betaFeedback

WorksMathematics

Half-line density and the structure of crystalline measures

Abstract

Let µ be a crystalline measure on the real line. Both µand its Fourier transform are locally finite sums of point masses. We prove that µ is a finite sum of translated and modulated lattice Dirac combs if and only if its support and its spectrum each have finite lower linear density on at least one half-line. The two half-lines and the sequences of radii may be chosen independently. If the support of a nonzero measure has lower half-line density d<∞, then d is the infimum of the frequency widths of the nonzero exponential polynomials vanishing on the support. In particular, the number of atoms per unit interval is uniformly bounded. If µ is not a generalized Dirac comb, its spectrum contains at least (1/(2d2)−o(1))T2 points in every interval of length T, uniformly in the position of the interval as T →∞. The proof combines a spectral gap theorem with a finite difference equation for the spectral masses. For a prescribed set Λ of lower density d<∞on one half-line, we describe the space of measures with support and spectrum in Λ and prove the sharp upper bound ⌊d2⌋for its dimension.

The counting theorems

Poisson summation identifies the Fourier transform of a lattice Dirac comb with a comb on the reciprocal lattice. Translations, modulations and finite linear combinations give a larger family of summation formulas. How can one recognize this family from the locations of the atoms alone? The criterion below uses two counts. On each side of the Fourier transform it suffices to count from the origin in a single direction, and only along a sequence of radii.

We use e(t)=exp⁡(2πit)e(t)=\exp(2\pi i t) and f^(ξ)=∫Rf(x)e(−xξ) dx\widehat{f}(\xi)=\int_{\mathbb{R}} f(x)e(-x\xi)\,\mathrm{d}x. A crystalline measure is a tempered distribution μ\mu on R\mathbb{R} such that, on compactly supported smooth test functions,

μ=∑a∈Auaδa,μ^=∑b∈Bvbδb.\mu=\sum_{a\in A}u_a\delta_a,\qquad\widehat{\mu}=\sum_{b\in B}v_b\delta_b.

Here AA and BB meet every compact interval in finitely many points, and the displayed coefficients are nonzero. Thus AA and BB are the actual support and spectrum, with each site counted once after cancellation. The coefficients may be complex. Temperedness is understood in the distributional sense throughout, including when the total variation has no polynomial growth bound.

For a locally finite set SS put NS(R)=#(S∩[−R,R])N_S(R)=\#(S\cap[-R,R]). Its lower densities on the two half-lines are

d+(S)=lim inf⁡R→∞#(S∩[0,R])R,d−(S)=lim inf⁡R→∞#(S∩[−R,0])R.(1)d_+(S)=\liminf_{R\to\infty}\frac{\#(S\cap[0,R])}{R},\qquad d_-(S)=\liminf_{R\to\infty}\frac{\#(S\cap[-R,0])}{R}. \tag*{(1)}

Write d∗(S)=min⁡{d+(S),d−(S)}d_*(S)=\min\{d_+(S),d_-(S)\}. We say that SS is relatively separated if sup⁡x#(S∩[x,x+1])<∞\sup_{x}\#(S\cap[x,x+1])<\infty. This allows distinct points to be arbitrarily close. We use

D+(S)=lim sup⁡L→∞sup⁡x∈R#(S∩[x,x+L])L(2)D^+(S)=\limsup_{L\to\infty}\sup_{x\in\mathbb{R}}\frac{\#(S\cap[x,x+L])}{L} \tag*{(2)}

for the upper uniform density. An exponential polynomial here is a finite sum P(x)=∑jcje(sjx)P(x)=\sum_j c_j e(s_jx) with real frequencies. For P≠0P\ne0 in reduced form, define its frequency width by

width⁡(P)=max⁡jsj−min⁡jsj.(3)\operatorname{width}(P)=\max_j s_j-\min_j s_j. \tag*{(3)}

Theorem 1.1 (Density on one half-line). Let μ≠0\mu\ne0 be a crystalline measure with support AA. If d=d∗(A)<∞d=d_*(A)<\infty, then d>0d>0. For every ℓ>d\ell>d there is a nonzero exponential polynomial PP of width less than ℓ\ell which vanishes on AA. In particular, AA is relatively separated. The counting densities on the positive and negative half-lines both exist and equal D+(A)=dD^+(A)=d. Moreover,

d=inf⁡{width⁡(P):P≠0 is an exponential polynomial vanishing on A}.d=\inf\{\operatorname{width}(P):P\ne0\text{ is an exponential polynomial vanishing on }A\}.

Following Meyer [10], a generalized Dirac comb is a finite linear combination of measures of the form

e(sx)∑n∈Zδa+hn,h>0,a,s∈R.e(sx)\sum_{n\in\mathbb{Z}}\delta_{a+hn},\qquad h>0,\quad a,s\in\mathbb{R}.

The summands may have different lattice spacings. For instance, the sum of the combs on Z\mathbb{Z} and 2Z\sqrt{2}\mathbb{Z} has arbitrarily close atoms. Its Fourier transform has the same property, since the two reciprocal spacings are 11 and 1/21/\sqrt{2}.

Theorem 1.2 (Quadratic lower bound for the spectrum). Let μ≠0\mu\ne0 be a crystalline measure with support AA and spectrum BB. Suppose d=d∗(A)<∞d=d_*(A)<\infty. If μ\mu is not a generalized Dirac comb, then

lim inf⁡R→∞NB(R)R2≥2d2.(4)\liminf_{R\to\infty}\frac{N_B(R)}{R^2}\ge\frac{2}{d^2}. \tag*{(4)}

The same statement holds with the support and spectrum interchanged.

Corollary 5.1 gives the corresponding estimate uniformly over all translates of the counting interval and, in particular, on each half-line separately.

Corollary 1.3 (A half-line counting criterion). A crystalline measure is a generalized Dirac comb if and only if d∗(A)<∞d_*(A)<\infty and d∗(B)<∞d_*(B)<\infty.

Proof. For a generalized Dirac comb both supports are contained in finite unions of lattice translates, by Poisson summation. Conversely, Theorem 1.1 applied to μ^\widehat{\mu} gives NB(R)=O(R)N_B(R)=O(R). For a nonzero measure, Theorem 1.2 now gives the conclusion. The zero measure is the empty sum.

