betaFeedback

WorksMathematics

Nonnegative weak tilings and the Coven-Meyerowitz condition (T2)

Abstract

We develop a general method, in finite cyclic groups, for constructing nonnegative rational functions ff, normalized by f(0)=1f(0)=1, that satisfy the Coven-Meyerowitz condition (T1)(T1) but fail (T2)(T2), such that ff is a weak complement for a set tile AA. Motivated by the unsupported-divisor questions in arXiv:2507.11672, we show that if a set tile AA admits an unsupported divisor, then a suitable dilation of AA admits such a rational weak complement ff. Using this method, we construct an explicit example in the group Z180\mathbb{Z}_{180}. The weak complements in our explicit perturbative constructions are not positive definite because of restrictions on the Fourier support of the perturbations. Thus these constructions do not resolve the proposed extension of the Coven-Meyerowitz conditions to positive-definite weak tilings (arXiv:2411.03854, Problem 1.2).

Introduction

By D. Newman’s periodicity theorem [15], tilings of the integers by a finite set can be studied through factorizations A⊕B=ZMA \oplus B = \mathbb{Z}_M, in which each residue has a unique representation as a+ba+b. For A⊂ZMA \subset\mathbb{Z}_M, choose representatives in {0,…,M−1}\{0,\ldots,M-1\} and write A(X)=∑a∈AXaA(X)=\sum_{a\in A}X^a for its mask polynomial. The tiling identity becomes

A(X)B(X)≡UM(X)(mod(XM−1)),UM(X)=1+X+⋯+XM−1.(1)A(X)B(X) \equiv U_M(X) \pmod{(X^M-1)}, \qquad U_M(X)=1+X+\cdots+X^{M-1}. \tag*{(1)}

E. Coven and A. Meyerowitz [1] introduced two conditions on the cyclotomic divisors of mask polynomials. Condition (T1)(T1) determines the value at X=1X=1 from the prime-power divisors; (T2)(T2) requires products of prime-power orders associated with distinct primes to occur as well. Together they imply tiling. Every integer tile satisfies (T1)(T1), and (T2)(T2) is also necessary when its cardinality has at most two distinct prime divisors (see [1 Theorems A, B1, and B2]). Precise definitions of these properties are given in § 2 of this work.

Remark 1.1. Recent preprints by J. Prajapati [16] and C. Li et al. [13] claim the unrestricted necessity of (T2)(T2) for integer tiles, which, if correct, would resolve the Coven–Meyerowitz conjecture for integer tilings. Our work does not use these proposed solutions or their methods.

Our starting point is the study of unsupported cyclotomic divisors in G. Kiss, I. Łaba, C. Marshall, and G. Somlai [5 Section 1.3]. They construct tiles satisfying (T1)(T1) and (T2)(T2) with unsupported divisors [5 Theorem 1.11]. We use such mixed zeros to construct rational weak complements failing (T2)(T2).

Date: October 20, 2026.

2020 Mathematics Subject Classification. Primary 05B45, 20K01, 11C08; Secondary 11B75, 42A99.

Key words and phrases. integer tilings, weak tilings, functional tilings, cyclotomic polynomials, Coven–Meyerowitz conditions.

Following [9, 12], a normalized nonnegative weak complement of A⊂ZMA \subset\mathbb{Z}_M, with 0∈A0 \in A, is a function satisfying

f:ZM⟶R≥0,f(0)=1,1A∗f=1ZM.(2)f:\mathbb{Z}_M \longrightarrow\mathbb{R}_{\ge0},\qquad f(0)=1,\qquad\mathbf{1}_A * f=\mathbf{1}_{\mathbb{Z}_M}. \tag*{(2)}

Equivalently, 1A∗(f−δ0)=1Ac\mathbf{1}_A * (f-\delta_0)=\mathbf{1}_{A^c}, so AA weakly tiles its complement. We keep the set factor equal to a tile and allow rational weights in its complement, leading to the following question.

Question 1.2. Suppose that A⊂ZMA \subset\mathbb{Z}_M is a tile satisfying (T1)(T1) and (T2)(T2). Must every normalized nonnegative rational weak complement of AA that satisfies (T1)(T1) also satisfy (T2)(T2)?

The answer is negative, as we shall see in the sequel. However, the explicit weak complements in our construction are not positive definite; the distinction from positive-definite (PD) weak tilings is discussed in § 7.

An explicit counterexample

We begin with an example of period 180. The polynomial appearing below is given explicitly in § 5.

Theorem 1.3. The set

A∗={0,12,18,24,36}⊂Z180A_*=\{0,12,18,24,36\}\subset\mathbb{Z}_{180}

tiles Z180\mathbb{Z}_{180} and satisfies ordinary (T1)(T1) and (T2)(T2). There is an ordinary polynomial F∈Q[X]F\in\mathbb{Q}[X] with nonnegative coefficients and F(0)=1F(0)=1 such that

A∗(X)F(X)≡U180(X)(mod(X180−1)),A_*(X)F(X)\equiv U_{180}(X)\pmod{(X^{180}-1)},

while FF satisfies ordinary (T1)(T1) but fails (T2)(T2). Its cyclic coefficient function wF180:Z180⟶Q≥0w_F^{180}:\mathbb{Z}_{180}\longrightarrow\mathbb{Q}_{\ge0}, which satisfies

F(X)≡∑y=0179wF180(y)Xy(mod(X180−1)),F(X)\equiv\sum_{y=0}^{179}w_F^{180}(y)X^y\pmod{(X^{180}-1)},

is not positive definite.

The set A∗A_* represents each residue class modulo five exactly once, so that the subgroup B∗=5Z180B_* = 5\mathbb{Z}_{180} is a tiling complement of cardinality 36. To obtain FF, we use a different complement, formed from translated layers, and verify the required cyclotomic divisibility properties directly in § 5. We do not claim that 180 is the least possible period.

The example comes from the five-point set E={0,2,3,4,6}E=\{0,2,3,4,6\}, since A∗=6EA_*=6E and E(X)=Φ5(X)Φ6(X)E(X)=\Phi_5(X)\Phi_6(X). This is an instance of the recombination construction in [5 Example 8.1], with primes 2, 3, 5. H. Tan and Y. Zhang [17] also use this set EE in their unsupported-divisor constructions.

A construction without new primes

A mixed order is an integer with at least two distinct prime divisors. If N=∏ipiαiN=\prod_i p_i^{\alpha_i}, its constituent prime-power orders are the piαip_i^{\alpha_i}. The following theorem gives a general construction from a mixed cyclotomic divisor whose constituent prime-power divisors are absent.

Theorem 1.4. Suppose that A⊕B=ZMA \oplus B = \mathbb{Z}_M and that N∣MN \mid M is a mixed order. Assume that ΦN∣A(X)\Phi_N \mid A(X), but that Φs∤A(X)\Phi_s \nmid A(X) for every constituent prime-power order ss of NN. Then there exist an integer q≥1q \ge1 and a polynomial P∈Q[X]P \in\mathbb{Q}[X] with nonnegative coefficients and P(0)=1P(0) = 1 such that

A(Xq)P(X)≡UqM(X)(mod(XqM−1)),A(X^q)P(X) \equiv U_{qM}(X) \pmod{(X^{qM}-1)},

and PP satisfies ordinary (T1)(T1) but fails (T2)(T2). The dilation qq can be chosen to be 11 or a power of a prime dividing NN.

A more thorough description of the polynomial PP and the dilated set tile qA⊂ZqMqA \subset\mathbb{Z}_{qM} with mask polynomial A(Xq)A(X^q) is given in Theorem 4.2.

The tiles constructed by G. Kiss, I. Łaba, C. Marshall, and G. Somlai [5 Theorem 10.1] satisfy the hypotheses of Theorem 1.4, so their dilations admit the weak complements constructed here. The set qA⊂ZqMqA \subset\mathbb{Z}_{qM} is again a set tile. Furthermore, qAqA satisfies ordinary (T2)(T2) if and only if AA does, as shown in Theorem 4.2; see also [1 Lemma 1.4]. Thus dilation preserves the cardinality, tiling property, and Coven–Meyerowitz conditions of the set factor.

For the main step of the construction, we choose q>1q > 1 and form a set complement CC of qAqA by placing translated copies of BB in the residue classes modulo qq. Its normalized autocorrelation, ∣C∣−11C∗1−C|C|^{-1}1_C * 1_{-C}, is a weak complement of qAqA. We then add a signed function whose Fourier support consists of a single mixed order. This function, which we call a pure-order cuboid filter, is defined in Section 3. The translations are chosen so that the autocorrelation has enough positive weight to absorb the negative coefficients of the filter, while the perturbation does not change the value at the origin. The proof is given in Section 4. The example in Theorem 1.3 uses the same idea with the composite dilation q=6q = 6 and fewer layers than the general bound would give.

Our use of a larger ambient group was inspired by the stacking construction of R. Greenfeld and T. Tao [3 Theorem 3.1]. As part of their disproof of the periodic tiling conjecture, they use additional layers to encode a system of tiling equations as a single equation in a larger group. Here the additional residue layers allow us to place translated complements so that their autocorrelation absorbs the signed perturbation while preserving normalization.

The weak complement construction can fail without dilations

One may ask whether the dilation in Theorem 1.4 can be avoided by allowing a larger ambient period while leaving the set tile AA fixed. The following result shows that this is not possible in general.

Theorem 1.5. Let

T={0,2,3,4,6}+30{0,1,2,3,4,5}⊂Z180.T = \{0,2,3,4,6\} + 30\{0,1,2,3,4,5\} \subset\mathbb{Z}_{180}.

This thirty-element tile satisfies ordinary (T1)(T1) and (T2)(T2) and has unsupported divisor Φ6\Phi_6.

For every L≥2L \ge2, suppose that P∈Q[X]P \in\mathbb{Q}[X] has nonnegative coefficients, satisfies P(0)=1P(0) = 1 and ordinary (T1)(T1), and obeys

T(X)P(X)≡UL(X)(mod(XL−1)).T(X)P(X) \equiv U_L(X) \pmod{(X^L-1)}.

Then PP satisfies ordinary (T2)(T2).

Here unsupported divisor includes the cardinality condition in Definition 2.6; Theorem 1.4 does not require that additional condition. Theorem 1.5 is proved as Proposition 6.2.

We also consider what dilation can accomplish when the complement is required to be a set. Proposition 6.3 shows that, for every tile A⊂ZMA \subset\mathbb{Z}_M and every q≥1q \ge1, the set qA⊂ZqMqA \subset\mathbb{Z}_{qM} has a set complement failing (T2)(T2) if and only if AA has one. Thus dilation cannot produce a new failure at the level of set tilings.

There is a further restriction on the ambient period. If MM has at most two distinct prime divisors and A⊂ZMA \subset\mathbb{Z}_M is a tile, then every rational polynomial weak complement of AA satisfying ordinary (T1)(T1) also satisfies ordinary (T2)(T2) (Proposition 6.4). In fact, this conclusion requires neither nonnegativity nor normalization of the complement.

Finally, we explain why the perturbation method does not give positive-definite examples. In the nontrivial constructions, the perturbation vanishes at the origin and its Fourier transform is supported on the Fourier zeros of the initial PD complement. Under these assumptions, Proposition 7.1 shows that every nonzero perturbation destroys positive definiteness. For the five-point set underlying Theorem 1.3, we prove a stronger statement: for every q≥1q \ge1, a normalized nonnegative rational PD complement of qE⊂Z30qqE \subset\mathbb{Z}_{30q} satisfying ordinary (T1)(T1) must also satisfy ordinary (T2)(T2) (Corollary 7.2).

Organization

Section 2 gives the Fourier and cyclotomic conventions. Section 3 develops the cuboid filter, and Sections 4 and 5 prove Theorems 1.4 and 1.3. Section 6 proves Theorem 1.5 and the additional boundary results. Section 7 explains the PD obstruction. Appendix A presents an alternative construction using an auxiliary prime coprime to the original period.

Mask polynomials, Fourier transforms, and tiling conditions

Cyclic coefficient functions and convolution

We will use both polynomials and their coefficient functions on finite cyclic groups, extending the multiset notation of [5 Section 2.1] to rational weights. To fix the correspondence, write [Xn]Q[X^n]Q for the coefficient of XnX^n in a polynomial QQ. For Q∈Q[X]Q \in\mathbb{Q}[X] and M≥2M \ge2, define its cyclic coefficient function by

wQM(x)=∑n≥0n≡x mod M[Xn]Q,x∈ZM.w_Q^M(x)=\sum_{\substack{n\ge0\\n\equiv x\ \mathrm{mod}\ M}}[X^n]Q,\qquad x\in\mathbb{Z}_M.

