2]\bigr)\ge m. \tag{11}

This interval has length mm. As mm is arbitrary, (11) implies

D+(Z(f))≥1.D^+(Z(f))\ge1. \tag{12}

Even if the density definition is restricted to integer starting points and lengths, the containing interval [Lm,Lm+m+1][L_m,L_m+m+1] gives the ratio m/(m+1)→1m/(m+1)\to1 and the same conclusion.

5. Exact upper density of coefficient sign changes

Every index in M(c)M(c) is an index of a nonzero coefficient. Since the support of cc is contained in 2ℤ2\mathbb Z,

M(c)⊆2ℤ.M(c)\subseteq2\mathbb Z.

An interval of integer length nn contains at most n/2+1n/2+1 even integers. Hence

D+(c)≤12.D^+(c)\le\frac12. \tag{13}

Conversely, within the NNth block the consecutive nonzero coefficients

cLN,cLN+2,…,cLN+2Nc_{L_N},c_{L_N+2},\ldots,c_{L_N+2N}

have alternating signs, with the intervening odd coefficients zero. The source's definition of M(c)M(c) therefore gives

{LN,LN+2,…,LN+2(N−1)}⊆M(c).\{L_N,L_N+2,\ldots,L_N+2(N-1)\}\subseteq M(c).

The interval [LN,LN+2N][L_N,L_N+2N] has length 2N2N and contains at least NN coefficient sign changes. Letting N→∞N\to\infty yields

D+(c)≥12.D^+(c)\ge\frac12. \tag{14}

An additional sign change between two blocks can only occur at an even endpoint and does not affect either bound. Equations (13) and (14) prove D+(c)=1/2D^+(c)=1/2.

Combining this with (12) gives

D+(Z(f))≥1>12=D+(c).D^+(Z(f))\ge1>\frac12=D^+(c).

This is a counterexample to the upper-density assertion in Gröchenig's Problem 4.12. It uses one fixed Gaussian generator and one infinite coefficient sequence, with all sufficiently long density scales supplied by the recursive construction. It is not a finite-dimensional or fixed-size numerical instance.