Relation to earlier results. Lev and Olevskii proved rigidity for measures on the line with uniformly discrete support and spectrum [8]. Their later Theorem 2.2 in [9] already permits bounded unit-interval counts on the spectral side, assuming a positive minimum spacing on the physical side. Baake, Strungaru and Terauds [2 Theorem 6.6] derived a representation on translates of a common lattice when the support lies in a Meyer set and the spectrum has finite upper density along a van Hove sequence. In our one-dimensional setting, Corollary 1.3 allows both sets to have arbitrarily small spacings. Theorem 1.2 also allows the spectrum to have unbounded unit-interval counts.

For the spectral gap argument, Lev and Olevskii [9 Proposition 6.1] obtain a positive lower bound for the centered counting density of a set that supports a nonzero slowly increasing measure with a spectral gap. Their bound depends on the gap length and a uniform bound on the unit-interval counts. In Lemma 2.1 we adapt the oscillation theorem of Eremenko and Novikov [3 Theorem 1] to count locally finite atoms. This gives the gap length itself as a lower bound on each half-line, assuming only distributional temperedness and local finiteness of the atom set.

Meyer [11 Theorem 6] showed that if the support is uniformly discrete, a nonzero compactly supported continuous function annihilates μ^\widehat{\mu} under convolution. The proof following Lemma 15 of [11] constructs such a function for finite unions of uniformly discrete sets by convolving annihilators for the individual sets. Theorem 1.1 gives an annihilator with finitely many point masses in an interval of any length greater than d∗(A)d_{*}(A). Meyer’s Lemma 14 studies the support of a solution to a particular finite-difference equation. Proposition 3.1 gives a counting alternative for arbitrary finite shifts with nonzero extreme coefficients.

Olevskii and Ulanovskii characterize supports of unit-mass Fourier quasicrystals as zero sets of exponential polynomials [12]. Alon and Kummer obtain a Lee–Yang description for uniformly discrete crystalline measures with unit weights. For arbitrary weights on uniformly discrete supports they also prove finite generation of the spectral group and polynomial vanishing [1 Theorems 3.1 and 4.2]. Their Lemma 4.7 estimates the dimension of a space of modulations by restricting the Fourier transform to a fixed window. Here a half-line count determines how long that interval must be. By adjoining one frequency to the group generated by the relevant spectral sites, we obtain a polynomial whose width is arbitrarily close to the lower half-line density. Its zero set contains the support and may contain additional points.

Gonçalves and Vedana parameterize summation formulas under weighted growth conditions by almost periodic holomorphic functions [4]. The counting conditions above concern the locations of nonzero atoms and are independent of their sizes. § 6 shows that one finite union of lattice translates contains the supports and spectra of all measures carried on both sides by a prescribed set of finite lower half-line density. The points of that union missing from the prescribed set impose finitely many linear conditions. This gives a reduction of the dimension question raised by Meyer [10 Section 6] and a sharp upper bound in terms of the lower half-line density.

Outline of the argument. A gap in the Fourier transform forces a minimum density of atoms on each half-line. We use this fact to turn a relation visible on one bounded spectral interval into an identity on the whole line. The resulting finite difference equation constrains the spectral masses. Either their locations lie in a finite union of lattice translates, or multiplying the masses by polynomials produces enough independent solutions to force quadratic growth. We prove this alternative in § 3. § 4 recovers the weights on the lattices and § 5 proves Theorem 1.2 and Corollary 5.1.

From a half-line count to an exponential polynomial

A gap in frequency forces many atoms in physical space, even if only one half-line is observed. We use the oscillation theorem of Eremenko and Novikov, who also discuss its extension to tempered distributions [3 Theorem 1 and the subsequent discussion]. The proof below avoids a growth assumption on the total variation of the measure.

Lemma 2.1 (Point counts in the presence of a spectral gap). Let T≠0T \ne0 be a tempered distribution that is locally a sum of point masses on a locally finite set AA. If T^\widehat{T} vanishes on a bounded open interval II, then

lim inf⁡R→∞#(A∩[0,R])R≥∣I∣,lim inf⁡R→∞#(A∩[−R,0])R≥∣I∣.\liminf_{R\to\infty}\frac{\#(A\cap[0,R])}{R}\ge|I|,\qquad\liminf_{R\to\infty}\frac{\#(A\cap[-R,0])}{R}\ge|I|.

Proof. Modulate TT to center the gap at the origin, and write it as (−h,h)(-h,h). After multiplication by a constant of modulus one, its real part UU is nonzero. Taking the real part replaces the Fourier transform by its average with its conjugate reflection, so the symmetric gap persists. The atom set of UU is contained in AA.

We need a real, continuous function ff of polynomial growth with f(m)=Uf^{(m)}=U for some m≥2m\ge2, whose Fourier transform still has the gap. Here is one construction. Temperedness gives integers p,qp,q and a bound

∣U(φ)∣≤Cmax⁡0≤j≤qsup⁡t∈R(1+t2)p∣φ(j)(t)∣.|U(\varphi)|\le C\max_{0\le j\le q}\sup_{t\in\mathbb{R}}(1+t^2)^p|\varphi^{(j)}(t)|.

Apply Hahn–Banach to the map φ↦((1+t2)pφ(j))j=0q\varphi\mapsto((1+t^2)^p\varphi^{(j)})_{j=0}^q into a finite product of copies of C0(R)C_0(\mathbb{R}). The Riesz representation theorem then writes U=∑j=0qDjνjU=\sum_{j=0}^q D^j\nu_j, where the real measures νj\nu_j satisfy ∫(1+t2)−p d∣νj∣(t)<∞\int(1+t^2)^{-p}\,d|\nu_j|(t)<\infty. Each νj\nu_j has a locally bounded primitive of polynomial growth, obtained by taking its cumulative mass from the origin. One more integration gives a continuous primitive. Integrating the terms to a common order m=q+2m=q+2 thus gives a real continuous FF of polynomial growth with DmF=UD^mF=U.