Thus coefficients whose exponents agree modulo the period are added together, and the canonical representative of QQ modulo XM−1X^M-1 is ∑x=0M−1wQM(x)Xx\sum_{x=0}^{M-1}w_Q^M(x)X^x. We call such a weight function a rational multiset on ZM\mathbb{Z}_M, and denote the collection by MQ(ZM)\mathcal{M}_{\mathbb{Q}}(\mathbb{Z}_M), using MQ+(ZM)\mathcal{M}_{\mathbb{Q}}^+(\mathbb{Z}_M) when all weights are nonnegative. Sets correspond to weights in {0,1}\{0,1\}.

Conversely, for a function f:ZM→Qf:\mathbb{Z}_M\to\mathbb{Q}, its canonical mask polynomial F∈Q[X]F\in\mathbb{Q}[X] is given by the formula

F(X)=∑x=0M−1f(x)Xx.F(X)=\sum_{x=0}^{M-1}f(x)X^x.

Thus F(0)=f(0)F(0)=f(0) and F(1)=∑xf(x)F(1)=\sum_x f(x); polynomial evaluation at one is distinct from the value f(1)f(1) of the coefficient function.

For functions f,g:ZM→Cf,g:\mathbb{Z}_M\to\mathbb{C}, write

(f∗g)(x)=∑y∈ZMf(y)g(x−y),mass⁡(f)=∑xf(x),f~(x)=f(−x)‾.(f*g)(x)=\sum_{y\in\mathbb{Z}_M}f(y)g(x-y),\qquad\operatorname{mass}(f)=\sum_x f(x),\qquad\widetilde{f}(x)=\overline{f(-x)}.

Here the overline denotes complex conjugation. With these conventions, multiplication of mask polynomials corresponds to convolution over the corresponding cyclic group, while reduction modulo XM−1X^M-1 preserves mass: Q(1)=mass⁡(wQM)Q(1)=\operatorname{mass}(w_Q^M).

Lemma 2.1. If A⊂ZMA \subset\mathbb{Z}_M is nonempty and b:ZM→Z≥0b:\mathbb{Z}_M \to\mathbb{Z}_{\ge0} satisfies 1A∗b=1ZM\mathbf{1}_A*b=\mathbf{1}_{\mathbb{Z}_M}, then bb is the indicator function of a set.

Proof. Fix a0∈Aa_0 \in A. Then, for every xx, one must have

1=(1A∗b)(x+a0):=∑a∈Ab(x+a0−a)≥b(x)≥0.1=(\mathbf{1}_A*b)(x+a_0):=\sum_{a\in A}b(x+a_0-a)\ge b(x)\ge0.

Hence, b(x)∈{0,1}b(x)\in\{0,1\} for every x∈ZMx\in\mathbb{Z}_M, as claimed. □

The lemma explains why passing from sets to nonnegative integer weights does not enlarge the class of complements at tiling level one. For a rational weak complement, however, clearing a denominator produces a multiple tiling: if qfqf is integer-valued, then 1A∗(qf)=q1ZM\mathbf{1}_A*(qf)=q\mathbf{1}_{\mathbb{Z}_M}.

Fourier and cyclotomic criteria

The tiling congruence can be tested at roots of unity, or equivalently on the Fourier side. We use the unnormalized Fourier transform:

f^(k)=∑x=0M−1f(x)e−2πikx/M,f(x)=1M∑k=0M−1f^(k)e2πikx/M.\widehat{f}(k)=\sum_{x=0}^{M-1}f(x)e^{-2\pi ikx/M},\qquad f(x)=\frac{1}{M}\sum_{k=0}^{M-1}\widehat{f}(k)e^{2\pi ikx/M}.

The character indexed by kk has order M/gcd⁡(k,M)M/\gcd(k,M), with the trivial character having order one. Character orthogonality gives the inversion formula and the identity f∗g^=f^g^\widehat{f*g}=\widehat{f}\widehat{g}. To express Fourier vanishing in terms of polynomial divisibility, recall that for d≥1d\ge1 the cyclotomic polynomial is

Φd(X)=∏1≤u≤d(u,d)=1(X−e2πiu/d).\Phi_d(X)=\prod_{\substack{1\le u\le d\\(u,d)=1}}\left(X-e^{2\pi iu/d}\right).

It is monic and irreducible over Q\mathbb{Q}, and XM−1=∏d∣MΦd(X)X^M-1=\prod_{d\mid M}\Phi_d(X).

Proposition 2.2 (cf. E. Coven and A. Meyerowitz, [1 Lemma 1.3]). For A⊂ZMA\subset\mathbb{Z}_M and a rational-valued function ff with canonical mask polynomial FF, the following are equivalent:

(i) 1A∗f=1ZM\mathbf{1}_A*f=\mathbf{1}_{\mathbb{Z}_M};

(ii) A(X)F(X)≡UM(X)(mod(XM−1))A(X)F(X)\equiv U_M(X)\pmod{(X^M-1)};

(iii) ∣A∣F(1)=M|A|F(1)=M and, for every d∣Md\mid M with d>1d>1, at least one of A(X)A(X) and F(X)F(X) is divisible by Φd(X)\Phi_d(X).

Proof. Comparing cyclic coefficients gives the equivalence of (i) and (ii). On the Fourier side, the constant function has value MM at the trivial character and zero elsewhere. The convolution identity is therefore equivalent to the mass identity together with 1A^(k)f^(k)=0\widehat{\mathbf{1}_A}(k)\widehat{f}(k)=0 at every nontrivial character. Since the mask polynomials have rational coefficients, vanishing at one primitive ddth root is equivalent to divisibility by Φd\Phi_d. This proves (iii). □

The two conventions for (T1) and (T2)

For rational mask polynomials it is useful to distinguish the conditions on an ordinary polynomial from the conditions seen at a fixed cyclic period. The latter involve only orders dividing that period.

Definition 2.3 (E. Coven and A. Meyerowitz [1]). For a nonzero polynomial Q∈Q[X]Q \in\mathbb{Q}[X], let

SQ∗={pa:p prime, a≥1, Φpa∣Q}.S_Q^*=\{p^a : p\ \text{prime},\ a\ge1,\ \Phi_{p^a}\mid Q\}.

The ordinary Coven–Meyerowitz conditions are:

(T1) Q(1)=∏s∈SQ∗Φs(1)Q(1)=\prod_{s\in S_Q^*}\Phi_s(1);

(T2) whenever k≥1k\ge1 and s1,…,sk∈SQ∗s_1,\ldots,s_k\in S_Q^* are powers of distinct primes, Φs1⋯sk∣Q\Phi_{s_1\cdots s_k}\mid Q.

For a function on ZM\mathbb{Z}_M, we call the versions restricted to the ambient period the period-relative conditions. They use SQ,M={s∈SQ∗:s∣M}S_{Q,M}=\{s\in S_Q^*:s\mid M\} in the same formulas instead of SQ∗S_Q^*. These depend only on the cyclic coefficient function, not on its ordinary polynomial representative.

Recall that Φpa(1)=p\Phi_{p^a}(1)=p. We write pn∥Mp^n\mathrel{\|}M when pn∣Mp^n\mid M and pn+1∤Mp^{n+1}\nmid M. For a prime pp and a positive integer mm, the pp-adic valuation vp(m)v_p(m) is the unique integer r≥0r\ge0 such that pr∥mp^r\mathrel{\|}m. The following explains how prime-power divisors pa∣Mp^a\mid M of the tiling period MM are distributed among the mask polynomials of the sets.

Lemma 2.4. If A⊕B=ZMA\oplus B=\mathbb{Z}_M, then every prime-power cyclotomic divisor of either mask polynomial has order dividing MM. For each prime power pa∣Mp^a\mid M, a≥1a\ge1, the factor Φpa\Phi_{p^a} divides exactly one mask polynomial. Both mask polynomials satisfy ordinary (T1).

Proof. The counting argument rests on integrality. For a nonzero Q∈Z[X]Q\in\mathbb{Z}[X] with Q(1)>0Q(1)>0, monicity and Gauss’s lemma imply that the product of any distinct cyclotomic divisors divides QQ in Z[X]\mathbb{Z}[X]. Evaluating at one shows that the number of prime-power factors based at pp is at most vp(Q(1))v_p(Q(1)).

Now let pn∥Mp^n\mathrel{\|}M. Proposition 2.2 requires each of the nn factors Φp,…,Φpn\Phi_p,\ldots,\Phi_{p^n} to divide at least one of the two masks. On the other hand, their cardinalities allow only vp(∣A∣∣B∣)=nv_p(|A||B|)=n such factors in total. Thus every required factor occurs exactly once, and there is no room for any further pp-power factor. A prime not dividing MM cannot occur in either mask polynomial for the same reason. Evaluating the assigned factors at one gives (T1). □

Remark 2.5. The restriction to integer mask polynomials is essential to this counting argument, as a small example shows. The tile {0,3}⊂Z6\{0,3\}\subset\mathbb{Z}_6 has the normalized weak complement f=(1,3/4,1/4,0,1/4,3/4)f=(1,3/4,1/4,0,1/4,3/4). Its mass is 3, but it vanishes at orders 2 and 3, whose cyclotomic values at one multiply to 6. Hence ff fails (T1), and its nonvanishing at order 6 also gives a failure of (T2). Without the hypothesis (T1) on ff, Question 1.2 would therefore have a trivial negative answer. The constructions below will retain this hypothesis.

We use the definition of an unsupported divisor from [5 Definition 1.5].

Definition 2.6. Let AA be a finite nonempty set and let N=∏i=1kpiαiN=\prod_{i=1}^k p_i^{\alpha_i}, with k≥2k\ge2 and distinct primes pip_i. Then ΦN\Phi_N is an unsupported divisor of AA if ΦN∣A\Phi_N\mid A, every pip_i divides ∣A∣|A|, and Φpiαi∤A\Phi_{p_i^{\alpha_i}}\nmid A for every ii. The cardinality condition is part of this definition. For the construction below, it will be enough to have the mixed zero and the absence of its constituent prime-power zeros.

Normalized autocorrelations and positive definiteness

The construction will start from the normalized autocorrelation of a set complement. We recall why this gives a weak complement and how its Fourier transform relates to that of the set.

Definition 2.7. A real-valued function f:ZM→Rf:\mathbb{Z}_M\to\mathbb{R} is positive definite (PD) if f^(k)≥0\widehat{f}(k)\geq0 for every k∈ZMk\in\mathbb{Z}_M. A PD weak complement satisfies this additional condition as well as (1.2). This is the finite-group notion of weak PD tiling in [8 Definition 1.1].

Lemma 2.8 (G. Kiss, D. Matolcsi, M. Matolcsi, and G. Somlai, [8 Lemma 1.2]). If A⊕B=ZMA\oplus B=\mathbb{Z}_M, the autocorrelation function

gB=1∣B∣1B∗1−Bg_B=\frac{1}{|B|}\mathbf{1}_B * \mathbf{1}_{-B}

is a normalized PD weak complement of AA with mass ∣B∣|B|. Moreover,

1B^(k)=0⟺gB^(k)=0,k∈ZM.\widehat{\mathbf{1}_B}(k)=0\Longleftrightarrow\widehat{g_B}(k)=0,\qquad k\in\mathbb{Z}_M.

As the proof is direct and short, we include it for completeness.

Proof. Each coefficient is a difference count divided by the normalizing factor, so it is nonnegative. Exactly ∣B∣|B| ordered pairs of points of BB have difference zero, giving the normalization. To check the weak-tiling identity, associativity gives 1A∗gB=∣B∣−11ZM∗1−B=1ZM\mathbf{1}_A*g_B=|B|^{-1}\mathbf{1}_{\mathbb{Z}_M}*\mathbf{1}_{-B}=\mathbf{1}_{\mathbb{Z}_M}. Finally, the identity gB^=∣1B^∣2/∣B∣\widehat{g_B}=|\widehat{\mathbf{1}_B}|^2/|B| proves positive definiteness, preserves all Fourier zeros, and gives the mass. □

A real PD function is even: its Fourier transform is real, so Fourier inversion and conjugation give f(−x)=f(x)f(-x)=f(x). Later we will choose perturbations that are not invariant under negation, and this observation will show that the resulting weak complements are not PD.