On (−h,h)(-h,h), the distribution F^\widehat{F} is killed by ξm\xi^m. Its restriction there is consequently a linear combination of δ0,…,δ0(m−1)\delta_0,\ldots,\delta_0^{(m-1)}. Subtract the inverse transform of this combination from FF. The subtracted function is a real polynomial of degree less than mm, by conjugate symmetry. The resulting ff has the required gap and Dmf=UD^mf=U.

Let s(f;[0,R])s(f;[0,R]) denote the greatest number of successive sign changes at ordered points where ff is nonzero. Eremenko and Novikov use the minimum degree of a nonzero real polynomial pp for which pfpf is nonnegative on (0,R)(0,R). For the continuous, piecewise polynomial function ff here the counts agree. Each sign transition forces a root of pp, and one root placed at each transition suffices. Continuity allows the testing points to be chosen in the open interval. The oscillation theorem of Eremenko and Novikov [3 Theorem 1] gives

lim inf⁡R→∞s(f;[0,R])R≥2h.\liminf_{R\to\infty}\frac{s(f;[0,R])}{R}\ge2h.

Indeed, if ∣f(t)∣=O((1+∣t∣)b)|f(t)|=O((1+|t|)^b) with b≥0b\ge0, the uniformly continuous weight ω(t)=(b+2)log⁡(2+∣t∣)\omega(t)=(b+2)\log(2+|t|) satisfies ∫Rω(t)/(1+t2) dt<∞\int_{\mathbb{R}}\omega(t)/(1+t^2)\,dt<\infty and ∫R∣f(t)∣e−ω(t) dt<∞\int_{\mathbb{R}}|f(t)|e^{-\omega(t)}\,dt<\infty. Their angular-frequency half-width is 2πh2\pi h in our convention.

To bound the sign changes from above, consider an interval slightly larger than [0,R][0,R]. There ff is the sum of a polynomial of degree at most m−1m-1 and the functions ua(x−a)m−1/(m−1)!u_a(x-a)^{m-1}/(m-1)! for the finitely many atoms uaδau_a\delta_a in that interval. This follows by taking mm distributional derivatives. If an absolutely continuous real function alternates sign at k+1k+1 ordered points, its kk successive increments alternate sign. Its derivative takes the respective signs on sets of positive measure in the intervening intervals. Each differentiation therefore costs at most one sign change. The derivatives through order m−2m-2 are locally absolutely continuous. After m−1m-1 differentiations we take the almost-everywhere step-function representative and choose the sign-testing points away from its jumps. Its jumps in [0,R][0,R] occur only at atoms. Thus

s(f;[0,R])≤#(A∩[0,R])+m−1.(5)s(f;[0,R])\le\#(A\cap[0,R])+m-1. \tag*{(5)}

Division by RR proves the first inequality. Reflection gives the second.

In particular, any interval of length greater than d∗(A)d_{*}(A) detects whether an exponential multiple PμP\mu vanishes. If Pμ^\widehat{P\mu} vanishes there, then Pμ=0P\mu=0. Indeed, multiplying by an exponential polynomial preserves temperedness and can only remove atoms from AA. We now construct a nonzero polynomial PP for which the Fourier transform of PμP\mu vanishes on such an interval.

Proposition 2.2. Suppose μ\mu is a crystalline measure with support AA, and d∗(A)<∞d_{*}(A)<\infty. For each ℓ>d∗(A)\ell>d_{*}(A) there is a nonzero exponential polynomial PP of frequency width less than ℓ\ell which vanishes on AA.

Proof. Choose J,L>0J,L>0 with d∗(A)<2J<2L<ℓd_{*}(A)<2J<2L<\ell, and let B=supp⁡μ^B=\operatorname{supp}\widehat{\mu}. We seek frequencies in [−L,L][-L,L]. Only the finitely many points of B0=B∩[−J−L,J+L]B_{0}=B\cap[-J-L,J+L] can be shifted by these frequencies into [−J,J][-J,J].

Choose a Z\mathbb{Z}-basis β1,…,βr\beta_{1},\ldots,\beta_{r} for the additive group generated by B0B_{0}. Choose 0<ε<L−J0<\varepsilon<L-J outside their rational linear span. For c>0c>0 and an integer N≥0N\geq0, put

Ec(N)={β⋅n+εk:n∈Zr, ∥n∥∞≤N, k∈Z, ∣β⋅n+εk∣≤c}.E_{c}(N)=\{\beta\cdot n+\varepsilon k:n\in\mathbb{Z}^{r},\ \lVert n\rVert_{\infty}\leq N,\ k\in\mathbb{Z},\ \lvert\beta\cdot n+\varepsilon k\rvert\leq c\}.

Each frequency has a unique representation. For fixed nn, the eligible integers kk lie in an interval of length 2c/ε2c/\varepsilon. Hence

(2N+1)r(2c/ε−1)≤∣Ec(N)∣≤(2N+1)r(2c/ε+1).(2N+1)^{r}(2c/\varepsilon-1)\leq\lvert E_{c}(N)\rvert\leq(2N+1)^{r}(2c/\varepsilon+1).

These bounds also hold for the trivial group, with r=0r=0. Choose an integer H≥0H\geq0 bounding all coordinates of all points of B0B_{0} in absolute value. Take H=0H=0 when there are no coordinates. All possible points in (B+EL(N))∩[−J,J](B+E_{L}(N))\cap[-J,J] lie in EJ(N+H)E_{J}(N+H). Since 2L/ε−1>2J/ε+12L/\varepsilon-1>2J/\varepsilon+1, for large NN the latter set contains fewer points than EL(N)E_{L}(N).

Fix such an NN, and write by=μ^({y})b_y=\widehat{\mu}(\{y\}), taking by=0b_y=0 off BB. The equations

∑s∈EL(N)csby−s=0,y∈(B+EL(N))∩[−J,J],\sum_{s\in E_{L}(N)}c_s b_{y-s}=0,\qquad y\in(B+E_{L}(N))\cap[-J,J],

have a nonzero solution because there are fewer equations than unknowns. The distinct characters in P(x)=∑scse(sx)P(x)=\sum_s c_s e(sx) are linearly independent, so PP is nonzero. The equations say exactly that the mass of Pμ^\widehat{P\mu} vanishes at every possible point of [−J,J][-J,J]. Lemma 2.1 now gives Pμ=0P\mu=0. Each atom of μ\mu has nonzero mass, hence PP vanishes on AA. Its frequency width is at most 2L<ℓ2L<\ell.

Lemma 2.3 (Zero count for an exponential polynomial). For a nonzero exponential polynomial PP of frequency width ww, its number of real zeros in any interval of length TT, counted with multiplicity, is at most wT+CPwT+C_P.

Proof. This is the elementary real-frequency case of Langer’s zero count [6]. Write P(z)=∑j=1kcje(sjz)P(z)=\sum_{j=1}^{k}c_j e(s_jz), with s1<⋯<sks_1<\cdots<s_k and all cj≠0c_j\ne0. On Im⁡z=H\operatorname{Im}z=H, for fixed sufficiently large HH, the first summand dominates all the others combined. On Im⁡z=−H\operatorname{Im}z=-H, the last summand does. Along a horizontal segment of length TT, oriented from left to right, the changes of argument are therefore 2πs1T+O(1)2\pi s_1T+O(1) at the top and 2πskT+O(1)2\pi s_kT+O(1) at the bottom. The error is uniform in the endpoints, since the quotient by the dominant term stays in a disk about 11 of radius less than 11.

On any zero-free vertical edge z=x+iyz=x+iy, rotate the values by a constant phase so that their imaginary part is not identically zero. It is a real linear combination of the kk functions exp⁡(−2πsjy)\exp(-2\pi s_j y), so it has at most k−1k-1 zeros, by division by the first exponential and induction using Rolle’s theorem. Between consecutive zeros the argument changes by at most π\pi. The vertical contributions to the argument principle are thus bounded independently of the edge. Enlarge the real interval by less than one at each end so that the vertical edges avoid zeros. The argument principle on the resulting rectangle bounds the number of enclosed zeros by wT+OP(1)wT+O_P(1).

Completion of the proof of Theorem 1.1. If d∗(A)<∞d_*(A)<\infty, Proposition 2.2 and Lemma 2.3 give D+(A)≤ℓD^+(A)\le\ell for every ℓ>d∗(A)\ell>d_*(A). The zero count with T=1T=1 also bounds the number of points of AA in every unit interval. Conversely, each half-line counting ratio has limsup at most D+(A)D^+(A), since its interval is among those in the uniform supremum. Both ray ratios therefore converge to d∗(A)=D+(A)d_*(A)=D^+(A). The two preceding results also identify this common value with the infimum of the widths of nonzero exponential polynomials vanishing on AA. Finally, a nonzero crystalline measure has strictly positive lower density on each ray. Its locally finite spectrum leaves some nonempty open interval free of atoms, to which Lemma 2.1 applies.

A counting alternative for a finite difference equation

The following proposition concerns a function on the real line. Its proof uses only the positions of its nonzero values and a linear relation among its translates. The constant comes from counting monomials in two variables.

Proposition 3.1. Let u:R→Cu:\mathbb{R}\to\mathbb{C} have locally finite support S={t:u(t)≠0}S=\{t:u(t)\ne0\}. Suppose that

(Du)(t):=∑j=0kaju(t−sj)=0(t∈R),s0<s1<⋯<sk,aj≠0.(Du)(t):=\sum_{j=0}^{k}a_j u(t-s_j)=0\quad(t\in\mathbb{R}),\qquad s_0<s_1<\cdots<s_k,\quad a_j\ne0.

If W=sk−s0>0W=s_k-s_0>0, then either SS is contained in a finite union of translates of lattices in R\mathbb{R}, or

lim inf⁡R→∞#(S∩[−R,R])R2≥2W2.\liminf_{R\to\infty}\frac{\#(S\cap[-R,R])}{R^2}\ge\frac{2}{W^2}.

Proof. Translating the equation, we may take s0=0s_0=0 and sk=Ws_k=W. We may also assume S≠∅S\ne\varnothing.

Integer coordinates for the support. If t∈St\in S, the equation at tt shows that t−sj∈St-s_j\in S for at least one j>0j>0. The equation at t+Wt+W shows that t+W−sj∈St+W-s_j\in S for at least one j<kj<k. In either direction, the step is positive, at most WW, and bounded away from zero. Starting at any t>Wt>W in SS and stepping to the left, we must therefore reach S∩[0,W]S\cap[0,W]. From t<0t<0 we reach the same interval by stepping to the right.

It follows that SS is contained in the additive subgroup GG generated by the shifts sjs_j and the finite set S∩[0,W]S\cap[0,W]. This group is finitely generated and torsion-free. Fix a basis over Z\mathbb{Z} and write n(t)∈Zrn(t)\in\mathbb{Z}^r for the coordinates of t∈Gt\in G. Denote the inverse coordinate map by ℓ:Zr→G\ell:\mathbb{Z}^r\to G. The same stepping argument also proves uniqueness from an interval. If a solution of a difference equation of width ww vanishes on an open interval of length greater than ww, it vanishes everywhere. A chain of nonzero values starting on either side would have to enter that interval.

Polynomial multiples and observation on an interval. For a polynomial pp in rr variables, set up(t)=p(n(t))u(t)u_p(t)=p(n(t))u(t) for t∈Gt\in G, and set up(t)=0u_p(t)=0 elsewhere. If deg⁡p≤m\deg p\le m, then

Dm+1up=0.D^{m+1}u_p=0.

For constants this is the original equation. For the induction step, subtract p(n(t))(Du)(t)p(n(t))(Du)(t) from (Dup)(t)(D u_p)(t) for t∈Gt \in G. The summand corresponding to sjs_j becomes

aj[p(n(t−sj))−p(n(t))]u(t−sj).a_j\left[p(n(t-s_j))-p(n(t))\right]u(t-s_j).

The expression in brackets is a polynomial of degree at most m−1m-1 in n(t−sj)n(t-s_j). Thus DupD u_p is a sum of translates of lower-degree polynomial multiples of uu. The induction hypothesis applies since DD commutes with translations. Outside GG every summand is zero.