A pure-order cuboid filter

We use the cuboid formalism introduced in [11 Section 5] and reviewed in [5 Section 2.3]. The construction below develops the cuboid perturbation proposed in [5 Section 10.2]: we construct a signed mask that changes the Fourier values at a chosen order while leaving every other Fourier value at the ambient period unchanged. Let 1<N∣M1<N\mid M and write

N=∏i=1kpiαi,αi≥1,N=\prod_{i=1}^{k}p_i^{\alpha_i},\qquad\alpha_i\geq1,

where the pip_i are distinct primes. Define the signed cuboid on ZN\mathbb{Z}_N by its canonical mask

ΔN(X)=∏i=1k(1−XN/pi)(mod(XN−1)),deg⁡ΔN<N.\Delta_N(X)=\prod_{i=1}^{k}(1-X^{N/p_i})\pmod{(X^N-1)},\qquad\deg\Delta_N<N.

Its 2k2^k vertices carry alternating signs. The pure-order cuboid filter on ZM\mathbb{Z}_M is the periodic extension

RN,M(X)=ΔN(X)∑j=0M/N−1XNj.(3)R_{N,M}(X)=\Delta_N(X)\sum_{j=0}^{M/N-1}X^{Nj}. \tag*{(3)}

This polynomial has degree less than MM; its coefficient function is RN,M(x)=ΔN(x mod N)R_{N,M}(x)=\Delta_N(x\bmod N).

Lemma 3.1. The filter R=RN,MR = R_{N,M} has the following properties:

(1) R(x)∈{−1,0,1}R(x) \in\{-1, 0, 1\}, R(0)=1R(0) = 1, and mass⁡(R)=0\operatorname{mass}(R) = 0;

(2) it has exactly 2k−1M/N2^{k-1}M/N positive and the same number of negative coefficients;

(3) its Fourier support consists exactly of the characters of order NN;

(4) if ΦN∣A(X)\Phi_N \mid A(X), then 1A∗R=0\mathbf{1}_A * R = 0.

Proof. To verify the coefficient claims, first consider the subset sums ∑iϵiN/pi\sum_i \epsilon_iN/p_i, ϵi∈{0,1}\epsilon_i \in\{0,1\}. These are distinct modulo NN: reducing an equality modulo piαip_i^{\alpha_i} eliminates all terms except the iith, and the pip_i-adic valuation of N/piN/p_i is αi−1\alpha_i - 1, so the equality determines ϵi\epsilon_i. It follows that the canonical mask polynomial ΔN\Delta_N has 2k2^k nonzero coefficients, each equal to 11 or −1-1, with equally many of each sign and with coefficient 11 at the origin.

Repeating the cuboid does not change these coefficients. Indeed, the shifted copies XNjΔN(X)X^{Nj}\Delta_N(X) in (3) occupy the disjoint exponent intervals [Nj,N(j+1)−1][Nj, N(j+1)-1], so their supports do not overlap. Each coefficient therefore occurs in a separate block, proving (1) and (2).

For the Fourier support, let ζ\zeta have order d∣Md \mid M. The geometric sum ∑jζNj\sum_j \zeta^{Nj} vanishes unless d∣Nd \mid N, in which case it equals M/NM/N. If d∣Nd \mid N is proper, some ii satisfies d∣N/pid \mid N/p_i, so the factor 1−ζN/pi1-\zeta^{N/p_i} vanishes. At order NN, none of these factors vanishes. This proves (3), including vanishing at the trivial character. If ΦN∣A\Phi_N \mid A, the Fourier transform of 1A\mathbf{1}_A vanishes on the support of R^\widehat{R}, so the convolution theorem gives (4). □

Corollary 3.2. Let Q∈Q[X]Q \in\mathbb{Q}[X] and t∈Q∖{0}t \in\mathbb{Q} \setminus\{0\}. Adding tRN,MtR_{N,M} leaves the values of QQ unchanged at every MMth root whose order differs from NN. If ΦN∣Q\Phi_N \mid Q, then ΦN∤Q+tRN,M\Phi_N \nmid Q + tR_{N,M}. When NN is mixed, this change preserves every prime-power Fourier zero and nonzero at the ambient period MM.

Proof. By Lemma 3.1(3), the filter vanishes at every MMth root except those of order NN, where it is nonzero. The two assertions follow by evaluating the perturbed polynomial at these roots. □

This gives a way to remove a mixed cyclotomic zero while retaining the data entering the period-relative condition (T1)(T1). To use it for weak complements, we must also preserve nonnegativity and normalization. Thus the negative filter sites need enough positive weight, and the perturbation must leave the origin unchanged. We arrange these two requirements by placing translated complements in separate layers.

A cyclic construction without new primes

We now use the filter to construct a normalized nonnegative weak complement satisfying (T1)(T1) and failing (T2)(T2), starting from a mixed zero of a set tile whose constituent prime-power zeros are absent. The construction allows a dilation by a power of a prime already dividing the mixed order, so no new prime is introduced into the period or the coefficient denominators.

The use of the residue layers j+qZMj+q\mathbb{Z}_M of ZqM\mathbb{Z}_{qM} is what permits qq and MM to share prime factors. For comparison, the construction in Appendix A uses a direct product ZM×Zr\mathbb{Z}_M\times\mathbb{Z}_r, with a new prime rr coprime to MM.

Removing off-period factors using an old prime

Before constructing the weak complement, we show how to remove off-period cyclotomic factors without changing the cyclic coefficient function. We begin with the following coefficient reweighting result over Q[X]\mathbb{Q}[X].

Lemma 4.1. Let F∈Q[X]F\in\mathbb{Q}[X] have nonnegative coefficients and degree less than DD. Suppose that mF∈Z[X]mF\in\mathbb{Z}[X] for a positive integer mm, and choose an integer xx with 0<x<D0<x<D and [Xx]F>0[X^x]F>0. For any odd prime ℓ\ell, the polynomial

P(X)=F(X)+Xx+D−XxℓmP(X)=F(X)+\frac{X^{x+D}-X^x}{\ell m}

has nonnegative coefficients, satisfies P(0)=F(0)P(0)=F(0) and P(1)=F(1)P(1)=F(1), and represents the same cyclic coefficient function on ZD\mathbb{Z}_D. Every cyclotomic divisor of PP has order dividing DD.

Proof. The only coefficient being decreased is the one at xx. Since [Xx](mF)[X^x](mF) is a positive integer, we have [Xx]F≥1/m[X^x]F\geq1/m, so subtracting 1/(ℓm)1/(\ell m) preserves nonnegativity. The new coefficient at x+Dx+D is positive, and the two exponents agree modulo DD. Consequently the cyclic function is unchanged. As xx is nonzero, the transfer also preserves the constant coefficient, and its two terms cancel at one.

To locate the cyclotomic divisors, we work modulo the chosen odd prime. The reduction identities used below also appear in T. Fallon, G. Kiss, A. Mayeli, and G. Somlai [2 Lemma 2.3] and R. Malikiosis [14 proof of Theorem 6.1]; we include the details to keep track of root multiplicities. Assume that Φn∣P\Phi_n\mid P. Since Φn\Phi_n is a monic polynomial, Gauss’s lemma implies that Φn\Phi_n divides the integer polynomial

ℓmP(X)=ℓmF(X)+Xx(XD−1)\ell mP(X)=\ell mF(X)+X^x(X^D-1)

in Z[X]\mathbb{Z}[X]. For an integer polynomial QQ, write Q‾\overline{Q} for its coefficientwise reduction modulo ℓ\ell. Reducing the divisibility gives

Φn‾(X)∣Xx(XD−1)in Fℓ[X].\overline{\Phi_n}(X)\mid X^x(X^D-1)\quad\text{in }\mathbb{F}_\ell[X].

Since Φn‾(0)≠0\overline{\Phi_n}(0)\ne0, the polynomials XxX^x and Φn‾\overline{\Phi_n} are coprime, so Φn‾∣XD−1\overline{\Phi_n}\mid X^D-1 in Fℓ[X]\mathbb{F}_\ell[X]. Write

D=ℓbD0,n=ℓan0,(D0,ℓ)=(n0,ℓ)=1.D=\ell^bD_0,\qquad n=\ell^a n_0,\qquad(D_0,\ell)=(n_0,\ell)=1.

This separates the order of a root from its multiplicity in the chosen characteristic. Indeed, in characteristic ℓ\ell we have XD−1=(XD0−1)ℓbX^D-1=(X^{D_0}-1)^{\ell^b}, whose nonzero roots are the D0D_0th roots of unity, each with multiplicity ℓb\ell^b.

Since ℓ∤n0\ell\nmid n_0, the polynomial Xn0−1X^{n_0}-1 is separable (i.e., it has n0n_0 distinct roots in an algebraic closure). Its factorization into reduced cyclotomic polynomials shows that Φn0‾\overline{\Phi_{n_0}} has precisely the primitive n0n_0th roots, each with multiplicity one. For a≥1a\geq1, the cyclotomic identity

Φℓan0(X)Φn0(Xℓa−1)=Φn0(Xℓa)\Phi_{\ell^a n_0}(X)\Phi_{n_0}(X^{\ell^{a-1}})=\Phi_{n_0}(X^{\ell^a})

and the Frobenius identity Q(Xℓj)=Q(X)ℓjQ(X^{\ell^j})=Q(X)^{\ell^j} for Q∈Fℓ[X]Q\in\mathbb{F}_\ell[X] give

Φℓan0‾  Φn0‾ℓa−1=Φn0‾ℓa.\overline{\Phi_{\ell^a n_0}}\;\overline{\Phi_{n_0}}^{\ell^{a-1}}=\overline{\Phi_{n_0}}^{\ell^a}.

Cancellation in Fℓ[X]\mathbb{F}_{\ell}[X] therefore yields

Φℓan0‾=Φn0‾ℓa−1(ℓ−1)\overline{\Phi_{\ell^{a}n_0}}=\overline{\Phi_{n_0}}^{\ell^{a-1}(\ell-1)}

The reduced cyclotomic polynomials also have nonzero roots. Comparing their orders with the roots allowed by the divisibility above forces n0∣D0n_0 \mid D_0. If a≥1a \ge1, comparing their multiplicities gives

ℓa−1(ℓ−1)≤ℓb\ell^{a-1}(\ell-1) \le\ell^b

Because ℓ\ell is odd, a≥b+1a \ge b+1 would make the left-hand side strictly larger than ℓb\ell^b. Hence a≤ba \le b, so n∣Dn \mid D. □

The oddness of ℓ\ell is needed only for the last multiplicity comparison.

The prime-preserving construction

We now give our main result. Given a set tile AA of period MM that admits an unsupported divisor ΦN\Phi_N with N∣MN \mid M, one can construct a polynomial P∈Q[X]P \in\mathbb{Q}[X] with nonnegative coefficients such that

A(Xq)P(X)≡UqM(X)(modXqM−1)A(X^q)P(X) \equiv U_{qM}(X) \pmod{X^{qM}-1}

where q=1q=1 or q=phq=p^h for a prime p∣Np \mid N and an integer h≥1h \ge1. The polynomial PP satisfies the Coven–Meyerowitz condition (T1)(T1) but fails (T2)(T2).

Theorem 4.2. Suppose that A⊕B=ZMA \oplus B = \mathbb{Z}_M and that

N=∏i=1kpiαi∣M,k≥2,ΦN∣A,Φpiαi∤A(1≤i≤k).N=\prod_{i=1}^{k}p_i^{\alpha_i}\mid M,\qquad k\ge2,\qquad\Phi_N\mid A,\qquad\Phi_{p_i^{\alpha_i}}\nmid A\quad(1\le i\le k).

Then there are an integer q≥1q \ge1, which is either 1 or a power of a prime dividing NN, and a polynomial P∈Q[X]P \in\mathbb{Q}[X] such that the following hold.

(1) The coefficients of PP are nonnegative, P(0)=1P(0)=1, and

A(Xq)P(X)≡UqM(X)(modXqM−1).A(X^q)P(X)\equiv U_{qM}(X)\pmod{X^{qM}-1}.

(2) The polynomial PP satisfies the Coven–Meyerowitz condition (T1)(T1) but fails (T2)(T2).