The extreme shifts of Dm+1D^{m+1} are 00 and (m+1)W(m+1)W, with nonzero coefficients a0m+1a_0^{m+1} and am+1m+1a_{m+1}^{m+1}. Consequently, if II is any open interval of length greater than (m+1)W(m+1)W, the values of upu_p on II determine upu_p everywhere. Put E=n(S)E=n(S), and let H(m)H(m) be the dimension of the space of restrictions to EE of polynomials of total degree at most mm. Multiplication by uu preserves this dimension, since uu is nonzero at every point of SS. Restriction to II is injective on these multiples, so

H(m)≤#(S∩I).(6)H(m) \le\#(S \cap I). \tag*{(6)}

Quadratic growth or a union of lines. Let V⊂CrV \subset\mathbb{C}^r be the Zariski closure of EE, the common zero set of all polynomials vanishing on EE. If an irreducible component of VV has dimension at least two, two coordinate functions on it are algebraically independent. Indeed, the coordinate functions generate its function field, whose transcendence degree is its dimension. The monomials of degree at most mm in these two coordinates are therefore linearly independent on EE. Indeed, a relation on EE would hold on VV and hence on this component. We obtain H(m)≥(m+1)(m+2)/2H(m) \ge(m+1)(m+2)/2. In (5), take I=(−R,R)I=(-R,R) and m=⌊2R/W⌋−2m=\lfloor2R/W\rfloor-2 for all sufficiently large RR. Division by R2R^2 gives the asserted lower bound.

It remains to treat the case in which every component is a curve or a point. By the forward step above, for every n∈En \in E there is a vector

v∈F:={n(W−sj):0≤j<k}v \in F := \{n(W-s_j): 0 \le j < k\}

with n+v∈En+v \in E.

Here FF is finite and ℓ(v)>0\ell(v)>0 for every v∈Fv \in F. Taking Zariski closures gives V⊂⋃v∈F(V−v)V \subset\bigcup_{v\in F}(V-v). If CC is an irreducible curve component of VV, it lies in C′−vC'-v for some component C′C' and some v∈Fv \in F. The component C′C' must itself be a curve, and the inclusion is equality, since every proper closed subset of an irreducible algebraic curve is finite.

Starting at CC we may thus repeatedly pass to a component that is a translate of the preceding one by a vector from FF. There are finitely many components, so one repeats. The repeated curve is invariant under translation by the sum v∗v_* of the intervening vectors. Since ℓ(v∗)>0\ell(v_*)>0, this sum is nonzero. For any point zz on the repeated curve, it contains z+nv∗z+n v_* for every positive integer nn. Every polynomial vanishing on the curve therefore vanishes identically on the complex line z+Cv∗z+\mathbb{C}v_*. The curve equals that line. Translating back along the chain shows that CC was a line as well.

Finally, the integer points on any complex affine line are empty, a singleton, or a translate of a cyclic subgroup of Zr\mathbb{Z}^r. Indeed, if there are two such points, their difference defines a primitive integer direction for the line. Applying ℓ\ell sends these integer points into a translated lattice in R\mathbb{R}. The finitely many point components can also be covered by translated lattices. This proves the other alternative.

From lattice supports to lattice combs

Support on several lattices allows their spacings to have irrational ratios. Consequently the different lattices can come arbitrarily close. The following argument works directly with the coefficients at their points.

Proposition 4.1. Let μ\mu be a crystalline measure on R\mathbb{R}. Suppose that each of supp⁡μ\operatorname{supp}\mu and supp⁡μ^\operatorname{supp}\widehat{\mu} is contained in a finite union of translates of lattices. Then μ\mu is a finite linear combination of translated and modulated lattice Dirac combs.

Proof. Put ν=μ^\nu=\widehat{\mu}, and cover supp⁡μ\operatorname{supp}\mu by the sets aj+hjZa_j+h_j\mathbb{Z}, with hj>0h_j>0. The function

P(x)=∏j(e(x/hj)−e(aj/hj))P(x)=\prod_j\left(e(x/h_j)-e(a_j/h_j)\right)

vanishes on supp⁡μ\operatorname{supp}\mu. Write τtv(x)=v(x−t)\tau_t v(x)=v(x-t) for translation of the coefficient function of a point measure. Fourier transformation of Pμ=0P\mu=0 gives

Qν=0,Q=∏j(τtj−zj),tj=hj−1,zj=e(aj/hj).Q\nu=0,\qquad Q=\prod_j(\tau_{t_j}-z_j),\qquad t_j=h_j^{-1},\qquad z_j=e(a_j/h_j).

Group the lattices covering supp⁡ν\operatorname{supp}\nu according to the rational ratios of their spacings. Write ν=∑iνi\nu=\sum_i\nu_i accordingly, assigning an atom at an intersection to one group. For now we use each νi\nu_i only through its coefficients. For each ii, its support is contained in finitely many cosets of hZh\mathbb{Z} for some h>0h>0 depending on ii. The same holds for QνiQ\nu_i. Lattices from different groups still intersect in at most one point after arbitrary translations, and only finitely many translations occur in QQ. Since ∑iQνi=0\sum_iQ\nu_i=0, every nonzero coefficient of QνiQ\nu_i occurs at one of these finitely many intersections. Thus QνiQ\nu_i has finite support.

Fix ii. We can remove from QQ every factor with tj/h∉Qt_j/h\notin\mathbb{Q} while preserving this conclusion. Indeed, suppose ww is supported on finitely many cosets of hZh\mathbb{Z}, and (τt−z)w(\tau_t-z)w has finite support, where t/h∉Qt/h\notin\mathbb{Q} and z≠0z\ne0. The translates of these cosets by ktkt are disjoint from the original cosets for some positive integer kk. For each pair of cosets at most one kk can give an intersection. By telescoping, (τkt−zk)w(\tau_{kt}-z^k)w has finite support. Its two terms have disjoint supports, so ww has finite support. Applying this observation successively is legitimate because each remaining product of translations preserves containment in finitely many cosets of hZh\mathbb{Z}.