(3) If the coefficients [Xx]P[X^x]P are written in lowest terms as nx/dx∈Qn_x/d_x\in\mathbb{Q}, with dx≥1d_x\ge1, then every prime divisor of dxd_x divides MM.

(4) The dilated set qA⊂ZqMqA\subset\mathbb{Z}_{qM} is a set tile, has cardinality ∣qA∣=∣A∣|qA|=|A|, and satisfies ordinary (T2)(T2) if and only if AA does.

Proof. Since A⊕B=ZMA\oplus B=\mathbb{Z}_M, by Lemma 2.4, every Φpiαi\Phi_{p_i^{\alpha_i}} divides BB. We prove (1)–(3) by considering whether the complement vanishes at the same mixed order as the tile.

Case 1: ΦN∤B\Phi_N\nmid B. Here the complement already has the required prime-power zeros and lacks their mixed zero. We need only arrange that its indicator is normalized. Choose b0∈Bb_0\in B and replace BB by the translate B′=(B−b0) mod MB'=(B-b_0)\bmod M, so that 0∈B′0\in B' and A⊕B′=ZMA\oplus B'=\mathbb{Z}_M. For every d∣Md\mid M, translation multiplies the mask polynomial value at a primitive ddth root by a nonzero monomial value. It therefore preserves vanishing at every such order. Both set complements have all their ordinary prime-power factors at orders dividing MM, by Lemma 2.4; hence their ordinary prime-power signatures agree. Every mixed order required by (T2)(T2) also divides MM, so translation preserves (T2)(T2) and its failure.

Set q=1q=1 and let PP be the canonical mask polynomial of B′B'. Then PP has nonnegative integer coefficients, P(0)=1P(0)=1, and A(X)P(X)≡UM(X)(modXM−1)A(X)P(X) \equiv U_M(X) \pmod{X^M-1}. It satisfies ordinary (T1)(T1) by allocation. All the constituent prime-power factors of NN divide PP, whereas ΦN∤P\Phi_N \nmid P, so (T2)(T2) fails at NN.

Case 2: ΦN∣B\Phi_N \mid B. In this case both mask polynomials vanish at the mixed order. We will build a layered complement whose autocorrelation has enough positive weight at every negative filter site. Let R=RN,MR=R_{N,M} be the filter of Lemma 3.1; its negative support has cardinality

s=2k−1M/N.s=2^{k-1}M/N.

List these positions as x1,…,xs∈{0,…,M−1}x_1,\ldots,x_s \in\{0,\ldots,M-1\}. We allow a consecutive pair of layers for each of these positions by choosing a prime p∣Np \mid N and a power q=phq=p^h such that q≥max⁡(3,s+1)q \ge\max(3,s+1). Put D=qMD=qM and K=q∣B∣K=q|B|. Choose shifts in ZM\mathbb{Z}_M by

t0=0,tj−tj−1=xj(1≤j≤s),tj−tj−1=0(s<j<q).t_0=0,\qquad t_j-t_{j-1}=x_j\quad(1\leq j\leq s),\qquad t_j-t_{j-1}=0\quad(s<j<q).

The layered set

C=⋃j=0q−1{j+q(b+tj):b∈B}⊂ZD(4)C=\bigcup_{j=0}^{q-1}\{j+q(b+t_j):b\in B\}\subset\mathbb{Z}_D \tag*{(4)}

has KK elements and complements qAqA. Indeed, its layers occupy distinct residue classes modulo qq. Subtracting jj and dividing by qq identifies the tiling in layer jj with A⊕(B+tj)=ZMA\oplus(B+t_j)=\mathbb{Z}_M. Equivalently,

C(X)≡B(Xq)∑j=0q−1Xj+qtj(modXD−1).C(X)\equiv B(X^q)\sum_{j=0}^{q-1}X^{j+qt_j}\pmod{X^D-1}.

Multiplication by A(Xq)A(X^q) gives UM(Xq)Uq(X)=UD(X)U_M(X^q)U_q(X)=U_D(X), because UM(Xq)U_M(X^q) is invariant under multiplication by XqtjX^{qt_j} in the quotient ring Q[X]/(XD−1)\mathbb{Q}[X]/(X^D-1).

To keep the origin fixed, we place the filter in a nonzero residue layer of the autocorrelation. Define functions on ZD\mathbb{Z}_D by

g=1K1C∗1−C,H(X)=XR(Xq),f=g+1KH.(5)g=\frac{1}{K}\mathbf{1}_C * \mathbf{1}_{-C},\qquad H(X)=XR(X^q),\qquad f=g+\frac{1}{K}H. \tag*{(5)}

The polynomial HH has degree less than DD, and its coefficient function is supported in the residue class 11 mod qq. Explicitly, H(1+qx)=R(x)H(1+qx)=R(x) for 0≤x≤M−10\leq x\leq M-1. Since 1A∗R=0\mathbf{1}_A*R=0, substitution gives

A(Xq)H(X)=XA(Xq)R(Xq)=X(AR)(Xq)≡0(modXD−1).A(X^q)H(X)=XA(X^q)R(X^q)=X(AR)(X^q)\equiv0\pmod{X^D-1}.

Also H(0)=0H(0)=0 and mass⁡(H)=0\operatorname{mass}(H)=0.

We now address the nonnegativity of the weight function ff. For each negative position xjx_j and every b∈Bb\in B, the two points

j+q(b+tj),(j−1)+q(b+tj−1)j+q(b+t_j),\qquad(j-1)+q(b+t_{j-1})

of CC have difference 1+qxj1+qx_j modulo DD. These are ∣B∣|B| distinct ordered pairs, so

g(1+qxj)≥∣B∣K,f(1+qxj)≥∣B∣−1K≥0.g(1+qx_j)\geq\frac{|B|}{K},\qquad f(1+qx_j)\geq\frac{|B|-1}{K}\geq0.

These are all the negative sites of HH, and consequently f≥0f \ge0. Lemma 2.8 gives g(0)=1g(0)=1, 1qA∗g=1ZD\mathbf{1}_{qA} * g=\mathbf{1}_{\mathbb{Z}_D}, and mass⁡(g)=K\operatorname{mass}(g)=K. Since H(0)=mass⁡(H)=0H(0)=\operatorname{mass}(H)=0 and 1qA∗H=0\mathbf{1}_{qA} * H=0, the same identities hold for ff:

f(0)=1,1qA∗f=1ZD,mass⁡(f)=K.f(0)=1,\qquad\mathbf{1}_{qA} * f=\mathbf{1}_{\mathbb{Z}_D},\qquad\operatorname{mass}(f)=K.

We next check that the perturbation preserves the prime-power zeros while removing the mixed zero. For a root of unity ζ\zeta of order dividing DD, the value H(ζ)H(\zeta) is nonzero exactly when ζq\zeta^q has order NN. If ζ\zeta has order dd, then ζq\zeta^q has order d/gcd⁡(d,q)d/\gcd(d,q). Since qq is a power of a prime dividing NN, this order equals NN exactly when d=qNd=qN. Thus HH has pure Fourier order qNqN. As qNqN has at least two distinct prime factors, HH vanishes at every prime-power character. The prime-power Fourier zeros of ff are therefore exactly those of gg, hence of CC. Lemma 2.4 and ∣C∣=K|C|=K prove period-relative (T1)(T1).

At a primitive qNqNth root ζ\zeta, the layered factorization of CC gives C(ζ)=0C(\zeta)=0, since B(ζq)=0B(\zeta^q)=0. The autocorrelation gg therefore also vanishes at this character. In contrast, H(ζ)=ζR(ζq)≠0H(\zeta)=\zeta R(\zeta^q)\ne0, so ff does not vanish at order qNqN. To obtain the (T2)(T2) failure, we also need its constituent prime-power zeros. The constituent prime powers of qNqN are piαi+vpi(q)p_i^{\alpha_i+v_{p_i}(q)}. Evaluating A(Xq)A(X^q) at each such order reduces to a nonzero evaluation of AA at order piαip_i^{\alpha_i}. Allocation of prime-power divisors puts all these prime-power factors in CC, and hence in ff. This then proves the (T2)(T2) failure.

It remains to pass from the cyclic function to an ordinary polynomial with the required conditions. Let FF be the canonical mask polynomial of ff. We have KF∈Z[X]KF\in\mathbb{Z}[X] and [X]F≥1/K[X]F\ge1/K, because [X]H=[X0]R=1[X]H=[X^0]R=1. Since k≥2k\ge2, some odd prime ℓ\ell divides NN. Lemma 4.1, with m=Km=K and x=1x=1, gives the ordinary representative

P(X)=F(X)+XD+1−XℓK.(6)P(X)=F(X)+\frac{X^{D+1}-X}{\ell K}. \tag*{(6)}

By Lemma 4.1, PP has nonnegative coefficients, satisfies P(0)=1P(0)=1, and is congruent to FF modulo XD−1X^D-1, proving (1). It has no off-period cyclotomic divisors. Thus period-relative (T1)(T1) becomes ordinary (T1)(T1), while the cyclic normalization and (T2)(T2) failure remain unchanged. This proves (2) in Case 2. Every prime in the denominator ℓK=ℓq∣B∣\ell K=\ell q|B| divides the original period MM, verifying (3).

To prove (4), observe that for every q≥1q\ge1, evaluation at prime-power roots gives

SA(Xq)∗={ℓa+vℓ(q):ℓa∈SA∗}.(7)S^*_{A(X^q)}=\{\ell^{a+v_\ell(q)}:\ell^a\in S_A^*\}. \tag*{(7)}

At order ℓb\ell^b with b≤vℓ(q)b\le v_\ell(q), the value is ∣A∣|A|, not zero; at larger orders, taking the qqth power reduces the root order by the factor ℓvℓ(q)\ell^{v_\ell(q)}. For products of powers of distinct primes, the same calculation applies to all constituent exponents simultaneously. The mixed orders required by (T2)(T2) for A(Xq)A(X^q) therefore correspond exactly to those required for AA, and evaluation at their primitive roots proves (T2)(T2) in both directions. The set qAqA has cardinality ∣A∣|A| and is a tile: the complement is B′B' in Case 1 and CC in Case 2. □\square

Remark 4.3. In any nontrivial-dilation branch where q=phq=p^h, gg is even whereas the cyclic coefficients of HH at 11 and −1-1 are respectively 11 and 00, since q>2q>2. Thus f(1)−f(−1)=1/Kf(1)-f(-1)=1/K, and ff is not positive definite.

The next remark explains how a set complement failing (T2)(T2) would yield a PD weak complement with the same failure. If the set factor also satisfies (T2)(T2), this would give a counterexample to the positive-definite analogue of Question 1.2, and hence to the extension proposed in [6 Problem 1.2]. No such set complement is constructed here.

Remark 4.4. Case 1 begins with a set complement BB that already fails (T2)(T2). If such a complement exists, its normalized autocorrelation is a PD weak complement with the same Fourier zeros and the same failure. Its mass is ∣B∣>1|B| > 1 and its value at zero is 1, so it has a positive coefficient away from zero. Applying Lemma 4.1 with m=∣B∣m = |B| gives an ordinary (T1)(T1) representative without changing that cyclic PD function. Thus Case 1 yields a PD weak complement failing (T2)(T2) only if one already has a set complement failing (T2)(T2). If AA also satisfies (T2)(T2), this would answer the positive-definite analogue of Question 1.2 negatively.

A three-prime counterexample of period 180

We now prove Theorem 1.3 by giving the weak complement explicitly. We start from the five-point set {0,2,3,4,6}\{0, 2, 3, 4, 6\}, whose mask polynomial is Φ5Φ6\Phi_{5}\Phi_{6}, and follow the construction of the preceding section with mixed order N=6N = 6. In this example, six translated layers suffice to keep the perturbed complement nonnegative. This gives the composite dilation q=6q = 6, although Theorem 4.2 chooses a prime power and uses the larger bound q≥s+1=11q \ge s + 1 = 11. The resulting period is 180, with only three distinct prime divisors. The set realizes [5 Example 8.1] with primes 2, 3, 5, as noted in the introduction; we verify below all of its properties that we use.