Choose a finer lattice h′Zh'\mathbb{Z}, h′>0h'>0, so that hh and all remaining tjt_j are integer multiples of h′h'. Along each of the finitely many relevant cosets of h′Zh'\mathbb{Z} the coefficients of νi\nu_i now satisfy, outside a finite set, a constant-coefficient linear recurrence. Write v(n)v(n) for the coefficient at x0+nh′x_0+nh' on a fixed coset and let (Tv)(n)=v(n−1)(Tv)(n)=v(n-1) be the backward shift. With mj=tj/h′m_j=t_j/h', the remaining factors give

Qi(T)=∏j remaining(Tmj−zj).Q_i(T)=\prod_{j\ \mathrm{remaining}}(T^{m_j}-z_j).

For a solution v(n)=λnv(n)=\lambda^n of the homogeneous recurrence, Qi(λ−1)=0Q_i(\lambda^{-1})=0. Since all roots of QiQ_i have modulus one, so do the exponential bases λ\lambda. The elementary solution formula for linear recurrences shows that the positive tail is a sum of terms p(n)λnp(n)\lambda^n, with pp a polynomial and ∣λ∣=1|\lambda|=1. An empty product gives the zero tail.

Extend these formulas to all n∈Zn\in\mathbb{Z} on every coset and sum over the groups. The resulting measure σ\sigma is tempered, by polynomial growth, and agrees with ν\nu on a right half-line. By Poisson summation, σ^\widehat{\sigma} is a finite sum of modulated derivatives of translated lattice combs. Put η=ν−σ\eta=\nu-\sigma. Its transform η^=ν^−σ^\widehat{\eta}=\widehat{\nu}-\widehat{\sigma} has locally finite support, hence vanishes on some open interval. On the positive ray the support of η\eta is finite. Lemma 2.1 gives η=0\eta=0, so ν=σ\nu=\sigma.

Finally, ν^\widehat{\nu} is the reflection of μ\mu and consists of point masses. Expand the smooth modulation factors multiplying the derivatives of Dirac masses by the product rule. At each site the derivatives of different orders are linearly independent, so all positive-order terms cancel. The order-zero terms are modulated lattice combs, since every derivative of an exponential is a constant multiple of that exponential. Reflection gives the asserted form of μ\mu. □\square

Proof of the quadratic bound and its consequences

Proof of Theorem 1.2. Write d=d∗(A)d=d_{*}(A) and suppose that the left side of (4) is strictly smaller than 2/d22/d^{2}. Choose ℓ>d\ell>d so close to dd that

lim inf⁡R→∞NB(R)R2<2ℓ2.\liminf_{R\to\infty}\frac{N_{B}(R)}{R^{2}}<\frac{2}{\ell^{2}}.

By Theorem 1.1, some nonzero exponential polynomial PP of width W<ℓW<\ell vanishes on AA. A single exponential has no zeros, so W>0W>0. Taking Fourier transforms of Pμ=0P\mu=0 gives a finite difference equation of width WW for the function b↦vbb\mapsto v_b, extended by zero off BB. Proposition 3.1 forces BB to lie in a finite union of lattice translates.

We also need a lattice cover on the physical side. Choose a cover B⊂⋃j=1m(bj+wjZ)B\subset\bigcup_{j=1}^{m}(b_j+w_j\mathbb{Z}), and take the product of the factors e(ξ/wj)−e(bj/wj)e(\xi/w_j)-e(b_j/w_j). This nonzero exponential polynomial kills μ^\widehat{\mu}. Its transform gives a difference equation of positive width for the coefficients of μ^\widehat{\mu} on −A-A. By Theorem 1.1, AA is relatively separated, hence N−A(R)=O(R)N_{-A}(R)=O(R). A second application of Proposition 3.1 covers −A-A, and thus AA, by finitely many lattice translates. Proposition 4.1 now identifies μ\mu as a generalized Dirac comb. Reflection preserves centered counts and interchanges the two half-line densities, so the argument applies in the opposite direction as well.

Corollary 5.1 (Quadratic growth in every interval). Under the hypotheses of Theorem 1.2, if μ\mu is not a generalized Dirac comb, then

lim inf⁡T→∞inf⁡x∈R#(B∩[x,x+T])T2≥12d2.\liminf_{T\to\infty}\inf_{x\in\mathbb{R}}\frac{\#(B\cap[x,x+T])}{T^{2}}\ge\frac{1}{2d^{2}}.

In particular, each spectral half-line satisfies

lim inf⁡R→∞#(B∩[0,R])R2≥12d2,lim inf⁡R→∞#(B∩[−R,0])R2≥12d2.\liminf_{R\to\infty}\frac{\#(B\cap[0,R])}{R^{2}}\ge\frac{1}{2d^{2}},\qquad \liminf_{R\to\infty}\frac{\#(B\cap[-R,0])}{R^{2}}\ge\frac{1}{2d^{2}}.

The conclusions also hold with the support and spectrum interchanged.

Proof. The second paragraph of the proof of Theorem 1.2 shows that a finite lattice cover of BB would force μ\mu to be a generalized Dirac comb. Thus BB has no such cover. For any ℓ>d\ell>d, choose an annihilator of width 0<W<ℓ0<W<\ell as in Theorem 1.1. The resulting recurrence for the spectral coefficients falls in the quadratic case of Proposition 3.1. For any sufficiently large TT put m=⌊T/W⌋−2m=\lfloor T/W\rfloor-2, so that T>(m+1)WT>(m+1)W. The restriction estimate (6) and the monomial count in that proof give, for every x∈Rx\in\mathbb{R},

#(B∩[x,x+T])≥#(B∩(x,x+T))≥(m+1)(m+2)2.\#(B\cap[x,x+T])\ge\#(B\cap(x,x+T))\ge\frac{(m+1)(m+2)}{2}.

The right side is independent of xx. Dividing by T2T^{2}, taking the infimum over xx and then the lower limit in TT gives at least 1/(2W2)1/(2W^{2}), hence at least 1/(2ℓ2)1/(2\ell^{2}). Let ℓ↓d\ell\downarrow d. Taking x=0x=0 and x=−Tx=-T yields the half-line estimates. Reflection gives the assertion in the opposite Fourier orientation.