For a positive integer ss, write Us(Y)=1+Y+⋯+Ys−1U_s(Y) = 1 + Y + \cdots+ Y^{s-1}. Define

A∗={0,12,18,24,36}⊂Z180,T={0,1,5,8,10,21}⊂Z30.(8)A_* = \{0, 12, 18, 24, 36\} \subset\mathbb{Z}_{180}, \qquad T = \{0, 1, 5, 8, 10, 21\} \subset\mathbb{Z}_{30}. \tag*{(8)}

Let Q∈Z[X]Q \in\mathbb{Z}[X] be the canonical cyclic autocorrelation mask polynomial of TT:

Q(X)=∑t,t′∈TX(t−t′) mod 30,0≤(t−t′) mod 30<30.(9)Q(X) = \sum_{t,t' \in T} X^{(t-t') \bmod30}, \qquad0 \le(t-t') \bmod30 < 30. \tag*{(9)}

Using QQ, define the following ordinary polynomials:

H(X)=X(1−X12)(1−X18)U5(X36),(10)H(X) = X(1-X^{12})(1-X^{18})U_5(X^{36}), \tag*{(10)}
W(X)=Q(X)U6(X30)+H(X),F(X)=16W(X).(11)W(X) = Q(X)U_6(X^{30}) + H(X), \qquad F(X) = \frac{1}{6}W(X). \tag*{(11)}

The polynomials HH, WW, and FF all have degree less than 180.

Theorem 5.1. The set A∗A_\ast defined in (8) tiles Z180\mathbb{Z}_{180} and satisfies ordinary (T1) and (T2). The polynomial FF defined in (11) has degree 179, nonnegative rational coefficients, and F(0)=1F(0)=1, and it satisfies

A∗(X)F(X)≡U180(X)(mod(X180−1)).A_\ast(X)F(X) \equiv U_{180}(X) \pmod{(X^{180}-1)}.

Moreover,

F(1)=36,SF∗={2,3,4,9},Φ36∤F.F(1)=36,\qquad S_F^\ast=\{2,3,4,9\},\qquad\Phi_{36}\nmid F.

Thus FF satisfies ordinary (T1) and fails (T2). Its cyclic coefficient function is not positive definite.

Proof. The base tiling and the lifted set complement. We first check that A∗A_\ast tiles. An explicit subgroup complement is

B∗=5Z180={0,5,10,…,175},∣B∗∣=36.B_\ast=5\mathbb{Z}_{180}=\{0,5,10,\ldots,175\},\qquad|B_\ast|=36.

Indeed, the elements of A∗A_\ast represent the five residue classes modulo 5 exactly once, and hence A∗⊕B∗=Z180A_\ast\oplus B_\ast=\mathbb{Z}_{180}. To construct the required weak complement, we use the autocorrelation of a different set complement CC, whose layers we choose as follows.

In Z30\mathbb{Z}_{30} take

A0={0,2,3,4,6},B0={0,5,10,15,20,25}.A_0=\{0,2,3,4,6\},\qquad B_0=\{0,5,10,15,20,25\}.

The set A0A_0 represents each residue class modulo 5 exactly once, so A0⊕B0=Z30A_0\oplus B_0=\mathbb{Z}_{30}. Direct multiplication gives

A0(X)=1+X2+X3+X4+X6=Φ5(X)Φ6(X).A_0(X)=1+X^2+X^3+X^4+X^6=\Phi_5(X)\Phi_6(X).

Choose the six shifts

(t0,t1,t2,t3,t4,t5)=(0,0,1,3,6,10)(t_0,t_1,t_2,t_3,t_4,t_5)=(0,0,1,3,6,10)

and define, with all residues reduced modulo 180,

C=⋃j=05{j+6(b+tj):b∈B0}.(12)C=\bigcup_{j=0}^{5}\{j+6(b+t_j):b\in B_0\}. \tag*{(12)}

To see that CC complements A∗A_\ast, observe that its six layers lie in distinct residue classes modulo 6. In the class indexed by jj, subtracting jj and dividing by 6 reduces the required tiling identity to A0⊕(B0+tj)=Z30A_0\oplus(B_0+t_j)=\mathbb{Z}_{30}. Thus A∗⊕C=Z180A_\ast\oplus C=\mathbb{Z}_{180}, with ∣C∣=36|C|=36.

Modulo 30, the residues j+6tjj+6t_j are 0,1,8,21,10,50,1,8,21,10,5. Consequently

C=T+{0,30,60,90,120,150},C(X)=T(X)U6(X30).C=T+\{0,30,60,90,120,150\},\qquad C(X)=T(X)U_6(X^{30}).

Writing qn=[Xn]Qq_n=[X^n]Q for 0≤n<300\le n<30, we obtain

(1C∗1−C)(n)=6qn(mod30),n∈Z180.(13)(1_C\ast1_{-C})(n)=6q_n\pmod{30},\qquad n\in\mathbb{Z}_{180}. \tag*{(13)}

For each ordered pair of representatives from TT, there are six ordered pairs from CC with any prescribed difference in the corresponding coset modulo 30. This proves (13) and shows that the first term of FF in (11) is the canonical mask polynomial of the normalized autocorrelation g=(1C∗1−C)/36g=(1_C\ast1_{-C})/36.

The signed perturbation and nonnegativity. On Z30\mathbb{Z}_{30} let

R(X)=(1−X2)(1−X3)U5(X6).R(X)=(1-X^2)(1-X^3)U_5(X^6).

The cyclic coefficient function of RR repeats the vector (1,0,−1,−1,0,1)(1,0,-1,-1,0,1) five times, and its Fourier transform is supported exactly at characters of order 6. Indeed, U5(X6)U_5(X^6) vanishes at the 30th roots whose orders do not divide 6, while the remaining two factors vanish at orders 1, 2, and 3, but not at order 6. Since Φ6∣A0\Phi_6 \mid A_0, it follows that

A0(X)R(X)≡0(modX30−1).A_0(X)R(X) \equiv0 \pmod{X^{30}-1}.

The polynomial (10) is exactly XR(X6)XR(X^6). Hence

A∗(X)H(X)=XA0(X6)R(X6)≡0(modX180−1).(14)A_*(X)H(X)=XA_0(X^6)R(X^6)\equiv0 \pmod{X^{180}-1}. \tag*{(14)}

The exponents occurring in HH lie in the residue class 11 mod 6, and [X1+6x]H=[Xx]R∈{−1,0,1}[X^{1+6x}]H=[X^x]R\in\{-1,0,1\} for 0≤x<300\le x<30. To check that the negative coefficients can be absorbed by the first term of WW, we count the corresponding differences in TT:

x mod 5x\bmod501234
q(1+6x) mod 30q(1+6x)\bmod3011112

The entries in the second row are all at least one, so every negative coefficient of HH is absorbed and WW has nonnegative coefficients. Since the constant coefficient of HH is zero and q0=∣T∣=6q_0=|T|=6, we also have F(0)=1F(0)=1. Moreover, Q(1)=∣T∣2=36Q(1)=|T|^2=36 and H(1)=0H(1)=0, from which W(1)=216W(1)=216 and F(1)=36F(1)=36 follow. Finally, the set tiling gives 1A∗∗1g=1Z180\mathbf{1}_{A_*}*\mathbf{1}_g=\mathbf{1}_{\mathbb{Z}_{180}}, and (14) shows that the perturbation preserves this identity. This proves the required weak tiling.

The prime-power zeros and ordinary (T1). Multiplication by 6 permutes the residue classes modulo 5, so Φ5∣A∗\Phi_5\mid A_*. Since A∗(1)=5A_*(1)=5, this is its only prime-power cyclotomic factor: a second such factor would force an additional prime factor in the integer value A∗(1)A_*(1). This proves ordinary (T1) for A∗A_*, and (T2) has no nontrivial mixed requirement. The weak-tiling identity now forces

Φ2Φ4Φ3Φ9∣W.\Phi_2\Phi_4\Phi_3\Phi_9\mid W.

It remains to exclude any further prime-power divisors, including those whose orders do not divide 180. Since WW has integer coefficients and W(1)=216=2333W(1)=216=2^3 3^3, a divisor Φpa\Phi_{p^a} must have p∈{2,3}p\in\{2,3\}. Furthermore, q29=1q_{29}=1 and deg⁡H=175\deg H=175, so that deg⁡W=179\deg W=179. Comparing this with the degree of Φpa\Phi_{p^a} leaves only the orders

2,4,8,16,32,64,128,256,and3,9,27,81,243.2,4,8,16,32,64,128,256,\quad\text{and}\quad3,9,27,81,243.

The constant coefficients of the monic remainders at the nine unwanted orders are as follows:

d81632641282562781243[X0](W mod Φd)626565686(15)\begin{array}{c|ccccccccc} d & 8 & 16 & 32 & 64 & 128 & 256 & 27 & 81 & 243\\ [X^0](W\bmod\Phi_d) & 6 & 2 & 6 & 5 & 6 & 5 & 6 & 8 & 6 \tag*{(15)} \end{array}

The table can be checked by hand as follows. Put wn=[Xn]Ww_n=[X^n]W for 0≤n<1800\le n<180 and wn=0w_n=0 otherwise. For d=pad=p^a, the entry in the second row is

∑k≥0(wkd−w(p−1)pa−1+kd).\sum_{k\ge0}\left(w_{kd}-w_{(p-1)p^{a-1}+kd}\right).

This follows by first reducing modulo Xd−1X^d-1 and then using Φpa=1+Xpa−1+⋯+X(p−1)pa−1\Phi_{p^a}=1+X^{p^{a-1}}+\cdots+X^{(p-1)p^{a-1}}. Each sum is finite, and every entry of (15) is nonzero. Consequently SF∗={2,3,4,9}S_F^*=\{2,3,4,9\}, and the values of the corresponding cyclotomic polynomials at one have product 2⋅2⋅3⋅3=36=F(1)2\cdot2\cdot3\cdot3=36=F(1). This proves ordinary (T1).

Failure of (T2) and positive definiteness. At a primitive 36th root, the geometric sum U6(X30)U_6(X^{30}) vanishes. Since

Φ36(X)=X12−X6+1,X18≡−1(modΦ36),\Phi_{36}(X)=X^{12}-X^6+1,\qquad X^{18}\equiv-1\pmod{\Phi_{36}},

formula (10) gives

W≡H≡5X(1−X12)(1−X18)≡20X−10X7(modΦ36).W \equiv H \equiv5X(1-X^{12})(1-X^{18}) \equiv20X-10X^7 \pmod{\Phi_{36}}.

The remainder is nonzero and has degree less than deg⁡Φ36=12\deg\Phi_{36}=12, proving that Φ36∤F\Phi_{36}\nmid F. Since both Φ4\Phi_4 and Φ9\Phi_9 divide FF, this gives the failure of (T2)(T2). Finally, w1=q1+R(0)=2w_1=q_1+R(0)=2, whereas w179=q29=1w_{179}=q_{29}=1. Thus the cyclic coefficient function of FF is not even, and hence is not positive definite. □

Remark 5.2. The weak complement FF belongs to the dilated set A∗=6A0A_* = 6A_0, not to A0A_0 itself. The next section shows that this is unavoidable in general: some tiles with an unsupported divisor have no such weak complement at any period.

What period enlargement does and does not preserve

The construction of Theorem 4.2 replaces A(X)A(X) by A(Xq)A(X^q) before constructing a weak complement. We now examine the role of this change in the set factor. First, we give a tile for which enlarging the ambient period, while keeping its mask polynomial fixed, can never produce a normalized nonnegative (T1)(T1) weak complement failing (T2)(T2). We then show that dilation cannot create a set complement failing (T2)(T2), and that the rational construction requires a period with at least three distinct prime divisors.

Allocation for a rational complement satisfying (T1)

We begin with the allocation of prime-power factors. For a rational complement, integrality no longer gives the counting argument used in Lemma 2.4. The condition (T1)(T1) supplies the missing constraint through the value of the complement at one.

Lemma 6.1. Let A∈Z[X]A\in\mathbb{Z}[X] satisfy a=A(1)>0a=A(1)>0, and suppose that P∈Q[X]P\in\mathbb{Q}[X] satisfies ordinary (T1)(T1) and

A(X)P(X)=UM(X)(mod(XM−1)).A(X)P(X)=U_M(X)\pmod{(X^M-1)}.

Then AA satisfies ordinary (T1)(T1), and SA∗S_A^* and SP∗S_P^* form a disjoint partition of the prime-power orders dividing MM. No positivity assumption is needed.

Proof. We compare the prime-power factors forced into PP by the cyclic identity with those already present in AA. Put

QA=∏pb∣MΦpb∣Ap.Q_A=\prod_{\substack{p^b\mid M\\ \Phi_{p^b}\mid A}}p.