The quadratic scale is attained. Kurasov and Sarnak [5 Theorems 1 and 2] construct a positive unit-mass crystalline measure whose support is uniformly discrete and which is not a generalized Dirac comb. In their example the spectrum is contained in

{0}∪{±(mu+nv):m,n∈Z≥0, m+n>0}.\{0\}\cup\{\pm(mu+nv):m,n\in\mathbb{Z}_{\ge0},\ m+n>0\}.

for two positive frequencies uu, vv, after choosing the Fourier normalization. Counting the integer pairs with mu+nv≤Rmu+nv\le R gives NB(R)=O(R2)N_B(R)=O(R^2). Theorem 1.2 gives the matching lower order bound for the actual spectrum. Counting all semigroup points gives an upper bound even if some Fourier coefficients vanish. The example therefore establishes optimality of the exponent. The optimal universal coefficient in (4) remains to be determined.

The case of two subquadratic counts. Suppose that NA(R)+NB(R)=o(R2)N_A(R)+N_B(R)=o(R^2). If any one of the four half-line lower densities is finite, Theorem 1.2, applied in the appropriate direction, forces μ\mu to be a generalized Dirac comb. Thus a counterexample to rigidity under this joint bound would have to satisfy

#(S∩[0,R])R⟶∞and#(S∩[−R,0])R⟶∞(S=A,B).\frac{\#(S\cap[0,R])}{R}\longrightarrow\infty \quad\text{and}\quad \frac{\#(S\cap[-R,0])}{R}\longrightarrow\infty \qquad(S=A,B).

Whether such a crystalline measure exists is open.

Measures with prescribed support and spectrum

For a locally finite set Λ⊂R\Lambda\subset\mathbb{R}, let MΛM_\Lambda be the complex vector space of crystalline measures μ\mu such that supp⁡μ⊂Λ\operatorname{supp}\mu\subset\Lambda and supp⁡μ^⊂Λ\operatorname{supp}\widehat{\mu}\subset\Lambda. Meyer [10 Section 6] asked which sets admit a nonzero measure and how the dimension depends on the geometry of the set. For d∗(Λ)<∞d_*(\Lambda)<\infty, Corollary 1.3 identifies the measures as generalized Dirac combs. We show that their supports and spectra have a common cover by finitely many lattice translates, with only finitely many points of the cover outside Λ\Lambda. This reduces the dimension question to the space on that cover and the linear conditions at its missing points. Meyer also asked about the sets {±k+θ:k∈N}\{\pm\sqrt{k}+\theta:k\in\mathbb{N}\} for 0<θ<10<\theta<1. Their half-line counts grow quadratically, so that question lies outside the hypothesis below.

Let LL range over finite unions of full lattice translates for which FL=L∖ΛF_L=L\setminus\Lambda is finite. Include the empty set, with M∅=0M_\varnothing=0. On MLM_L define the map

EFL:μ⟼(μ({x}),μ^({x}))x∈FL.(7)E_{F_L}:\mu\longmapsto\bigl(\mu(\{x\}),\widehat{\mu}(\{x\})\bigr)_{x\in F_L}. \tag*{(7)}

Its rank counts independent conditions. In particular, a single missing point may impose two independent conditions.

Theorem 6.1. If d=d∗(Λ)<∞d=d_*(\Lambda)<\infty, then every member of MΛM_\Lambda is a generalized Dirac comb and dim⁡CMΛ≤⌊d2⌋\dim_{\mathbb{C}}M_\Lambda\le\lfloor d^2\rfloor. One finite union LL of lattice translates contains the supports and spectra of all members of MΛM_\Lambda, and FL=L∖ΛF_L=L\setminus\Lambda is finite. For this LL, the space MΛM_\Lambda is the kernel of (7) in MLM_L. In particular,

dim⁡CMΛ=max⁡L: #(L∖Λ)<∞(dim⁡CML−rank⁡EFL),(8)\dim_{\mathbb{C}}M_\Lambda= \max_{L:\,\#(L\setminus\Lambda)<\infty} \bigl(\dim_{\mathbb{C}}M_L-\operatorname{rank}E_{F_L}\bigr), \tag*{(8)}

where LL ranges over the unions specified above.

Proof. Both actual supports of every μ∈MΛ\mu\in M_\Lambda inherit the half-line bound. Corollary 1.3 makes μ\mu a generalized Dirac comb. For every bounded open interval II with ∣I∣>d|I|>d, the map

μ⟼(μ^({b}))b∈Λ∩I\mu\longmapsto\bigl(\widehat{\mu}(\{b\})\bigr)_{b\in\Lambda\cap I}

is injective by Lemma 2.1. Thus k=dim⁡CMΛk=\dim_{\mathbb{C}}M_\Lambda is finite and every open interval of length T>dT>d contains at least kk points of Λ\Lambda. In either [0,R][0,R] or [−R,0][-R,0] one can fit ⌊R/T⌋\lfloor R/T\rfloor disjoint open intervals of length TT. Taking the lower density on each ray gives d≥k/Td \ge k/T. Hence k≤dTk \le dT for every T>dT > d. Letting T↓dT \downarrow d proves k≤⌊d2⌋k \le\lfloor d^2 \rfloor. The same reasoning gives finite dimension for MLM_L.

We record why finitely many missing points suffice. The support of any generalized Dirac comb is a finite union of full lattice translates with at most finitely many points removed. To see this, group its terms by commensurable spacings. Within a group, take a positive common multiple of the finitely many spacings and split each lattice into cosets of that common period. On each coset the coefficient sequence has the form ∑jcjλjn\sum_j c_j\lambda_j^n, with ∣λj∣=1|\lambda_j| = 1. By the Skolem–Mahler–Lech theorem [7], its zeros on the integers consist of finitely many full arithmetic progressions and a finite set. For the two-sided formulation, apply the theorem on each half-line. Vanishing on a tail of a progression implies vanishing on the entire progression by the invertible recurrence for its subsequence. Removing the full zero progressions from the coset leaves a finite union of full subprogressions. The remaining zeros remove only finitely many points. Between different groups only finitely many intersection sites can undergo further cancellation. Poisson summation gives the same description for the spectrum.