The product of the distinct monic cyclotomic factors counted in QAQ_A has an integral quotient in AA. Evaluating at one gives QA∣aQ_A\mid a, and in particular QA≤aQ_A\leq a. On the other hand, each prime-power order dividing MM that is absent from AA must occur in PP. The values at one of these forced factors multiply to M/QAM/Q_A. Since the cyclic identity also determines P(1)P(1), ordinary (T1)(T1) for PP gives

Ma=P(1)=∏s∈SP∗Φs(1)≥MQA.\frac{M}{a}=P(1)=\prod_{s\in S_P^*}\Phi_s(1)\geq\frac{M}{Q_A}.

It follows that QA≥aQ_A\geq a, so both inequalities must be equalities. The prime-power factors already counted in AA therefore exhaust its value at one, leaving no further prime-power factor. Equality for PP likewise excludes both a factor shared with AA and a factor whose order does not divide MM. This proves the asserted partition and (T1)(T1) for AA. □

An unchanged tile obstructed at every period

The fixed-tile question is motivated by the discussion in [5 Section 10.2]. To show why the set factor cannot in general be left unmodified (and, in our specific context, undilated), we use the period-180 example of H. Tan and Y. Zhang [17 Section 8].

This tile consists of six translates of the five-point set from § 5. Its nonnegative normalized complements satisfy a rigidity property that persists when the ambient period is enlarged.

Proposition 6.2. Let

E={0,2,3,4,6},T=E+30{0,1,2,3,4,5}.E = \{0,2,3,4,6\}, \qquad T = E + 30\{0,1,2,3,4,5\}.

The set TT tiles Z180\mathbb{Z}_{180}, satisfies ordinary (T1)(T1) and (T2)(T2), and has unsupported divisor Φ6\Phi_{6}. Nevertheless, for every integer L≥2L \ge2, every polynomial P∈Q[X]P \in\mathbb{Q}[X] with nonnegative coefficients satisfying

P(0)=1,T(X)P(X)≡UL(X)(mod(XL−1)),P(0)=1,\qquad T(X)P(X) \equiv U_{L}(X) \pmod{(X^{L}-1)},

and ordinary (T1)(T1) must also satisfy ordinary (T2)(T2).

Proof. We first verify the tiling and cyclotomic data. The set EE represents each residue class modulo five exactly once, so E⊕B=Z30E \oplus B = \mathbb{Z}_{30} for B=5{0,1,2,3,4,5}B = 5\{0,1,2,3,4,5\}. Consequently T⊕B=Z180T \oplus B = \mathbb{Z}_{180}. Direct multiplication gives

T(X)=X180−1X30−1Φ5(X)Φ6(X),E(X)=Φ5(X)Φ6(X).(16)T(X)=\frac{X^{180}-1}{X^{30}-1}\Phi_{5}(X)\Phi_{6}(X), \qquad E(X)=\Phi_{5}(X)\Phi_{6}(X). \tag*{(16)}

In particular, ST∗={4,5,9}S_{T}^{*} = \{4,5,9\} and ∣T∣=30|T|=30. Every mixed product of prime powers in this signature divides 180 but not 30, and hence occurs among the factors in the quotient in (16). This proves (T1)(T1) and (T2)(T2). The same factorization gives Φ6∣T\Phi_{6} \mid T and Φ2,Φ3∤T\Phi_{2},\Phi_{3} \nmid T. Since 2,3∣∣T∣2,3 \mid|T|, the divisor Φ6\Phi_{6} is unsupported.

The main point of the proof is that the set EE has only one normalized nonnegative weak complement on Z30\mathbb{Z}_{30}. To see this, suppose that F≥0F \ge0 on Z30\mathbb{Z}_{30}, F(0)=1F(0)=1, and 1E∗F=1Z30\mathbf{1}_{E} * F = \mathbf{1}_{\mathbb{Z}_{30}}. Its mass is six. By (16), the nontrivial Fourier support of FF can occur only at orders five and six. We may therefore separate these two components in the CRT coordinates Z30≃Z5×Z6\mathbb{Z}_{30} \simeq\mathbb{Z}_{5} \times\mathbb{Z}_{6}. Assigning the constant Fourier term to the first component, Fourier inversion gives real functions uu and vv such that

F(x,y)=u(x)+v(y),∑x∈Z5u(x)=1,∑y∈Z6v(y)=0.F(x,y)=u(x)+v(y), \qquad\sum_{x\in\mathbb{Z}_{5}}u(x)=1, \qquad\sum_{y\in\mathbb{Z}_{6}}v(y)=0.

where vv has Fourier support only at order six. Put c=−min⁡yv(y)≥0c=-\min_{y}v(y)\ge0. Nonnegativity of FF gives u(x)≥cu(x)\ge c for every xx, and normalization gives

v(0)=1−u(0)=∑x≠0u(x)≥4c.v(0)=1-u(0)=\sum_{x\ne0}u(x)\ge4c.

The order-six support imposes the additional relation v(0)+v(2)+v(4)=0v(0)+v(2)+v(4)=0, from which v(0)≤2cv(0)\le2c. The two bounds on v(0)v(0) are compatible only if c=0c=0. Thus v≥0v\ge0, and its zero mass implies v=0v=0. We then have u(0)=1u(0)=1 and u(x)=0u(x)=0 for x≠0x\ne0, proving that the only normalized nonnegative weak complement of EE on Z30\mathbb{Z}_{30} is

F=15Z30.(17)F=\mathbf{1}_{5\mathbb{Z}_{30}}. \tag*{(17)}

We next transfer this rigidity to complements of TT. Suppose that h≥0h \ge0 on Z180\mathbb{Z}_{180}, h(0)=1h(0)=1, and 1T∗h=1Z180\mathbf{1}_{T} * h=\mathbf{1}_{\mathbb{Z}_{180}}, and fold hh modulo 30 by setting

F(y)=∑j=05h(y+30j),y∈Z30.F(y)=\sum_{j=0}^{5}h(y+30j), \qquad y\in\mathbb{Z}_{30}.

The six translates in the definition of TT give 1E∗F=1Z30\mathbf{1}_{E} * F=\mathbf{1}_{\mathbb{Z}_{30}}. Moreover, F(0)≥h(0)=1F(0)\ge h(0)=1, whereas the convolution equation at zero, together with 0∈E0\in E and nonnegativity, gives F(0)≤1F(0)\le1. We may therefore apply (6.2). Since folding preserves Fourier values at characters whose orders divide 30, it follows that the Fourier transform of hh vanishes at order six. This conclusion required no (T1)(T1) assumption on hh.

It remains to show that a larger ambient period gives no additional freedom relevant to (T2)(T2). Let PP satisfy the stated hypotheses at period LL. By Lemma 6.1,

180∣L,SP∗={pa:pa∣L}∖{4,5,9}.(18)180\mid L,\qquad S_{P}^{*}=\{p^{a}:p^{a}\mid L\}\setminus\{4,5,9\}. \tag*{(18)}

Write f=wPLf=w_{P}^{L}. The factorization (6.1) shows that every cyclotomic factor of TT has order dividing 180. The cyclic identity therefore confines the Fourier support of ff to characters of these orders, together with the trivial character, and hence f(n+180)=f(n)f(n+180)=f(n). We must also check that this 180-periodic pattern retains the normalization. Nonnegative coefficients and P(0)=1P(0)=1 give f(0)≥1f(0)\ge1. The convolution equation at zero, using 0∈T0\in T, gives the reverse inequality. The 180-periodic pattern h(n)=f(n)h(n)=f(n), 0≤n<1800\le n<180, therefore satisfies

h≥0,h(0)=1,1T∗h=1Z180.h\ge0,\qquad h(0)=1,\qquad\mathbf{1}_{T}*h=\mathbf{1}_{\mathbb{Z}_{180}}.

Applying the preceding argument to hh gives the required order-six vanishing, and therefore Φ6∣P\Phi_{6}\mid P.

We can now verify (T2)(T2) using the complete list of cyclotomic factors of TT. Let d>1d>1 be a product of prime-power orders of distinct primes in SP∗S_{P}^{*}. By (6.3), d∣Ld\mid L. If Φd∤P\Phi_{d}\nmid P, the cyclic identity forces Φd∣T\Phi_{d}\mid T. By (6.1), the possible orders are 5, 6, and the divisors of 180 that do not divide 30. Order five is excluded by (6.3); every order in the last class has constituent prime power four or nine, also excluded. At the remaining order six, we have already proved Φ6∣P\Phi_{6}\mid P. This contradiction proves (T2)(T2). □

The period reduction used in this proof reflects a general restriction on a fixed mask. For a nonzero integer mask polynomial AA, let LAL_{A} be the least common multiple of the orders of its cyclotomic factors, with LA=1L_{A}=1 if there are none. At ambient period LL, the Fourier support of every cyclic complement of AA has orders dividing gcd⁡(L,LA)\gcd(L,L_{A}). The complement consequently has a period dividing this number. Thus enlarging the period cannot introduce Fourier orders beyond those already present among the cyclotomic divisors of AA. The allocation and normalization arguments above are additionally needed for the stated (T1)(T1)–(T2)(T2) obstruction.

Dilation and set complements

For set complements, the residue layers of a dilation give a further restriction: a failure of (T2)(T2) in a complement of qAqA must already occur in one of its layers. The following argument uses only allocation of prime-power factors and does not depend on the integer tiling characterization claimed by C. Li et al. [13].

Proposition 6.3. Let A⊆ZMA\subseteq\mathbb{Z}_{M} tile ZM\mathbb{Z}_{M} and let q≥1q\ge1. Then AA has a set complement failing (T2)(T2) if and only if qA⊆ZqMqA\subseteq\mathbb{Z}_{qM} has such a complement.

Proof. A translation by an element of qAqA preserves each residue class modulo qq. Thus every set complement CC of qAqA decomposes uniquely into residue layers

C=⋃j=0q−1(j+qBj),A⊕Bj=ZM(0≤j<q).C = \bigcup_{j=0}^{q-1}(j+qB_j), \qquad A \oplus B_j = \mathbb{Z}_M \quad(0 \le j < q).

We first show that (T2)(T2) for every layer implies (T2)(T2) for CC. Suppose, accordingly, that all the BjB_j satisfy (T2)(T2). Their ordinary prime-power signatures agree by Lemma 2.4. Let d>1d>1 be a product of prime-power orders of distinct primes in SC∗S_C^*, choose a primitive ddth root ζ\zeta, and put e=d/gcd⁡(d,q)e=d/\gcd(d,q). Then ζq\zeta^q has order ee and

C(ζ)=∑j=0q−1ζjBj(ζq).C(\zeta)=\sum_{j=0}^{q-1}\zeta^j B_j(\zeta^q).

If e=1e=1, then d∣qd\mid q, and the values of the layer mask polynomials in this sum are simply their cardinalities. Each is equal to M/∣A∣M/|A|, so the sum becomes (M/∣A∣)Uq(ζ)=0(M/|A|)U_q(\zeta)=0.

For e>1e>1, we use the common prime-power signature of the layers. A constituent pβ−vp(q)p^{\beta-v_p(q)} of ee comes from a constituent pβp^\beta of dd with β>vp(q)\beta>v_p(q). Since pβ∈SC∗p^\beta\in S_C^*, allocation places this order outside the signature of A(Xq)A(X^q). Formula (4.4) then shows that pβ−vp(q)p^{\beta-v_p(q)} is absent from SA∗S_A^* and belongs to every SBj∗S_{B_j}^*. By (T2)(T2) for the layers, or by their prime-power vanishing if ee is a prime power, Bj(ζq)=0B_j(\zeta^q)=0 for every jj. Thus CC satisfies (T2)(T2). Contrapositively, a complement of qAqA that fails (T2)(T2) has a layer that is already a complement of AA failing (T2)(T2).

For the converse, we use identical layers. Suppose that BB complements AA and fails (T2)(T2) at order ee. The set with mask

C(X)=Uq(X)B(Xq)C(X)=U_q(X)B(X^q)

complements qAqA. Put d=e∏p∣epvp(q)d=e\prod_{p\mid e}p^{v_p(q)}. The constituent prime powers of dd belong to SC∗S_C^* by allocation and (4.4). At a primitive ddth root ζ\zeta, the root ζq\zeta^q has order ee, so B(ζq)≠0B(\zeta^q)\ne0. Also

Uq(ζ)=1−ζq1−ζ≠0,U_q(\zeta)=\frac{1-\zeta^q}{1-\zeta}\ne0,

because e>1e>1. Hence C(ζ)≠0C(\zeta)\ne0, giving its (T2)(T2) failure. □\square

The rational construction avoids this set-level obstruction by adding a fractional perturbation to a normalized autocorrelation. Its weak complement need not be an indicator in any residue layer.

The two-prime boundary

Finally, we discuss what happens when the tiling period MM has at most two distinct prime divisors. As the following simple result shows, no such nonnegative normalized weak tiling complement satisfying (T1)(T1) but failing (T2)(T2) can exist.

Proposition 6.4. Suppose that MM has at most two distinct prime divisors and A⊕B=ZMA\oplus B=\mathbb{Z}_M. If P∈Q[X]P\in\mathbb{Q}[X] satisfies ordinary (T1)(T1) and

A(X)P(X)≡UM(X)(mod(XM−1)),A(X)P(X)\equiv U_M(X)\pmod{(X^M-1)},

then PP satisfies ordinary (T2)(T2). Positivity and normalization of PP are not required.

Proof. By Lemma 6.1, the prime-power orders dividing MM are partitioned between AA and PP. If MM has only one prime divisor, there is no mixed requirement. Otherwise, write its two prime divisors as pp and rr, and suppose that PP fails (T2)(T2) at N=pαrβ∣MN = p^{\alpha}r^{\beta} \mid M. Both constituent prime powers must then be absent from AA, whereas the cyclic identity forces ΦN∣A\Phi_N \mid A.

Since ∣A∣∣M|A| \mid M, the two-prime theorem of E. Coven and A. Meyerowitz [1] gives (T2)(T2) for AA. By G. Kiss, I. Łaba, C. Marshall, and G. Somlai [5 Theorem 9.4], a nonnegative integer multiset on a two-prime cyclic group satisfying (T2)(T2), with a mixed zero but neither constituent prime-power zero, has mass strictly larger than the product of its prime-power cyclotomic values at one. Applied to AA, this contradicts (T1)(T1). This version of the cited theorem does not require the additional cardinality condition in Definition 2.6. □

A period with at least three distinct prime divisors is therefore necessary for a rational complement satisfying (T1)(T1) to fail (T2)(T2). Without the (T1)(T1) assumption this restriction disappears, as the example in Z6\mathbb{Z}_6 following Lemma 2.4 shows.

Positive definiteness and the present constructions

Every spectral set in a finite cyclic group admits a normalized nonnegative weak complement that is positive definite [8 Lemma 1.3]. G. Kiss, I. Londner, M. Matolcsi, and G. Somlai [6 Problems 1.1 and 1.2] ask whether every set in a finite cyclic group admitting a PD weak complement is a tile, and whether the Coven–Meyerowitz conditions extend to such tilings. They answer the first question affirmatively for prime-power periods [6 Proposition 4.1]. They also construct functional PD tilings f∗g=1ZMf * g = 1_{\mathbb{Z}_M}, with M=p4q2M = p^4q^2 for primes p<q<p2p < q < p^2, in which both factors satisfy (T1)(T1) and neither satisfies (T2)(T2) [6 Theorem 1.6]. Neither factor is the indicator function of a set. Related weak-tiling questions are considered in [4, 7]. Geometric consequences of weak tiling in Euclidean spaces are studied in [10].

The explicit weak complements were shown to fail positive definiteness by comparing their values at opposite points. There is also a Fourier reason why the perturbation cannot preserve this property. We start with an autocorrelation, change only Fourier values that are initially zero, and leave the value at the origin unchanged. The following proposition shows that these requirements exclude every nonzero perturbation on that Fourier support.

Proposition 7.1. Let g,H:ZD→Rg, H : \mathbb{Z}_D \to\mathbb{R}, with gg positive definite, H≠0H \ne0, and H(0)=0H(0) = 0. Suppose that

supp⁡H^⊆{k∈ZD:g^(k)=0}.\operatorname{supp}\widehat{H} \subseteq\{k \in\mathbb{Z}_D : \widehat{g}(k) = 0\}.

Then g+εHg + \varepsilon H is not positive definite for any real ε≠0\varepsilon\ne0.

Proof. Suppose that g+εHg + \varepsilon H were positive definite. At a Fourier zero of gg, this would require εH^(k)\varepsilon\widehat{H}(k) to be real and nonnegative. Outside that zero set, H^\widehat{H} vanishes by assumption. Every Fourier value of εH\varepsilon H would therefore be real and nonnegative, and inversion at the origin would give

0=εH(0)=1D∑k∈ZDεH^(k).0 = \varepsilon H(0) = \frac{1}{D}\sum_{k\in\mathbb{Z}_D}\varepsilon\widehat{H}(k).

The summands are nonnegative and have sum zero, so they must all vanish. Fourier inversion then gives H=0H = 0, a contradiction. □

To apply the proposition to Case 2 of Theorem 4.2, recall that the perturbation HH has pure Fourier order qNqN, while the layered complement CC vanishes at that order. The Fourier transform g^\widehat{g} of its normalized autocorrelation therefore vanishes on supp⁡H^\operatorname{supp}\widehat{H}, and H(0)=0H(0) = 0. The same argument applies to the perturbation of the period-180 example: its Fourier support has order 36, at which the autocorrelation term vanishes.

The appendix construction has the same feature in product coordinates. Under the identification ZD≃ZM×Zr\mathbb{Z}_D \simeq\mathbb{Z}_M \times\mathbb{Z}_r, the support of H^\widehat{H} consists of characters whose old component has order NN. Each translated layer of B↑B^\uparrow vanishes at every such old character because ΦN∣B\Phi_N \mid B. Hence g^\widehat{g} is zero on this entire support, and again H(0,0)=0H(0,0)=0. Proposition 7.1 applies. Changing an ordinary polynomial representative, as in the coefficient transfer lemmas, leaves these cyclic functions and their Fourier transforms unchanged.

Symmetrization does not remove this obstruction while retaining a nonzero perturbation. Indeed, let Hev(x)=(H(x)+H(−x))/2H_{\mathrm{ev}}(x)=(H(x)+H(-x))/2. Since gg is real and positive definite, it is even and its Fourier zero set is invariant under negation. Thus Hev(0)=0H_{\mathrm{ev}}(0)=0 and its Fourier support is still contained in that zero set. If Hev≠0H_{\mathrm{ev}}\ne0, the proposition applies; if Hev=0H_{\mathrm{ev}}=0, symmetrization returns gg and restores the mixed Fourier zero.

The preceding proposition concerns the form of the perturbation. For the five-point set used in the explicit example, we can make a stronger statement about the complements themselves. Restriction to the subgroup corresponding to the original period allows us to use the rigidity argument from Proposition 6.2.

Corollary 7.2. Let E={0,2,3,4,6}E=\{0,2,3,4,6\} and q≥1q\ge1. Suppose that P∈Q[X]P\in\mathbb{Q}[X] has nonnegative coefficients, satisfies P(0)=1P(0)=1 and ordinary (T1), and obeys

E(Xq)P(X)≡U30q(X)(mod(X30q−1)).E(X^q)P(X)\equiv U_{30q}(X)\pmod{(X^{30q}-1)}.

If its cyclic coefficient function on Z30q\mathbb{Z}_{30q} is positive definite, then PP satisfies ordinary (T2). In particular, this exclusion applies to A∗=6E⊂Z180A_* = 6E \subset\mathbb{Z}_{180}.

Proof. We first restrict the complement to a copy of Z30\mathbb{Z}_{30}, where its values are determined by the set rigidity argument. Set D=30qD=30q and f=wPf=w_P. Nonnegativity and P(0)=1P(0)=1 give f(0)≥1f(0)\ge1, whereas the convolution equation at zero gives f(0)≤1f(0)\le1. Define h(x)=f(qx)h(x)=f(qx) on Z30\mathbb{Z}_{30}. Restricting the convolution equation to qZDq\mathbb{Z}_D gives

h≥0,h(0)=1,1E∗h=1Z30.h\ge0,\qquad h(0)=1,\qquad\mathbf{1}_E * h=\mathbf{1}_{\mathbb{Z}_{30}}.

The set rigidity argument in Proposition 6.2 yields h=15Z30h=\mathbf{1}_{5\mathbb{Z}_{30}}, so h^\widehat{h} vanishes at order six. Restriction to the subgroup also gives the Fourier identity

h^(ℓ)=1q∑k∈ZDk≡ℓ mod 30f^(k),ℓ∈Z30.(19)\widehat{h}(\ell)=\frac{1}{q}\sum_{\substack{k\in\mathbb{Z}_D\\ k\equiv\ell\ \mathrm{mod}\ 30}}\widehat{f}(k),\qquad\ell\in\mathbb{Z}_{30}. \tag*{(19)}

Positive definiteness makes every summand nonnegative. Thus a positive Fourier value of ff contributes a positive Fourier value at the order of its restriction, with no possibility of cancellation.

We use this observation to exclude a failure of (T2). If PP failed that condition, Lemma 6.1 would give an order d∣Dd\mid D whose constituent prime powers all belong to SP∗S_P^* but for which Φd∤P\Phi_d\nmid P. Choose a character of order dd at which f^(k)>0\widehat{f}(k)>0, and put e=d/gcd⁡(d,q)e=d/\gcd(d,q). The tiling identity forces E(Xq)E(X^q) to vanish at this character, so EE vanishes at order ee. Since E(X)=Φ5(X)Φ6(X)E(X)=\Phi_5(X)\Phi_6(X), we have e∈{5,6}e\in\{5,6\}.

If e=5e=5, then 51+ν5(q)5^{1+\nu_5(q)} is a constituent prime power of dd and belongs to the signature of E(Xq)E(X^q) by (7). This contradicts allocation, since every constituent of dd belongs to SP∗S_P^*. Therefore e=6e=6. The residue ℓ=k mod 30\ell=k\bmod30 then has character order six. Equation (19) gives h^(ℓ)>0\widehat{h}(\ell)>0, contradicting the order-six vanishing of hh. Hence PP satisfies (T2)(T2). □

The proposition explains the limitation of the signed perturbations, and the corollary gives an exclusion for the particular set EE. Whether other set tiles admit normalized nonnegative positive-definite (T1)(T1) complements failing (T2)(T2) remains open. Case 1 of Theorem 4.2 is separate from this perturbation argument: it starts from a set complement already failing (T2)(T2). Such a complement would also give a positive-definite weak complement by normalized autocorrelation, as discussed in Remark 4.4.

Acknowledgements

G. K. was supported by grants NKFIH-142993, NKFIH-150576, and NKFIH-154121.

C. M. received support for postdoctoral research during the Simons Special Semester at the Rényi Center and was also supported by an Arts & Science PDF Award and an NSERC Canadian PDF Award.

M. M. was supported by grants NKFIH-146387 and NKFIH-154121.

G. S. was supported by a Discovery Project grant (DP250104965) of the Australian Research Council and grant NKFIH-150576.

Normalization by shifted layers

We give an alternative construction on ZM×Zr\mathbb{Z}_M \times\mathbb{Z}_r, where rr is a prime coprime to MM. The product coordinates allow us to place the filter in a nonzero layer, leaving the value at the origin unchanged. As before, translations of the set complement provide enough positive weight to absorb the negative filter coefficients.

The general construction

A.1. The general construction. Assume throughout this subsection that A⊕B=ZMA \oplus B=\mathbb{Z}_M, that AA satisfies (T2)(T2), and that

N=∏i=1kpiαi∣M,k≥2,ΦN∣A,ΦN∣B,Φpiαi∣B(1≤i≤k).(20)N=\prod_{i=1}^{k}p_i^{\alpha_i}\mid M,\qquad k\ge2,\qquad\Phi_N\mid A,\qquad\Phi_N\mid B,\qquad\Phi_{p_i^{\alpha_i}}\mid B\quad(1\le i\le k). \tag*{(20)}

In particular, the zero of BB at the common mixed order NN is one of the zeros required by (T2)(T2). We will remove this zero while retaining the constituent prime-power zeros. The additional cardinality condition in the definition of an unsupported divisor is not needed for this construction.