Take a basis of MΛM_\Lambda. For the support and spectrum of each basis vector choose the finite lattice union just obtained. Their union LL contains all these supports and spectra, and L∖ΛL \setminus\Lambda is finite. It follows that MΛM_\Lambda is precisely the kernel of EFLE_{F_L} in MLM_L. For every other eligible L′L' the corresponding kernel is a subspace of MΛM_\Lambda. Rank–nullity proves (8), including attainment. For the zero space take L=∅L = \varnothing.

Once an attaining LL is known, the two coefficient conditions at each point of FLF_L form a finite linear system on MLM_L. Finding such an LL from a description of Λ\Lambda is a separate problem. The proof obtains LL from the supports and spectra of a basis of MΛM_\Lambda.

Example 6.2 (Deleting the origin). The dimension bound is sharp for L=q−1/2ZL = q^{-1/2}\mathbb{Z}, where qq is a positive integer. For μ∈ML\mu\in M_L, multiplication by e(q ξ)e(\sqrt{q}\,\xi) fixes μ^\widehat{\mu}, so μ\mu is periodic with period q\sqrt{q}. Its coefficients on LL are specified by qq numbers repeated periodically. Conversely, every such choice belongs to MLM_L by Poisson summation. Thus dim⁡CML=q=d∗(L)2\dim_{\mathbb{C}} M_L = q = d_*(L)^2.

Now take q=3q = 3 and Λ=L∖{0}\Lambda= L \setminus\{0\}. Write (c0,c1,c2)(c_0,c_1,c_2) for the three repeated coefficients. The physical mass at zero is c0c_0, and the Fourier mass there is (c0+c1+c2)/3(c_0+c_1+c_2)/\sqrt{3}. The two conditions are independent. Thus dim⁡CML=3\dim_{\mathbb{C}} M_L = 3 and dim⁡CMΛ=1\dim_{\mathbb{C}} M_\Lambda= 1. The kernel is spanned by the measure with repeated weights (0,1,−1)(0,1,-1), the character modulo three used by Meyer [10 Lemma 4.2].

A carrier with unbounded unit-interval counts. Take Λ=Z∪{−n:n=1,2,…}\Lambda= \mathbb{Z} \cup\{-\sqrt{n}: n = 1,2,\ldots\}. Its positive half-line has density one, while [−m−1,−m][-m-1,-m] contains 2m+22m+2 distinct points of Λ\Lambda for each integer m≥1m \ge1. In particular, d∗(Λ)=1d_*(\Lambda) = 1 and the unit-interval counts are unbounded. The integer comb belongs to MΛM_\Lambda by Poisson summation. Theorem 6.1 now gives MΛ=C∑n∈Z∩ΛδnM_\Lambda= \mathbb{C}\sum_{n\in\mathbb{Z}\cap\Lambda}\delta_n. Here L=ZL = \mathbb{Z} attains (8) and FLF_L is empty.

Formalisation. The one-ray formalisation [13] proves that a finite lower half-line count on one carrier gives a nonzero exponential polynomial annihilating the measure. When the counted carrier is the actual support and the opposite count satisfies lim inf⁡R→∞NB(R)/R2=0\liminf_{R\to\infty} N_B(R)/R^2 = 0, it proves the comb conclusion on both sides, in either Fourier orientation. The common-carrier formalisation [14] proves the attained finite-hole formula, including the joint rank of the hole conditions, for bounded unit-interval counts. Theorems 1.1 and 1.2 here give the optimal width and the bound 2/d22/d^2, with the uniform estimate in Corollary 5.1. Theorem 6.1 gives (8) and the sharp bound ⌊d2⌋\lfloor d^2 \rfloor under the one-ray hypothesis.

References

  1. [1]L. Alon and M. Kummer, A characterization of uniformly discrete crystalline measures with unit weights, arXiv:2611.24286 (2026).arxiv.org/abs/2611.24286
  2. [2]M. Baake, N. Strungaru and V. Terauds, Pure point measures with sparse support and sparse Fourier–Bohr support, Trans. Lond. Math. Soc. 7 (2020), 1–32.
  3. [3]A. Eremenko and D. Novikov, Oscillation of Fourier integrals with a spectral gap, J. Math. Pures Appl. (9) 83 (2004), 313–365.
  4. [4]F. Gonçalves and G. Vedana, A complete classification of Fourier summation formulas on the real line, arXiv:2504.02741 (2025).arxiv.org/abs/2504.02741
  5. [5]P. Kurasov and P. Sarnak, Stable polynomials and crystalline measures, J. Math. Phys. 61 (2020), 083501.DOI
  6. [6]R. E. Langer, On the zeros of exponential sums and integrals, Bull. Amer. Math. Soc. 37 (1931), 213–239.DOI
  7. [7]C. Lech, A note on recurring series, Ark. Mat. 2 (1953), 417–421.DOI
  8. [8]N. Lev and A. Olevskii, Quasicrystals and Poisson’s summation formula, Invent. Math. 200 (2015), 585–606.DOI
  9. [9]N. Lev and A. Olevskii, Fourier quasicrystals and discreteness of the diffraction spectrum, Adv. Math. 315 (2017), 1–26.DOI
  10. [10]Y. Meyer, Measures with locally finite support and spectrum, Rev. Mat. Iberoam. 33 (2017), 1025–1036.
  11. [11]Y. Meyer, Crystalline measures and mean-periodic functions, Trans. R. Norw. Soc. Sci. Lett. (2021), no. 2.
  12. [12]A. Olevskii and A. Ulanovskii, Fourier quasicrystals with unit masses, C. R. Math. Acad. Sci. Paris 358 (2020), 1207–1211.arxiv.org/abs/2009.12810
  13. [13]Jamie Martin, Annihilation and finite-comb classification from finite lower density on one ray, Palomar Registry, PALOMAR-2026-11-20-000023, version 1 (2026). Registry record.
  14. [14]Jamie Martin, Finite-comb and finite-hole classification for relatively separated carriers, Palomar Registry, PALOMAR-2026-11-20-000019, version 1 (2026). Registry record.

Paper details

vlnc:W_4n9hbavg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
  • Mathematics

Anthropic Claude Fable 5.1, OpenAI Codex 6 Astra

View statement
Looking for related papers…
Contents