Let R=RN,MR=R_{N,M}, let S−={x:R(x)=−1}S_-=\{x:R(x)=-1\}, and choose a prime r>2r>2 coprime to MM. Put

D=rM,K=r∣B∣,ZD≃ZM×Zr.D=rM,\qquad K=r|B|,\qquad\mathbb{Z}_D\simeq\mathbb{Z}_M\times\mathbb{Z}_r.

Here and below this is the literal residue-coordinate identification: z↦(z mod M,z mod r)z\mapsto(z\bmod M,z\bmod r). Choose shifts t0,…,tr−1∈ZMt_0,\ldots,t_{r-1}\in\mathbb{Z}_M and define

A↑=A×{0},B↑=⋃j∈Zr(B+tj)×{j}.(21)A^\uparrow=A\times\{0\},\qquad B^\uparrow=\bigcup_{j\in\mathbb{Z}_r}(B+t_j)\times\{j\}. \tag*{(21)}

Each layer is a translated copy of the base tiling, and consequently A↑⊕B↑=ZDA^\uparrow\oplus B^\uparrow= \mathbb{Z}_D. We now take the normalized autocorrelation of this complement and place the filter in layer one:

g=1K1B↑∗1−B↑,H(x,j)=R(x)1{1}(j),f=g+εH.(22)g = \frac{1}{K}\mathbf{1}_{B^\uparrow} \ast\mathbf{1}_{-B^\uparrow}, \qquad H(x,j) = R(x)\mathbf{1}_{\{1\}}(j), \qquad f = g + \varepsilon H. \tag*{(22)}

Let FF be the canonical mask polynomial of ff. The following lower bound on gg will ensure nonnegativity; we arrange it in Lemma A.2.

Theorem A.1. Under (A.1), suppose that ε∈Q>0\varepsilon\in\mathbb{Q}_{>0} and

g(x,1)≥ε(x∈S−).(23)g(x,1) \ge\varepsilon\qquad(x \in S_-). \tag*{(23)}

Then the function in (A.3) satisfies

f≥0,f(0,0)=1,1A↑∗f=1ZD.f \ge0,\qquad f(0,0)=1,\qquad\mathbf{1}_{A^\uparrow}\ast f=\mathbf{1}_{\mathbb{Z}_D}.

Its mass is KK, and its period-relative prime-power zero set is

SF,D=SB∗∪{r}.S_{F,D}=S_B^*\cup\{r\}.

It satisfies period-relative (T1) and fails (T2) at order NN. The set A↑A^\uparrow satisfies ordinary (T1) and (T2), while ff is not positive definite.

Proof. By Lemma 2.8, gg has the required tiling identity, mass, and value at the origin. The perturbation preserves the tiling identity because 1A↑∗H=(1A∗R)⊗1{1}=0\mathbf{1}_{A^\uparrow}\ast H=(\mathbf{1}_A\ast R)\otimes\mathbf{1}_{\{1\}}=0 by Lemma 3.1. Its negative sites are exactly the points (x,1)(x,1) with x∈S−x\in S_-, where (A.4) ensures f≥0f\ge0. Finally, H(0,0)=0H(0,0)=0 and mass⁡(H)=0\operatorname{mass}(H)=0, so the normalization and mass of gg are also preserved.

Each translated layer of B↑B^\uparrow retains the prime-power zeros of BB, while equal layer sizes give the zero at order rr. Their cyclotomic values at one have product r∣B∣=K=∣B↑∣r|B|=K=|B^\uparrow|, by Lemma 2.4. Consequently these are all the prime-power factors of its integer mask, and autocorrelation preserves exactly these Fourier zeros. The perturbation cannot change them: a character in the support of H^\widehat{H} has old component of mixed order NN, and hence cannot have prime-power order. Thus ff has precisely the asserted prime-power zeros, and their values at one give its mass KK.

At the mixed order NN, the situation is different. Let χ\chi be any base character of order NN, and take the trivial character in the new coordinate. Since ΦN∣B\Phi_N\mid B, each translated layer has zero Fourier value at χ\chi. Hence g^(χ,0)=0\widehat{g}(\chi,0)=0, whereas

f^(χ,0)=εR^(χ)≠0.\widehat{f}(\chi,0)=\varepsilon\widehat{R}(\chi)\ne0.

Since the constituent prime-power zeros in (A.1) were preserved, this nonzero value gives the asserted failure of (T2).

For the set factor, every old prime-power zero of AA is inherited by A↑A^\uparrow. These factors already account for its unchanged cardinality, so there are no additional prime-power factors. Every product required by (T2) is therefore an old order, at which vanishing is inherited from AA. To finish, we check that the perturbation destroys evenness. The autocorrelation gg is even and R(0)=1R(0)=1, while r>2r>2, so

f(0,1)−f(0,−1)=ε.(24)f(0,1)-f(0,-1)=\varepsilon. \tag*{(24)}

Thus ff is not even and cannot be PD.

We can arrange the required lower bound by prescribing the differences between successive layer shifts.

Lemma A.2. For every prime r>2r > 2 coprime to MM with r−1≥∣S−∣r - 1 \ge|S_{-}|, one can choose shifts such that g(x,1)≥∣B∣/K=1/rg(x,1) \ge|B|/K = 1/r for every x∈S−x \in S_{-}. Consequently, condition (A.4) holds whenever 0<ε≤1/r0 < \varepsilon\le1/r.

Proof. List S−S_{-} as x1,…,xsx_1,\ldots,x_s. We prescribe these differences as successive increments of the shifts, using the last increment to close the cycle. Define

dj=xj (1≤j≤s),dj=0 (s<j<r),d0=−∑j=1r−1dj mod M,d_j = x_j\ (1 \le j \le s), \qquad d_j = 0\ (s < j < r), \qquad d_0 = -\sum_{j=1}^{r-1}d_j \mathbin{\bmod} M,

and put t0=0t_0 = 0, tj=∑h=1jdh mod Mt_j = \sum_{h=1}^{j}d_h \mathbin{\bmod} M for 1≤j<r1 \le j < r. Then the cyclic increment tj−tj−1t_j-t_{j-1} equals djd_j, including at j=0j=0. If an increment has the prescribed value xx, the ∣B∣|B| ordered pairs

((b+tj,j),(b+tj−1,j−1)),b∈B,((b+t_j,j),(b+t_{j-1},j-1)), \qquad b \in B,

all have difference (x,1)(x,1). Counting these pairs in the normalized autocorrelation gives the stated lower bound. Since only finitely many primes divide MM, we may choose a prime rr as large as required. □

With these shifts, we may in particular take the rational value ε=1/K\varepsilon= 1/K. At every negative filter site this gives f(x,1)≥(∣B∣−1)/Kf(x,1) \ge(|B|-1)/K, while the value at the origin remains unchanged.

An ordinary polynomial representative

To pass from period-relative to ordinary (T1)(T1), we use Lemma 4.1 to remove cyclotomic factors whose orders do not divide DD, without changing the cyclic function.

Corollary A.3. The construction of Theorem A.1 admits an ordinary nonnegative rational representative PP with P(0)=1P(0)=1 satisfying ordinary (T1)(T1) and failing (T2)(T2), without changing its cyclic coefficient function.

Proof. Since f(x)≥ε>0f(x) \ge\varepsilon> 0 at x=CRT⁡(0,1)x = \operatorname{CRT}(0,1), this nonzero residue can be used for the coefficient transfer. Choose a positive integer mm clearing the denominators of FF and an odd prime ℓ∣N\ell\mid N. Lemma 4.1 gives

P(X)=F(X)+Xx+D−Xxℓm.P(X) = F(X) + \frac{X^{x+D}-X^x}{\ell m}.

This polynomial has the same cyclic coefficient function and constant coefficient as FF, and all its cyclotomic factors have orders dividing DD. Hence its period-relative (T1)(T1) condition is ordinary (T1)(T1), while its failure of (T2)(T2) at order NN is unchanged. □

References

Gergely Kiss

HUN-REN Alfréd Rényi Institute of Mathematics, Analysis Department, Reáltanoda u. 13-15, 1053, Budapest, Hungary,

and

Corvinus University of Budapest, Fővám tér 13-15, 1093, Budapest, Hungary

Email address: [email protected]

Caleb Marshall,

University of Toronto, Department of Mathematics, St. George Campus, Toronto, ON, Canada

Email address: [email protected]

Máté Matolcsi,

HUN-REN Alfréd Rényi Institute of Mathematics, Reáltanoda u. 13-15, 1053, Budapest, Hungary,

and

Budapest University of Technology and Economics, Department of Analysis and Operations Research, Institute of Mathematics, Műegyetem rkp. 3., H-1111 Budapest, Hungary.

Email address: [email protected]

Gábor Somlai, The University of Melbourne, School of Mathematics and Statistics, Parkville VIC 3010, Australia,

and

on unpaid leave from Eötvös Loránd University

Email address: [email protected]

References

  1. [1]E. M. Coven and A. D. Meyerowitz, Tiling the integers with translates of one finite set, J. Algebra 212 (1999), 161–174. doi:10.1006/jabr.1998.7628.DOI
  2. [2]T. Fallon, G. Kiss, A. Mayeli, and G. Somlai, Spectral Tile Direction in the Group Zₚ² × Zq₄² × Zr, arXiv:2308.16277v4 (2026).
  3. [3]R. Greenfeld and T. Tao, A counterexample to the periodic tiling conjecture, Ann. of Math. (2) 200 (2024), 301–363. doi:10.4007/annals.2024.200.1.5.DOI
  4. [4]M. Kadir and K. Fan, Tiles and weak tiles in Zₚq, preprint (2026). arXiv:2607.02149.
  5. [5]G. Kiss, I. Laba, C. Marshall, and G. Somlai, Lower bounds for mask polynomials with many cyclotomic divisors, Adv. Math. 494 (2026), 110932. doi:10.1016/j.aim.2026.110932. arXiv:2507.11672.DOI
  6. [6]G. Kiss, I. Londner, M. Matolcsi, and G. Somlai, Functional tilings and the Coven–Meyerowitz tiling conditions, Discrete Analysis 2025:18, 20 pp. arXiv:2411.03854.arxiv.org/abs/2411.03854
  7. [7]G. Kiss, I. Londner, M. Matolcsi, and G. Somlai, A lonely weak tile, Expo. Math. 44 (2026), no. 1, 125636, 9 pp. doi:10.1016/j.exmath.2024.125636. arXiv:2410.04948.
  8. [8]G. Kiss, D. Matolcsi, M. Matolcsi, and G. Somlai, Tiling and weak tiling in (Zₚ)ᵈ, Sampl. Theory Signal Process. Data Anal. 22 (2024), Paper No. 1. doi:10.1007/s43670-023-00073-7.DOI
  9. [9]M. N. Kolountzakis, N. Lev, and M. Matolcsi, Spectral sets and weak tiling, Sampl. Theory Signal Process. Data Anal. 21 (2023), Paper No. 31.DOI
  10. [10]M. N. Kolountzakis, N. Lev, and M. Matolcsi, Geometric implications of weak tiling, Anal. Math. 51 (2025), 1375–1393. doi:10.1007/s10476-025-00121-5.DOI
  11. [11]I. Laba and I. Londner, Combinatorial and harmonic-analytic methods for integer tilings, Forum Math. Pi 10 (2022), e8, 1–46. doi:10.1017/fmp.2022.3.DOI
  12. [12]N. Lev and M. Matolcsi, The Fuglede conjecture for convex domains is true in all dimensions, Acta Math. 228 (2022), 385–420.DOI
  13. [13]C. Li, J. Wu, W. Hu, C. Wang, E. Wang, and F. Zhu, Reduction of integer tiles via CRT and base-p digits, preprint (2026). arXiv:2609.14026v1.arxiv.org/abs/2609.14026
  14. [14]R. D. Maliokisis, On the structure of spectral and tiling subsets of cyclic groups, Forum Math. Sigma 10 (2022), e23, 1–42. doi:10.1017/fms.2022.14.
  15. [15]D. J. Newman, Tessellation of integers, J. Number Theory 9 (1977), 107–111.
  16. [16]J. Prajapati, Coven–Meyerowitz T2 necessity through coprime stripe collapse, preprint (2026). arXiv:2609.22280.arxiv.org/abs/2609.22280
  17. [17]H. Tan and Y. Zhang, Unsupported Cyclotomic Divisors in Three-Prime Integer Tilings, preprint (2026). arXiv:2609.06677v1.arxiv.org/abs/2609.06677

Paper details

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

views · 7

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents