betaFeedback

WorksMathematics

The MZ-property for three-term differences

Abstract

We study the Marcinkiewicz–Zygmund property for finite differences. A difference D(f,h)=∑k=1nakf(bkh) D(f,h)=\sum_{k=1}^n a_k f(b_kh) is said to have the MZ-property if, for every measurable function ff, D(f,h)=o(h)andf(h)=o(1) D(f,h)=o(h) \quad\text{and}\quad f(h)=o(1) imply f(h)=o(h). f(h)=o(h). In this paper, we characterize all three-node differences D(f,h)=af(qh)+bf(rh)+cf(sh) D(f,h)=af(qh)+bf(rh)+cf(sh) having the MZ-property. In particular, for irrational differences with distinct nodes, we obtain explicit necessary and sufficient conditions in terms of the coefficients a,b,ca,b,c and the nodes q,r,sq,r,s. As an application, our results resolve an open problem concerning a second-order Riemann difference.

We consider three-term functional differences of the form

D(f,h)=af(qh)+bf(rh)+cf(sh),(1)D(f,h)=af(qh)+bf(rh)+cf(sh), \tag*{(1)}

where hh is a real variable and ff is an arbitrary complex-valued function defined for real numbers in a neighborhood of 0. Each such difference DD has fixed parameters. These include nonzero coefficients, aa, bb, cc, which may be real or complex, and distinct nodes, qq, rr, ss, which are nonzero real numbers. We consider three possible limit properties of ff as h→0h\to0:

  • A. D(f,h)=o(h)D(f,h)=o(h)

  • B. f(h)=o(1)f(h)=o(1)

  • C. f(h)=o(h)f(h)=o(h)

Property C trivially implies both A and B, for any choice of parameters. If A and B together imply C then the difference D(f,h)D(f,h) is said to have the MZ-property.

A case requiring special consideration is the following.

Definition 1. A difference (1) is called rational if there are three integers i,j,ki,j,k, not all zero, such that ilog⁡∣q∣+jlog⁡∣r∣+klog⁡∣s∣=i+j+k=0i\log|q|+j\log|r|+k\log|s|=i+j+k=0. Otherwise, it is called irrational.

A characterization of rational differences with the MZ-property is contained in the results of [6] and will be discussed below. In this paper we characterize the irrational three-term differences that have the MZ-property. Our main result is the following:

Theorem 2. Let D(f,h)=af(qh)+bf(rh)+cf(sh)D(f,h)=af(qh)+bf(rh)+cf(sh) be an irrational difference with distinct nodes, qq, rr, ss, where ∣q∣≤∣r∣≤∣s∣|q|\le|r|\le|s|. Then DD has the MZ-property if and only if at least one of the following conditions holds

  • ∣a∣+∣b∣≤∣c∣|a| + |b| \le|c|

  • ∣br∣+∣cs∣<∣aq∣|br| + |cs| < |aq|

  • ∣a∣+∣c∣≤∣b∣|a| + |c| \le|b| and ∣cs∣+∣aq∣<∣br∣|cs| + |aq| < |br|

Since the MZ-property is invariant under scaling either the coefficients {a,b,c}\{a,b,c\} or the nodes {q,r,s}\{q,r,s\}, we will often rewrite the difference in standard form, with q=c=1q=c=1, 1≤∣r∣≤∣s∣1 \le|r| \le|s|, and

D(f,h)=af(h)+bf(rh)+f(sh).(2)D(f,h)=af(h)+bf(rh)+f(sh). \tag*{(2)}

Then the definition of rational takes on the more familiar form:

Definition 3. A difference (2) is called rational if and only if log⁡∣r∣log⁡∣s∣∈Q\frac{\log|r|}{\log|s|} \in\mathbb{Q}.

Theorem 2 can then be re-written in a less symmetrical, but equivalent form:

Theorem 2 (Alternate version). Let D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh) be an irrational difference with distinct nodes, 1,r,s1,r,s where 1≤∣r∣≤∣s∣1 \le|r| \le|s|. Then DD has the MZ-property if and only if at least one of the following conditions holds

  • ∣a∣+∣b∣≤1|a|+|b| \le1

  • ∣br∣+∣s∣<∣a∣|br|+|s| < |a|

  • ∣a∣+1≤∣b∣|a|+1 \le|b| and ∣s∣+∣a∣<∣br∣|s|+|a| < |br|.

Real differences

The original motivation for the MZ-property was to study smoothness properties of real functions. However, as shown in [6], this property cannot be properly studied without also including complex numbers. Hence, we allow complex numbers for the coefficients, a,b,ca,b,c, as well as the output of the arbitrary function ff. In this section we will justify the use of complex numbers for the case that we are only interested in real numbers.

Definition 4. A difference (2) is called real if the coefficients, a,b,ca,b,c are all real numbers.

Definition 5. A difference (2) has the real MZ-property if A and B imply C for all real-valued functions.

Theorem 6. A real difference (2) has the MZ-property if and only if it has the real MZ-property.

Proof. First, suppose that a real difference, DD, has the MZ-property. Then properties A and B together imply C for any function f:R→Cf:\mathbb{R}\to\mathbb{C}. Trivially, A and B together imply C for any function f:R→Rf:\mathbb{R}\to\mathbb{R}.

Next, suppose that a real difference fails the MZ-property. Then there is a function f:R→Cf:\mathbb{R}\to\mathbb{C} such that D(f,h)=o(h)D(f,h)=o(h), f(h)=o(1)f(h)=o(1), but f(h)f(h) is not o(h)o(h). Let r(h)=Re⁡(f)r(h)=\operatorname{Re}(f), and c(h)=Im⁡(f)c(h)=\operatorname{Im}(f), so that r:R→Rr:\mathbb{R}\to\mathbb{R} and c:R→Rc:\mathbb{R}\to\mathbb{R}. Since f=o(1)f=o(1), we have both r=o(1)r=o(1) and c=o(1)c=o(1). Since D(f,h)=o(h)D(f,h)=o(h), we have Re⁡(D(f,h))=o(h)\operatorname{Re}(D(f,h))=o(h) and Im⁡(D(f,h))=o(h)\operatorname{Im}(D(f,h))=o(h). But since DD is a real difference, Re⁡(D(f,h))=D(Re⁡(f),h)=D(r,h)\operatorname{Re}(D(f,h))=D(\operatorname{Re}(f),h)=D(r,h), and Im⁡(D(f,h))=D(Im⁡(f),h)=D(c,h)\operatorname{Im}(D(f,h))=D(\operatorname{Im}(f),h)=D(c,h).

Finally, if r(h)r(h) and c(h)c(h) were both o(h)o(h), then we would have f(h)=o(h)f(h)=o(h). By choice of ff, this does not happen. So, at least one of r(h),c(h)r(h),c(h) witnesses that DD does not have the real MZ-property.

A consequence of this is that Theorem 2 also characterizes the real MZ-property for real differences. For the remainder of this paper, we will only discuss the complex variant.

Background

When an ordinary derivative is not available, one may consider generalized derivatives such as the symmetric derivative

fs′(c)=lim⁡h→0f(c+h)−f(c−h)2h.f'_s(c)=\lim_{h\to0}\frac{f(c+h)-f(c-h)}{2h}.

The symmetric derivative is a generalization of ordinary differentiation since if f′(c)f'(c) exists then so does fs′(c)f'_s(c) and they are equal. Furthermore, it is a proper generalization since

f(h)=∣h∣f(h)=|h|

is symmetrically differentiable at 0, but not ordinarily differentiable there.

More generally, derivatives of the form

D′f(c)=lim⁡h→0∑k=1nakf(c+bkh)h.(3)D'f(c)=\lim_{h\to0}\frac{\sum_{k=1}^{n}a_kf(c+b_kh)}{h}. \tag*{(3)}

can be considered, where the coefficients aia_i are real numbers.

To generalize the ordinary derivative, we need the derivative of a constant to be 0 and the derivative of the identity function to be 1. That is, we need

∑k=1nak=0and∑k=1nakbk=1.(4)\sum_{k=1}^{n}a_k=0 \qquad\text{and}\qquad\sum_{k=1}^{n}a_kb_k=1. \tag*{(4)}

It is not hard to see that these two conditions guarantee that D′f(c)D'f(c) is a generalization of the ordinary derivative. That is, under these two conditions, whenever f′(c)f'(c) exists, so does D′f(c)D'f(c), and D′f(c)=f′(c)D'f(c)=f'(c). For this reason, derivatives that satisfy these two conditions are called Riemann derivatives.

A Riemann derivative, however, might not yield anything new; it may be equivalent to ordinary differentiation. For example, let

S′f(c)=lim⁡h→02f(c+h)+f(c−h)−3f(c)h.S'f(c)=\lim_{h\to0}\frac{2f(c+h)+f(c-h)-3f(c)}{h}.

If S′f(c)S'f(c) exists, then f′(c)f'(c) exists, and f′(c)=S′f(c)f'(c)=S'f(c). So, unlike the symmetric derivative S′S' is not a proper generalization of the ordinary derivative.

It is natural to wonder, therefore, which Riemann derivatives are equivalent to the ordinary derivative and which are proper generalizations. Ash, Catoiu, and Csörnyei [1] gave a complete answer to this question. Specifically, a derivative (3) is equivalent to the ordinary derivative if and only if it can be written in the form

R′f(c)=lim⁡h→0af(c+qh)+af(c−qh)−2af(c)+bf(c+rh)−bf(c−rh)2h.R'f(c)=\lim_{h\to0}\frac{af(c+qh)+af(c-qh)-2af(c)+bf(c+rh)-bf(c-rh)}{2h}.

where aq≠0aq\ne0 and br=1br=1. The derivative S′S' above is obtained from R′R' by setting a=3a=3 and q=r=b=1q=r=b=1.

This raises the question: What if we add continuity as an additional assumption? That is, when does

D′f(c)=kandlim⁡h→0f(c+h)=f(c)(5)D'f(c)=k \qquad\text{and}\qquad\lim_{h\to0}f(c+h)=f(c) \tag*{(5)}

imply

f′(c)=k?f'(c)=k?

To study this question, we assume, without loss of generality, that c=0c=0 and subtract an affine function so that f(c)=k=0f(c)=k=0. This leads us to the following definition, named after Marcinkiewicz and Zygmund since they are believed to be the first to consider this property (see [7]).

Definition 7. A finite difference

D(f,h)=∑k=1nakf(bkh),D(f,h)=\sum_{k=1}^{n}a_kf(b_kh),

is said to have the MZ-property if and only if for any real function ff defined in a neighborhood of 00,

D(f,h)=o(h) and f(h)=o(1) ⟹ f(h)=o(h).(6)D(f,h)=o(h)\ \text{and}\ f(h)=o(1)\ \Longrightarrow\ f(h)=o(h). \tag*{(6)}

Note that although the MZ-property was motivated by the study of Riemann derivatives, it may have intrinsic and independent interest, so we make no tacit assumptions on the properties (4) in this paper. Also note, as discussed above, that we have changed this definition to include complex coefficients and functions f:R→Cf:\mathbb{R}\to\mathbb{C}. Finally for simplicity, we have ruled out the possibility that bk=0b_k=0. This is without loss of generality, since otherwise, D(f,h)=o(h)D(f,h)=o(h) and f(h)=o(1)f(h)=o(1) imply that f(bkh)=f(0)=0f(b_kh)=f(0)=0.

Remark 8. In their study of mm-th Riemann derivatives2, Marcinkiewicz and Zygmund needed, in place of (6), the implication

D(f,h)=o(hm)andf(h)=o(hm−1)⟹f(h)=o(hm).D(f,h)=o(h^m)\quad\text{and}\quad f(h)=o(h^{m-1})\quad\Longrightarrow\quad f(h)=o(h^m).

Here mm is a positive integer, and we call this the MZ-property of order mm. However, the MZ-property of order mm reduces to the MZ-property defined above. Indeed, define

D1(g,h)=∑k=1nakbkm−1g(bkh).D_1(g,h)=\sum_{k=1}^{n}a_kb_k^{m-1}g(b_kh).

Then DD has the MZ-property of order mm if and only if D1D_1 has the MZ-property. To see this, suppose first that DD fails the MZ-property of order mm for some ff, and set

g(h)=f(h)hm−1.g(h)=\frac{f(h)}{h^{m-1}}.

Then g(h)=o(1)g(h)=o(1), while

D(f,h)=hm−1D1(g,h).D(f,h)=h^{m-1}D_1(g,h).

Thus D1(g,h)=o(h)D_1(g,h)=o(h), whereas g(h)≠o(h)g(h)\ne o(h), so D1D_1 fails the MZ-property. Conversely, if D1D_1 fails the MZ-property for some gg, then setting

f(h)=hm−1g(h)f(h)=h^{m-1}g(h)

shows that DD fails the MZ-property of order mm.

The MZ-property has been studied for higher order derivatives and for various number of terms, (See [2], [3], and [4].) though a complete characterization has not yet appeared. The primary result of this study solves the case of geometric nodes, defined as follows.

Dmf(c)=lim⁡h→0D(fc,h)hm,fc(x)=f(c+x).D^mf(c)=\lim_{h\to0}\frac{D(f_c,h)}{h^m},\qquad f_c(x)=f(c+x).

Definition 9. Let λ\lambda be a real number with ∣λ∣>1|\lambda| > 1. Then the finite difference

D(f,h)=∑k=0n(akf(λkh)+ckf(−λkh))D(f,h)=\sum_{k=0}^{n}\left(a_kf(\lambda^kh)+c_kf(-\lambda^kh)\right)

is said to have geometric nodes with ratio λ\lambda. It has two characteristic polynomials,

P+(z)=∑k=0n(ak+ck)zkandP−(z)=∑k=0n(ak−ck)zk(7)P_{+}(z)=\sum_{k=0}^{n}(a_k+c_k)z^k \qquad\text{and}\qquad P_{-}(z)=\sum_{k=0}^{n}(a_k-c_k)z^k \tag*{(7)}

As an example, suppose D(f,h)=5f(−2h)+3f(−h)−5f(0)−7f(2h)+f(23h)D(f,h)=5f(-2h)+3f(-h)-5f(0)-7f(2h)+f(2^3h). Then λ=2\lambda=2, the characteristic polynomials are P+(z)=z3−2z+3P_{+}(z)=z^3-2z+3, and P−(z)=z3−12z−3P_{-}(z)=z^3-12z-3.

In the case of geometric nodes, the MZ-property has been completely characterized in terms of the roots of the characteristic polynomials.

Theorem 10. (Fejzić [6]) A difference with geometric nodes has the MZ-property if and only if neither of its characteristic polynomials has a complex root zz in the annulus 1<∣z∣≤∣λ∣1<|z|\leq|\lambda|.

In the course of that investigation, it was necessary to allow for complex coefficients, explaining why we retain this convention, even if our main interest is real-valued functions.

Any two-term difference can be written with geometric nodes after scaling. Therefore, Theorem 10 completely characterizes the two-term case:

Corollary 11. (Fejzić [6]) Let D(f,h)=af(h)+f(qh)D(f,h)=af(h)+f(qh) be a two-term difference with 1≤∣q∣1\leq|q|.

Case 1: If 1=∣q∣1=|q|, i.e. q=−1q=-1 then D(f,h)D(f,h) has the MZ-property if and only if a≠±1a\ne\pm1.

Case 2: If 1<∣q∣1<|q| then D(f,h)D(f,h) has the MZ-property if and only if aa is not in the annulus 1<∣a∣≤∣q∣1<|a|\leq|q|.

Proof. Case 1: This can be deduced from Theorem 10, but is also seen directly. Suppose D(f,h)=af(h)+f(−h)=o(h)D(f,h)=af(h)+f(-h)=o(h). Then aD(f,h)−D(f,−h)=(a2−1)f(h)=o(h)aD(f,h)-D(f,-h)=(a^2-1)f(h)=o(h). If a2≠1a^2\ne1 then f(h)=o(h)f(h)=o(h). For counterexamples, if a=1a=1 use f(h)=hf(h)=h and if a=−1a=-1 use f(h)=∣h∣f(h)=|h|.

Case 2: Let λ=∣q∣\lambda=|q|. Then, D(f,h)=af(λ0h)+f(λ1h)D(f,h)=af(\lambda^0h)+f(\lambda^1h) or D(f,h)=af(λ0h)+f(−λ1h)D(f,h)=af(\lambda^0h)+f(-\lambda^1h). In any event, the two characteristic polynomials are a+za+z and a−za-z, with roots z=±az=\pm a. By Theorem 10, the MZ-property fails if and only if 1<∣a∣≤∣q∣1<|a|\leq|q|. □

For rational three-term differences in standard form (2), if

log⁡∣r∣log⁡∣s∣=kj,\frac{\log|r|}{\log|s|}=\frac{k}{j},

for integers k≥0k\geq0 and j≥1j\geq1 then letting λ=∣s∣1/j\lambda=|s|^{1/j} the difference can be written with geometric nodes as D(f,h)=af(h)+bf(sgn⁡(r)λkh)+f(sgn⁡(s)λjh)D(f,h)=af(h)+bf(\operatorname{sgn}(r)\lambda^kh)+f(\operatorname{sgn}(s)\lambda^jh), where ∣λ∣>1|\lambda|>1. Therefore, Theorem 10 applies, giving the following corollary.

Corollary 12. (Fejzić [6]) Let D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh) be a rational difference with 1≤∣r∣≤∣s∣1\leq|r|\leq|s|, and let j>0j>0, k≥0k\geq0 be integers such that log⁡∣r∣log⁡∣s∣=kj\frac{\log|r|}{\log|s|}=\frac{k}{j}. Then DD has the MZ-property if and only if the trinomials zj+bzk+az^j+bz^k+a and sgn⁡(s)zj+sgn⁡(r)bzk+a\operatorname{sgn}(s)z^j+\operatorname{sgn}(r)bz^k+a have no roots in the annulus 1<∣z∣≤∣s∣1/j1<|z|\leq|s|^{1/j}.

The next examples illustrate important differences between the rational and irrational cases. Examples 13 and 14 illustrate that the MZ-property in the rational case can depend on the plus and minus signs attached to coefficients. Examples 13 and 15 illustrate that it can also depend on plus and minus signs attached to the nodes. This is in contrast to the irrational case, Theorem 2, where the MZ-property only depends on absolute values.

Example 13. The difference D(f,h)=f(8h)+4f(2h)+3f(h)D(f,h)=f(8h)+4f(2h)+3f(h) has only one distinct trinomial x3+4x+3x^3+4x+3 whose roots are outside the annulus 1<∣z∣≤21<|z|\leq2, so it has the MZ-property.

Example 14. The difference D(f,h)=f(8h)−4f(2h)+3f(h)D(f,h)=f(8h)-4f(2h)+3f(h) has only one distinct trinomial x3−4x+3x^3-4x+3. It contains a root, zz, where 1<∣z∣≤21<|z|\leq2, so it does not have the MZ-property. Indeed, here is a specific counterexample. Let f(h)=znf(h)=z^n if h=2nh=2^n and n∈Zn\in\mathbb{Z}. Let f(h)=0f(h)=0 elsewhere.

Example 15. The difference D(f,h)=f(8h)+4f(−2h)+3f(h)D(f,h)=f(8h)+4f(-2h)+3f(h) has two distinct trinomials, x3+4x+3x^3+4x+3 and x3−4x+3x^3-4x+3 and since z=13−12z=\frac{\sqrt{13}-1}{2} is a root of the second trinomial with 1<∣z∣≤21<|z|\leq2, it fails MZ-property. Indeed, one can verify that f(h)=±znf(h)=\pm z^n for h=±2nh=\pm2^n n∈Zn\in\mathbb{Z} and 00 elsewhere is a counterexample.

Accordingly, the main focus of the present paper is the case of irrational three-term differences. Together, Theorem 2 and Corollary 12 completely characterize the MZ-property for three-term differences. But there is a big difference in these results. Theorem 2 solves the irrational case with a direct computation on the coefficients and nodes. Corollary 12 solves the rational case, using the location of the complex roots of polynomial equations. We do not know of a direct characterization for the rational case. In particular, Theorem 2 does not hold for the rational case. This is seen in Example 13 where we have a=3a=3, b=4b=4, r=2r=2, s=8s=8. None of the three conditions of Theorem 2 hold, and yet D(f,h)D(f,h) has the MZ-property. Nevertheless, one direction of Theorem 2 does hold in the rational case.

Theorem 16. Let D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh) be a rational difference with distinct nodes, 1,r,s1,r,s where 1≤∣r∣≤∣s∣1\leq|r|\leq|s|. Then DD has the MZ-property if at least one of the following conditions holds

  • ∣a∣+∣b∣≤1|a|+|b|\leq1

  • ∣br∣+∣s∣<∣a∣|br|+|s|<|a|

  • ∣a∣+1≤∣b∣|a|+1\leq|b| and ∣s∣+∣a∣<∣br∣|s|+|a|<|br|.

We end this section with an application of Theorem 2 to an open problem posed in [5]. Consider the second-order Riemann difference

D(f,h)=f(5h)−2f(2h)+f(−h).D(f,h)=f(5h)-2f(2h)+f(-h).

We show that

D(f,h)=o(h2)andf(h)=o(h)D(f,h)=o(h^2) \quad\text{and}\quad f(h)=o(h)

imply

f(h)=o(h2).f(h)=o(h^2).

By Remark 8, it suffices to show that the difference

D1(g,h)=5g(5h)−4g(2h)−g(−h)D_1(g,h)=5g(5h)-4g(2h)-g(-h)

has the MZ-property. Indeed, using the notation of Theorem 2, the coefficients a=−1a=-1, b=−4b=-4, and c=5c=5 satisfy the first condition of the theorem,

∣a∣+∣b∣≤∣c∣.|a|+|b|\leq|c|.

The main goal now, is to establish Theorems 2 and 16. The rest of this paper is organized as follows. In Section 3 we establish Theorems 2 and 16 for the edge cases r=−1r=-1 and r=−sr=-s. This is a mere convenience, allowing us to assume strict inequalities 1<∣r∣<∣s∣1<|r|<|s| for the remainder of the paper. In Sections 4, 5, and 6 we prove the MZ-property for the three respective conditions stated in Theorem 2 and Theorem 16. In Section 7 we show that if none of the three conditions hold, and D(f,h)D(f,h) is irrational, then the MZ-property fails.

The edge cases r=−1r=-1 and r=−sr=-s

In this section we consider the difference

D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh)

where 1,r,s1,r,s are distinct, 1≤∣r∣≤∣s∣1\leq|r|\leq|s|, and one of the inequalities in not strict. That is, either r=−1r=-1 or r=−sr=-s.

The cases r=−1r=-1 and r=−sr=-s do not occur in the irrational case. Indeed, if r=−1r=-1 then log⁡∣r∣log⁡∣s∣=0\frac{\log|r|}{\log|s|}=0 and if r=−sr=-s then log⁡∣r∣log⁡∣s∣=1\frac{\log|r|}{\log|s|}=1. In both cases, D(f,h)D(f,h) is rational. The goal of this section, then, is to establish Theorem 16 for these two cases.

Let us first consider r=−1r=-1, so that

D(f,h)=af(h)+bf(−h)+f(sh).(8)D(f,h)=af(h)+bf(-h)+f(sh). \tag*{(8)}

with ∣s∣>1|s|>1.

Corollary 17. A difference (8) has the MZ-property if and only if neither z=∣a+b∣z=|a+b| nor z=∣a−b∣z=|a-b| is in the interval 1<∣z∣≤∣s∣1<|z|\leq|s|.

Proof. Apply Corollary 12, with k=0k=0 and j=1j=1. The two characteristic polynomials are z+(b+a)z+(b+a) and sgn⁡(s)z+(a−b)\operatorname{sgn}(s)z+(a-b). The absolute values of the respective roots are ∣z∣=∣a+b∣|z|=|a+b| and ∣z∣=∣a−b∣|z|=|a-b|. The result immediately follows from Corollary 12.

The following corollary is Theorem 16 when r=−1r=-1.

Corollary 18. If any of the following conditions hold

  • ∣a∣+∣b∣≤1|a|+|b|\leq1

  • ∣b∣+∣s∣<∣a∣|b|+|s|<|a|

  • ∣a∣+1≤∣b∣|a|+1\leq|b| and ∣s∣+∣a∣<∣b∣|s|+|a|<|b|,

then the difference (8) has the MZ-property.

Proof. Note that

max⁡{∣a∣−∣b∣,∣b∣−∣a∣}≤max⁡{∣a−b∣,∣a+b∣}≤∣a∣+∣b∣.\max\{|a|-|b|,|b|-|a|\}\leq\max\{|a-b|,|a+b|\}\leq|a|+|b|.

So, if ∣a∣+∣b∣≤1|a|+|b|\leq1 then ∣a+b∣≤1|a+b|\leq1 and ∣a−b∣≤1|a-b|\leq1. If ∣b∣+∣s∣<∣a∣|b|+|s|<|a|, then ∣s∣<∣a∣−∣b∣|s|<|a|-|b|, so ∣s∣<∣a+b∣|s|<|a+b| and ∣s∣<∣a−b∣|s|<|a-b|. Finally, if ∣s∣+∣a∣<∣b∣|s|+|a|<|b| then ∣s∣<∣b∣−∣a∣|s|<|b|-|a|, so ∣s∣<∣a+b∣|s|<|a+b| and ∣s∣<∣a−b∣|s|<|a-b|. In all three cases, by Corollary 17, D(f,h)D(f,h) has the MZ-property. □

Next, consider r=−sr=-s, so that

D(f,h)=af(h)+bf(−sh)+f(sh).(9)D(f,h)=af(h)+bf(-sh)+f(sh). \tag*{(9)}

with ∣s∣>1|s|>1.

Corollary 19. A difference (9) has the MZ-property if and only if ∣a∣|a| is not in either of the intervals:

  • (∣1+b∣,∣s∣∣1+b∣](|1+b|,|s||1+b|]

  • (∣1−b∣,∣s∣∣1−b∣](|1-b|,|s||1-b|]

Proof. Apply Corollary 12, with k=1k=1 and j=1j=1. The two characteristic polynomials are

P+(z)=(1+b)z+aandP−(z)=(sgn⁡(s)+sgn⁡(−s)b)z+a.P_{+}(z)=(1+b)z+a \qquad\text{and}\qquad P_{-}(z)=(\operatorname{sgn}(s)+\operatorname{sgn}(-s)b)z+a.

Since the parameters are nonzero, if b=−1b=-1 then P+P_{+} has no root. Otherwise, it has a single root with absolute value ∣z∣=∣a∣∣1+b∣|z|=\frac{|a|}{|1+b|}. In other words, P+P_{+} has no root, zz, in the annulus 1<∣z∣≤∣s∣1<|z|\leq|s| if and only if ∣a∣|a| is not in the interval (∣1+b∣,∣s∣∣1+b∣](|1+b|,|s||1+b|]. Similarly, P−P_{-} has no root with

1<∣z∣≤∣s∣1 < |z| \le|s| if and only if ∣a∣|a| is not in the interval (∣1−b∣,∣s∣∣1−b∣](|1-b|,|s||1-b|]. The result immediately follows by Corollary 12.

The following corollary is Theorem 16 when r=−sr=-s.

Corollary 20. If any of the following conditions hold

  • ∣a∣+∣b∣≤1|a|+|b|\le1

  • ∣bs∣+∣s∣<∣a∣|bs|+|s|<|a|

  • ∣a∣+1≤∣b∣|a|+1\le|b| and ∣s∣+∣a∣<∣bs∣|s|+|a|<|bs|,

then the difference (9) has the MZ-property.

Proof. If ∣a∣+∣b∣≤1|a|+|b|\le1 then ∣a∣≤1−∣b∣≤min⁡{∣1+b∣,∣1−b∣}|a|\le1-|b|\le\min\{|1+b|,|1-b|\}. If ∣bs∣+∣s∣<∣a∣|bs|+|s|<|a| then ∣s∣max⁡{∣1+b∣,∣1−b∣}<∣a∣|s|\max\{|1+b|,|1-b|\}<|a|. If ∣a∣+1≤∣b∣|a|+1\le|b| then ∣a∣≤∣b∣−1≤min⁡{∣1+b∣,∣1−b∣}|a|\le|b|-1\le\min\{|1+b|,|1-b|\}. In all three cases, the MZ-property follows by Corollary 19.

It remains to show Theorems 2 and 16 for the case 1<∣r∣<∣s∣1<|r|<|s|.

The case ∣a∣+∣b∣≤1|a|+|b|\le1

In this section we consider

D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh)

where 1<∣r∣<∣s∣1<|r|<|s| and ∣a∣+∣b∣≤1|a|+|b|\le1.

In the rational case, suppose log⁡∣r∣log⁡∣s∣=kj\frac{\log|r|}{\log|s|}=\frac{k}{j} for some positive integers, j,kj,k. Then, this result follows from Corollary 12. In fact, no polynomial P(z)=±zj±bzk±aP(z)=\pm z^j\pm bz^k\pm a has a root with ∣z∣>1|z|>1. To see this, let ∣z∣>1|z|>1. Then,

∣P(z)∣≥∣zj∣−∣bzk∣−∣a∣>∣z∣j(1−∣b∣)−∣a∣>1−∣b∣−∣a∣≥0.|P(z)|\ge|z^j|-|bz^k|-|a|>|z|^j(1-|b|)-|a|>1-|b|-|a|\ge0.

Therefore, P(z)P(z) has no zeros with ∣z∣>1|z|>1. By Corollary 12, D(f,h)D(f,h) has the MZ-property.

The proof below works directly and does not require any rational relation between log⁡∣r∣\log|r| and log⁡∣s∣\log|s|.

Theorem 21. Let 1<∣r∣<∣s∣1<|r|<|s|, and let a,b∈Ca,b\in\mathbb{C} satisfy

∣a∣+∣b∣≤1.|a|+|b|\le1.

Then

Df(h)=af(h)+bf(rh)+f(sh)D_f(h)=af(h)+bf(rh)+f(sh)

has the MZ-property

Proof. Assume Df(h)=o(h)D_f(h)=o(h) and f(h)=o(1)f(h)=o(1) and fix ε>0\varepsilon>0. Then, there exists δ>0\delta>0 such that for all tt with 0<∣t∣<δ0<|t|<\delta,

∣af(t)+bf(rt)+f(st)∣≤ε∣t∣<ε∣st∣.|af(t)+bf(rt)+f(st)|\le\varepsilon|t|<\varepsilon|st|.

Then,

∣f(st)∣<∣a∣∣f(t)∣+∣b∣∣f(rt)∣+ε∣st∣.(10)|f(st)|<|a||f(t)|+|b||f(rt)|+\varepsilon|st|. \tag*{(10)}

It follows that either

∣f(t)∣>∣f(st)∣−ε∣st∣,or∣f(rt)∣>∣f(st)∣−ε∣st∣.|f(t)|>|f(st)|-\varepsilon|st|,\qquad\text{or}\qquad|f(rt)|>|f(st)|-\varepsilon|st|.

Indeed, otherwise since ∣a∣+∣b∣≤1|a|+|b|\le1, we would get a contradiction to (10). Since also ∣ts∣<δ\left|\frac{t}{s}\right|<\delta, either

∣f(ts)∣>∣f(t)∣−ε∣t∣,or∣f(rts)∣>∣f(t)∣−ε∣t∣(11)\left|f\left(\frac{t}{s}\right)\right|>|f(t)|-\varepsilon|t|,\qquad\text{or}\qquad\left|f\left(\frac{rt}{s}\right)\right|>|f(t)|-\varepsilon|t| \tag*{(11)}

Now fix hh with ∣h∣<δ|h| < \delta. Repeatedly applying 11 define a sequence hmh_m recursively by

h0=h,h_0 = h,

and choosing

hm=1shm−1orhm=rshm−1h_m = \frac{1}{s}h_{m-1}\quad\text{or}\quad h_m = \frac{r}{s}h_{m-1}

so that

∣f(hm)∣>∣f(hm−1)∣−ε∣hm−1∣.(12)|f(h_m)| > |f(h_{m-1})|-\varepsilon|h_{m-1}|. \tag*{(12)}

In either case, since 1<∣r∣<∣s∣1 < |r| < |s|,

0<∣hm∣≤∣rs∣∣hm−1∣.(13)0 < |h_m| \le\left|\frac{r}{s}\right||h_{m-1}|. \tag*{(13)}

where 0<∣rs∣<10 < \left|\frac{r}{s}\right| < 1. Therefore, ∣hm∣|h_m| strictly decreases to 0, justifying the continued use of (11)

Let cc be the sum of the geometric series,

c=∑j=0∞∣rs∣j.c = \sum_{j=0}^{\infty}\left|\frac{r}{s}\right|^j.

Iterating (13) gives

0<∣hj∣≤∣rs∣j∣h∣0 < |h_j| \le\left|\frac{r}{s}\right|^j |h|

for each non-negative integer jj and then iterating (12) gives

∣f(hm)∣>∣f(h)∣−ε∑j=0m−1∣hj∣>∣f(h)∣−εc∣h∣.|f(h_m)| > |f(h)|-\varepsilon\sum_{j=0}^{m-1}|h_j| > |f(h)|-\varepsilon c|h|.

Thus

∣f(h)∣<∣f(hm)∣+εc∣h∣.|f(h)| < |f(h_m)|+\varepsilon c|h|.

Since hm→0h_m \to0 and f(h)=o(1)f(h)=o(1), we have f(hm)→0f(h_m)\to0. Therefore,

∣f(h)∣≤εc∣h∣.|f(h)| \le\varepsilon c|h|.

Because ε>0\varepsilon> 0 is arbitrary, it follows that

f(h)=o(h).f(h)=o(h).

The case ∣br∣+∣s∣<∣a∣|br|+|s|<|a|

In this section we consider

D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh)

where 1<∣r∣<∣s∣1 < |r| < |s|. We will show that if ∣br∣+∣s∣<∣a∣|br|+|s|<|a| then DD has the MZ-property.

In the rational case, suppose

log⁡∣r∣log⁡∣s∣=kj\frac{\log|r|}{\log|s|}=\frac{k}{j}

for some positive integers j,kj,k. Then, this result follows from Corollary 12. In fact, no polynomial P(z)=±zj±bzk±aP(z)=\pm z^j\pm bz^k\pm a has a root zz with ∣z∣≤∣r∣1/k|z|\le|r|^{1/k}. To see this, let λ=∣r∣1/k\lambda=|r|^{1/k}. We have λj=∣s∣\lambda^j=|s| and λk=∣r∣\lambda^k=|r|. If zz is a zero of P(z)P(z) then ∣±zj±bzk∣=∣a∣|\pm z^j\pm bz^k|=|a|. But when ∣z∣≤λ|z|\le\lambda,

∣±zj±bzk∣≤λj+∣b∣λk=∣s∣+∣b∣∣r∣<∣a∣.|\pm z^j\pm bz^k|\le\lambda^j+|b|\lambda^k=|s|+|b||r|<|a|.

So, P(z)P(z) has no zeros with ∣z∣≤λ|z|\le\lambda and by Corollary 12, D(f,h)D(f,h) has the MZ-property.

The following proof works without assuming any rational relation between log⁡∣s∣\log|s| and log⁡∣r∣\log|r|. Furthermore, instead of f(h)=o(1)f(h)=o(1) it requires only f(h)=O(1)f(h)=O(1).

Theorem 22. Let 1<∣r∣<∣s∣1 < |r| < |s|, and ∣br∣+∣s∣<∣a∣|br| + |s| < |a|, and

Df(h)=af(h)+bf(rh)+f(sh).Df(h) = af(h) + bf(rh) + f(sh).

If

Df(h)=o(h)andf(h)=O(1)Df(h) = o(h) \qquad\text{and} \qquad f(h) = O(1)

Then

f(h)=o(h).f(h) = o(h).

The proof of this theorem is based on repeated applications of the following lemma.

Lemma 23. Let

Df(h)=af(h)+bf(rh)+f(sh).Df(h) = af(h) + bf(rh) + f(sh).

with 1<∣r∣<∣s∣1 < |r| < |s|, and ∣br∣+∣s∣<∣a∣|br| + |s| < |a|. Let

t=log⁡∣r∣log⁡∣s∣andK=∑k=0∞(∣br∣+∣s∣∣a∣)k.t = \frac{\log|r|}{\log|s|} \qquad\text{and} \qquad K = \sum_{k=0}^{\infty} \left(\frac{|br| + |s|}{|a|}\right)^k.

Suppose for some α≥0\alpha\ge0

Df(h)=o(h)andf(h)=O(hα).Df(h) = o(h) \qquad\text{and} \qquad f(h) = O(h^\alpha).

Then for each ϵ>0\epsilon> 0 there is a δ>0\delta> 0 and L>0L > 0 such that whenever ∣h∣<δ|h| < \delta,

∣f(h)∣<Kϵ∣h∣+L∣h∣α+t.|f(h)| < K\epsilon|h| + L|h|^{\alpha+t}.

We first prove the theorem from the lemma:

Proof. Let 1<∣r∣<∣s∣1 < |r| < |s|, and ∣br∣+∣s∣<∣a∣|br| + |s| < |a|, and

Df(h)=af(h)+bf(rh)+f(sh).Df(h) = af(h) + bf(rh) + f(sh).

Suppose

Df(h)=o(h)andf(h)=O(1).Df(h) = o(h) \qquad\text{and} \qquad f(h) = O(1).

We will show that f(h)=o(h)f(h) = o(h).

Since f(h)=O(hα)f(h) = O(h^\alpha) for α=0\alpha= 0. We apply the lemma with ϵ=1\epsilon= 1 and get positive numbers δ1\delta_1 and L1L_1 such that whenever ∣h∣<δ1|h| < \delta_1,

∣f(h)∣<K∣h∣+L1∣h∣t.|f(h)| < K|h| + L_1|h|^t.

By shrinking δ1\delta_1, if necessary, we may assume that ∣h∣<1|h| < 1. Since t<1t < 1, ∣h∣<∣h∣t|h| < |h|^t. So, for all ∣h∣<δ1|h| < \delta_1,

∣f(h)∣<(K+L1)∣h∣t.|f(h)| < (K + L_1)|h|^t.

This implies that f(h)=O(ht)f(h) = O(h^t).

We apply the lemma a second time with α=t\alpha= t and ϵ=1\epsilon= 1. We get positive numbers δ2\delta_2 and L2L_2 such that whenever ∣h∣<δ2|h| < \delta_2,

∣f(h)∣<K∣h∣+L2∣h∣2t.|f(h)| < K|h| + L_2|h|^{2t}.

If 2t≤12t \le1 then again, by shrinking δ2\delta_2 if necessary, we get that for all ∣h∣<δ2|h| < \delta_2,

∣f(h)∣<(K+L2)∣h∣2t,|f(h)| < (K + L_2)|h|^{2t},

and therefore, f(h)=O(h2t)f(h) = O(h^{2t}).

This process continues as long as α+t≤1\alpha+ t \le1. With each application of the lemma, the value of α\alpha increases by the constant tt. Eventually, after a finite number of applications, we have f(h)=O(hα)f(h) = O(h^\alpha) with α≤1\alpha\le1 but α+t>1\alpha+ t > 1. At that point, we apply the lemma one final time, but this time we fix some arbitrary ϵ>0\epsilon> 0. By the lemma, there is a δ>0\delta> 0 and L>0L > 0 such that whenever ∣h∣<δ|h| < \delta,

∣f(h)∣<Kϵ∣h∣+L∣h∣α+t.|f(h)| < K\epsilon|h| + L|h|^{\alpha+t}.

Since this time, α+t>1\alpha+ t > 1, the first term dominates as h→0h \to0. By shrinking δ\delta if necessary, we can assume ∣h∣|h| is so small that L∣h∣α+t<Kϵ∣h∣L|h|^{\alpha+t} < K\epsilon|h|. Then we get

∣f(h)∣<2Kϵ∣h∣.|f(h)| < 2K\epsilon|h|.

Since ϵ\epsilon is arbitrary, it follows that f(h)=o(h)f(h) = o(h). □

We now prove the lemma:

Proof. Let

Df(h)=af(h)+bf(rh)+f(sh).Df(h) = af(h) + bf(rh) + f(sh).

with 1<∣r∣<∣s∣1 < |r| < |s|, and ∣br∣+∣s∣<∣a∣|br| + |s| < |a|. Let

t=log⁡∣r∣log⁡∣s∣andK=∑k=0∞(∣br∣+∣s∣∣a∣)k.t = \frac{\log|r|}{\log|s|} \qquad\text{and}\qquad K = \sum_{k=0}^{\infty}\left(\frac{|br|+|s|}{|a|}\right)^k.

Fix α≥0\alpha\ge0 such that

Df(h)=o(h)andf(h)=O(hα).Df(h) = o(h) \qquad\text{and}\qquad f(h) = O(h^\alpha).

Fix ϵ>0\epsilon> 0. We must find δ>0\delta> 0 and L>0L > 0 such that whenever ∣h∣<δ|h| < \delta,

∣f(h)∣<Kϵ∣h∣+L∣h∣α+t.|f(h)| < K\epsilon|h| + L|h|^{\alpha+t}.

Using that f(h)=O(hα)f(h) = O(h^\alpha), choose δ>0\delta> 0 and B>0B > 0, such that

∣f(h)∣≤B∣h∣αfor all ∣h∣<∣s∣δ.(14)|f(h)| \le B|h|^\alpha\qquad\text{for all } |h| < |s|\delta. \tag*{(14)}

Since

af(h)+bf(rh)+f(sh)=o(h),af(h) + bf(rh) + f(sh) = o(h),

shrink δ\delta, if necessary, so that for all ∣h∣<δ|h| < \delta,

∣f(h)∣≤1∣a∣∣f(sh)∣+∣b∣∣a∣∣f(rh)∣+ϵ∣h∣.(15)|f(h)| \le\frac{1}{|a|}|f(sh)| + \frac{|b|}{|a|}|f(rh)| + \epsilon|h|. \tag*{(15)}

If ∣sh∣<δ|sh| < \delta (and therefore ∣rh∣<δ|rh| < \delta), then applying the same estimate to rhrh and shsh gives

∣f(rh)∣≤1∣a∣∣f(rsh)∣+∣b∣∣a∣∣f(r2h)∣+ϵ∣rh∣,|f(rh)| \le\frac{1}{|a|}|f(rsh)| + \frac{|b|}{|a|}|f(r^2h)| + \epsilon|rh|,

and

∣f(sh)∣≤1∣a∣∣f(s2h)∣+∣b∣∣a∣∣f(rsh)∣+ϵ∣sh∣.|f(sh)| \le\frac{1}{|a|}|f(s^2h)| + \frac{|b|}{|a|}|f(rsh)| + \epsilon|sh|.

Substituting these into (15), we obtain

∣f(h)∣≤1∣a∣2(∣f(s2h)∣+2∣b∣∣f(rsh)∣+∣b∣2∣f(r2h)∣)+ϵ∣h∣(1+∣br∣+∣s∣∣a∣).(16)|f(h)| \le\frac{1}{|a|^2}\left(|f(s^2h)| + 2|b||f(rsh)| + |b|^2|f(r^2h)|\right) + \epsilon|h|\left(1+\frac{|br|+|s|}{|a|}\right). \tag*{(16)}

If also ∣s2h∣<δ|s^2h| < \delta, then applying (15) to ∣r2h∣|r^2h|, ∣rsh∣|rsh|, ∣s2h∣|s^2h| we obtain

∣f(r2h)∣≤1∣a∣∣f(r2sh)∣+∣b∣∣a∣∣f(r3h)∣+ϵ∣r2h∣,|f(r^2h)| \le\frac{1}{|a|}|f(r^2sh)| + \frac{|b|}{|a|}|f(r^3h)| + \epsilon|r^2h|,
∣f(rsh)∣≤1∣a∣∣f(rs2h)∣+∣b∣∣a∣∣f(r2sh)∣+ϵ∣rsh∣,|f(rsh)| \le\frac{1}{|a|}|f(rs^2h)| + \frac{|b|}{|a|}|f(r^2sh)| + \epsilon|rsh|,
∣f(s2h)∣≤1∣a∣∣f(s3h)∣+∣b∣∣a∣∣f(rs2h)∣+ϵ∣s2h∣.|f(s^2h)| \le\frac{1}{|a|}|f(s^3h)| + \frac{|b|}{|a|}|f(rs^2h)| + \epsilon|s^2h|.

Substituting into (16), we obtain

∣f(h)∣≤1∣a∣3(∣f(s3h)∣+3∣bf(rs2h)∣+3∣b2f(r2sh)∣+∣b3f(r3h)∣)+ϵ∣h∣(1+∣br∣+∣s∣∣a∣+(∣br∣+∣s∣∣a∣)2).\lvert f(h)\rvert\le\frac{1}{\lvert a\rvert^3}\left(\lvert f(s^3h)\rvert+3\lvert b f(rs^2h)\rvert+3\lvert b^2 f(r^2sh)\rvert+\lvert b^3 f(r^3h)\rvert\right) +\epsilon\lvert h\rvert\left(1+\frac{\lvert br\rvert+\lvert s\rvert}{\lvert a\rvert}+\left(\frac{\lvert br\rvert+\lvert s\rvert}{\lvert a\rvert}\right)^2\right).

Continuing inductively, if ∣sNh∣<δ\lvert s^N h\rvert<\delta, then

∣f(h)∣≤1∣a∣N+1∑k=0N+1(N+1k)∣bkf(rksN+1−kh)∣+ϵ∣h∣∑k=0N(∣br∣+∣s∣∣a∣)k.(17)\lvert f(h)\rvert\le\frac{1}{\lvert a\rvert^{N+1}}\sum_{k=0}^{N+1}\binom{N+1}{k}\lvert b^k f(r^k s^{N+1-k}h)\rvert+\epsilon\lvert h\rvert\sum_{k=0}^{N}\left(\frac{\lvert br\rvert+\lvert s\rvert}{\lvert a\rvert}\right)^k . \tag*{(17)}

Furthermore, since

∣rksN+1−kh∣≤∣sN+1h∣<∣s∣δ,\lvert r^k s^{N+1-k}h\rvert\leq\lvert s^{N+1}h\rvert<\lvert s\rvert\delta,

it follows from (14) that

∣f(rksN+1−kh)∣≤B∣h∣α.\lvert f(r^k s^{N+1-k}h)\rvert\leq B\lvert h\rvert^\alpha.

Hence

∣f(h)∣≤B∣h∣α∣a∣N+1∑k=0N+1(N+1k)∣bk∣+ϵ∣h∣∑k=0N(∣br∣+∣s∣∣a∣)k.\lvert f(h)\rvert\leq\frac{B\lvert h\rvert^\alpha}{\lvert a\rvert^{N+1}}\sum_{k=0}^{N+1}\binom{N+1}{k}\lvert b^k\rvert+\epsilon\lvert h\rvert\sum_{k=0}^{N}\left(\frac{\lvert br\rvert+\lvert s\rvert}{\lvert a\rvert}\right)^k .

Applying the definition of KK and the binomial theorem,

∑k=0N+1(N+1k)∣bk∣=(1+∣b∣)N+1,\sum_{k=0}^{N+1}\binom{N+1}{k}\lvert b^k\rvert=(1+\lvert b\rvert)^{N+1},

we get

∣f(h)∣<B∣h∣α(1+∣b∣∣a∣)N+1+Kϵ∣h∣.(18)\lvert f(h)\rvert<B\lvert h\rvert^\alpha\left(\frac{1+\lvert b\rvert}{\lvert a\rvert}\right)^{N+1}+K\epsilon\lvert h\rvert. \tag*{(18)}

Now fix NN such that

∣sNh∣<δ≤∣sN+1h∣.\lvert s^N h\rvert<\delta\leq\lvert s^{N+1}h\rvert.

Then,

(1+∣b∣∣a∣)N+1<(∣s∣+∣br∣∣ar∣)N+1<(∣a∣∣ar∣)N+1=1∣r∣N+1=1∣s∣t(N+1).\left(\frac{1+\lvert b\rvert}{\lvert a\rvert}\right)^{N+1} < \left(\frac{\lvert s\rvert+\lvert br\rvert}{\lvert ar\rvert}\right)^{N+1} < \left(\frac{\lvert a\rvert}{\lvert ar\rvert}\right)^{N+1} = \frac{1}{\lvert r\rvert^{N+1}} = \frac{1}{\lvert s\rvert^{t(N+1)}}.

Since

δ≤∣s∣N+1∣h∣,\delta\leq\lvert s\rvert^{N+1}\lvert h\rvert,

it follows that

1∣s∣t(N+1)≤(∣h∣δ)t,\frac{1}{\lvert s\rvert^{t(N+1)}}\leq\left(\frac{\lvert h\rvert}{\delta}\right)^t,

yielding

(1+∣b∣∣a∣)N+1<(∣h∣δ)t.\left(\frac{1+\lvert b\rvert}{\lvert a\rvert}\right)^{N+1}<\left(\frac{\lvert h\rvert}{\delta}\right)^t.

Substituting into (18) and letting L=BδtL=\frac{B}{\delta^t}, we obtain

∣f(h)∣≤B∣h∣α(∣h∣δ)t+Kϵ∣h∣=L∣h∣α+t+Kϵ∣h∣\begin{aligned} |f(h)| &\le B|h|^\alpha\left(\frac{|h|}{\delta}\right)^t+K\epsilon|h| \\ &=L|h|^{\alpha+t}+K\epsilon|h| \end{aligned}

The case ∣a∣+1≤∣b∣|a|+1\le|b| and ∣s∣+∣a∣<∣br∣|s|+|a|<|br|

In this section we prove the following theorem.

Theorem 24. Suppose that

1<∣r∣<∣s∣1<|r|<|s| \quad and \quad ∣a∣+1≤∣b∣|a|+1\le|b| \quad and \quad ∣s∣+∣a∣<∣br∣|s|+|a|<|br|.

Then the difference

D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh)

has the MZ-property.

This time, we do not have a proof that handles both the rational and irrational cases simultaneously. We will therefore split them. First, the rational case:

Theorem 25. Suppose that

1<∣r∣<∣s∣1<|r|<|s| \quad and \quad ∣a∣+1≤∣b∣|a|+1\le|b| \quad and \quad ∣s∣+∣a∣<∣br∣|s|+|a|<|br|.

Suppose also that log⁡∣r∣log⁡∣s∣∈Q\frac{\log|r|}{\log|s|}\in\mathbb{Q}. Then the difference

D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh)

has the MZ-property.

Proof. Let log⁡∣r∣log⁡∣s∣=kj\frac{\log|r|}{\log|s|}=\frac{k}{j} where k,jk,j are positive integers. Let λ=∣s∣1/j\lambda=|s|^{1/j}, so that λj=∣s∣\lambda^j=|s| and λk=∣r∣\lambda^k=|r|. By Corollary 12, it is enough to show that no trinomial of the form

P(z)=±zj±bzk±aP(z)=\pm z^j\pm bz^k\pm a

can have zeros in the annulus 1<∣z∣≤λ1<|z|\le\lambda.

If P(z)=0P(z)=0 then ∣bzk∣=∣±zj±a∣≤∣z∣j+∣a∣|bz^k|=|\pm z^j\pm a|\le|z|^j+|a|. In other words, φ(x)=xj−∣b∣xk+∣a∣≥0\varphi(x)=x^j-|b|x^k+|a|\ge0 when x=∣z∣x=|z|. We will show this can’t happen for x∈(1,λ]x\in(1,\lambda]. By Descartes Rule of Signs, φ\varphi has at most two positive real roots. To estimate them,

φ(0)=∣a∣>0.\varphi(0)=|a|>0.
φ(1)=1−∣b∣+∣a∣≤0\varphi(1)=1-|b|+|a|\le0
φ(λ)=∣s∣−∣b∣∣r∣+∣a∣<0.\varphi(\lambda)=|s|-|b||r|+|a|<0.

and for large enough xx,

φ(x)>0.\varphi(x)>0.

So, φ(x)\varphi(x) has a root in (0,1](0,1] and another root in (λ,∞)(\lambda,\infty) and is strictly negative in the region (1,λ](1,\lambda].

We now consider the same theorem for the irrational case. To prove this theorem, it will be convenient to use the following lemma. Examples 13 and 14 demonstrate that this next lemma is false in the rational case, explaining why we split into cases.

Lemma 26. Let

D1(f,h)=af(h)+bf(rh)+f(sh)D_1(f,h)=af(h)+bf(rh)+f(sh)

and

D2(f,h)=cf(h)+df(rh)+f(sh)D_2(f,h)=cf(h)+df(rh)+f(sh)

be two irrational differences with the same nodes. Suppose ∣a∣=∣c∣|a|=|c| and ∣b∣=∣d∣|b|=|d|. Then D1D_1 has the MZ-property if and only if D2D_2 has the MZ-property.

Proof. Suppose that D1D_1 does not have the MZ-property. Then for some function f:R→Cf:\mathbb{R}\to\mathbb{C} we have that D1(f,h)=o(h)D_1(f,h)=o(h), f=o(1)f=o(1), but f≠o(h)f\ne o(h). We will finish the proof by finding a function g:R→Cg:\mathbb{R}\to\mathbb{C} such that D2(g,h)=o(h)D_2(g,h)=o(h), g=o(1)g=o(1), but g≠o(h)g\ne o(h).

Since ∣a∣=∣c∣|a|=|c|, choose θ1∈R\theta_1\in\mathbb{R} such that c=aeiθ1c=ae^{i\theta_1}. Since ∣b∣=∣d∣|b|=|d|, choose θ2∈R\theta_2\in\mathbb{R} such that deiθ2=beiθ1de^{i\theta_2}=be^{i\theta_1}.

Define

H={rnsm:n,m∈Z}.H=\{r^ns^m:n,m\in\mathbb{Z}\}.

Since log⁡∣r∣log⁡∣s∣∉Q\frac{\log|r|}{\log|s|}\notin\mathbb{Q}, every element of HH has a unique representation of the form rnsmr^ns^m.

Using that ff is not o(h)o(h), let hk→0h_k\to0 be such that f(hk)≠o(hk)f(h_k)\ne o(h_k). Consider the collection of cosets {hkH}\{h_kH\}, and note that this collection may have duplicates. If h∈⋃hkHh\in\bigcup h_kH, we will define g(h)g(h) as follows. Find the first kk such that h∈hkHh\in h_kH, and express hh as h=hkrnsmh=h_kr^ns^m. Then let g(h)=ei(mθ1+nθ2)f(h)g(h)=e^{i(m\theta_1+n\theta_2)}f(h). If h∉⋃hkHh\notin\bigcup h_kH, simply let g(h)=0g(h)=0.

Since ∣g(h)∣≤∣f(h)∣|g(h)|\le|f(h)|, we get g=o(1)g=o(1). Since ∣g(h)∣=∣f(h)∣|g(h)|=|f(h)|, for h∈⋃hkHh\in\bigcup h_kH, we get g(hk)≠o(hk)g(h_k)\ne o(h_k). Therefore, g≠o(h)g\ne o(h). It remains to show that D2(g,h)=o(h)D_2(g,h)=o(h).

If h∈⋃hkHh\in\bigcup h_kH, then the first kk such that h∈hkHh\in h_kH is also the first kk such that rh∈hkHrh\in h_kH and the first kk such that sh∈hkHsh\in h_kH. Then writing h=hkrnsmh=h_kr^ns^m we have

g(h)=ei(mθ1+nθ2)f(h),g(rh)=eiθ2ei(mθ1+nθ2)f(rh),g(sh)=eiθ1ei(mθ1+nθ2)f(sh).g(h)=e^{i(m\theta_1+n\theta_2)}f(h),\qquad g(rh)=e^{i\theta_2}e^{i(m\theta_1+n\theta_2)}f(rh),\qquad g(sh)=e^{i\theta_1}e^{i(m\theta_1+n\theta_2)}f(sh).

Hence,

cg(h)+dg(rh)+g(sh)=ei(mθ1+nθ2)(cf(h)+deiθ2f(rh)+eiθ1f(sh)).cg(h)+dg(rh)+g(sh)=e^{i(m\theta_1+n\theta_2)}\left(cf(h)+de^{i\theta_2}f(rh)+e^{i\theta_1}f(sh)\right).

By choice of θ1,θ2\theta_1,\theta_2,

cf(h)+deiθ2f(rh)+eiθ1f(sh)=eiθ1(af(h)+bf(rh)+f(sh)),cf(h)+de^{i\theta_2}f(rh)+e^{i\theta_1}f(sh)=e^{i\theta_1}\left(af(h)+bf(rh)+f(sh)\right),

and therefore

D2(g,h)=cg(h)+dg(rh)+g(sh)=eiθ1(af(h)+bf(rh)+f(sh))=eiθ1D1(f,h)=o(h).\begin{aligned} D_2(g,h)&=cg(h)+dg(rh)+g(sh)\\ &=e^{i\theta_1}\left(af(h)+bf(rh)+f(sh)\right)\\ &=e^{i\theta_1}D_1(f,h)=o(h). \end{aligned}

This lemma will be combined with the following theorem. Although it is only needed in the irrational case, the proof of the next theorem goes through for either rational or irrational cases.

Theorem 27. Let c,dc,d be real numbers such that

0<∣c∣<1<∣d∣,0<|c|<1<|d|,

and let c1>0,d1>0c_1>0,d_1>0 be real numbers such that

c1+d1≤1andc1∣c∣+d1∣d∣<1.c_1+d_1\le1\qquad\text{and}\qquad c_1|c|+d_1|d|<1.

Then

D(f,h)=f(h)−c1f(ch)−d1f(dh)D(f,h)=f(h)-c_1f(ch)-d_1f(dh)

has the MZ-property.

Proof. Set

λ:=1−c1∣c∣∣d∣d1andρ:=log⁡λlog⁡∣d∣.\lambda:= \frac{1-c_1|c|}{|d|d_1} \qquad\text{and}\qquad \rho:= \frac{\log\lambda}{\log|d|}.

By hypothesis, λ>1\lambda> 1 and ρ>0\rho> 0.

Suppose D(f,h)=o(h)D(f,h)=o(h) and f(h)=o(1)f(h)=o(1). We will show f(h)=o(h)f(h)=o(h). Suppose on the contrary,

f(h)h↛0.\frac{f(h)}{h} \not\to0.

Replacing ff with −f-f if necessary, we may assume that there exists α>0\alpha> 0 such that for arbitrarily small hh,

f(h)∣h∣>α.\frac{f(h)}{|h|} > \alpha.

Since f(h)=o(1)f(h)=o(1) and D(f,h)=o(h)D(f,h)=o(h), there exists β>0\beta> 0 such that whenever ∣h∣<β|h|<\beta we have

∣f(h)∣<1|f(h)|<1

and

∣D(f,h)∣=∣f(h)−c1f(ch)−d1f(dh)∣<(λ−1)∣d∣d14α∣h∣.|D(f,h)|=|f(h)-c_1f(ch)-d_1f(dh)|<\frac{(\lambda-1)|d|d_1}{4}\alpha|h|.

Choose h0h_0 small enough that the following three properties hold:

∣h0∣<β∣d∣,f(h0)∣h0∣>α,and∣h0∣ρ<(β∣d∣)1+ρα4.|h_0|<\frac{\beta}{|d|},\qquad\frac{f(h_0)}{|h_0|}>\alpha,\qquad\text{and}\qquad|h_0|^\rho<\left(\frac{\beta}{|d|}\right)^{1+\rho}\frac{\alpha}{4}.

The last inequality is possible because ρ>0\rho>0, so ∣h∣ρ→0|h|^\rho\to0 as h→0h\to0. Now, set

ϵ0=f(h0)2>0,\epsilon_0=\frac{f(h_0)}{2}>0,

and define

G(h):=f(h)−ϵ0∣h∣.G(h):=\frac{f(h)-\epsilon_0}{|h|}.

We define a sequence h0,h1,h2,…h_0,h_1,h_2,\ldots recursively as follows. For each i≥0i\geq0:

Case 1. If

f(chi)≥∣c∣f(hi)+(1−∣c∣)ϵ0,f(ch_i)\geq|c|f(h_i)+(1-|c|)\epsilon_0,

then

hi+1=chi.h_{i+1}=ch_i.

Case 2. If

f(chi)<∣c∣f(hi)+(1−∣c∣)ϵ0,f(ch_i)<|c|f(h_i)+(1-|c|)\epsilon_0,

then

hi+1=dhi.h_{i+1}=dh_i.

We now state and prove several claims.

Claim 0. G(h0)>α2G(h_0)>\frac{\alpha}{2}.

Indeed,

G(h0)=f(h0)−ϵ0∣h0∣=f(h0)/2∣h0∣>α2.G(h_0)=\frac{f(h_0)-\epsilon_0}{|h_0|} =\frac{f(h_0)/2}{|h_0|} >\frac{\alpha}{2}.

Claim 1. In Case 1, G(hi+1)≥G(hi)G(h_{i+1})\geq G(h_i).

Indeed,

G(hi+1)=f(chi)−ϵ0∣c∣∣hi∣≥∣c∣f(hi)+(1−∣c∣)ϵ0−ϵ0∣c∣∣hi∣=f(hi)−ϵ0∣hi∣=G(hi).G(h_{i+1})=\frac{f(ch_i)-\epsilon_0}{|c||h_i|} \geq\frac{|c|f(h_i)+(1-|c|)\epsilon_0-\epsilon_0}{|c||h_i|} =\frac{f(h_i)-\epsilon_0}{|h_i|} =G(h_i).

Claim 2. In Case 2, if ∣hi∣<β|h_i| < \beta. Then

G(hi+1)>λG(hi)−λ−14α.G(h_{i+1}) > \lambda G(h_i) - \frac{\lambda- 1}{4}\alpha.

Proof of Claim 2: Since ∣hi∣<β|h_i| < \beta,

∣f(hi)−c1f(chi)−d1f(dhi)∣<(λ−1)∣d∣d14α∣hi∣.|f(h_i) - c_1f(ch_i) - d_1f(dh_i)| < \frac{(\lambda- 1)|d|d_1}{4}\alpha|h_i|.

Therefore,

d1f(dhi)>f(hi)−c1f(chi)−(λ−1)∣d∣d14α∣hi∣.d_1f(dh_i) > f(h_i) - c_1f(ch_i) - \frac{(\lambda- 1)|d|d_1}{4}\alpha|h_i|.

Subtracting d1ϵ0d_1\epsilon_0 from both sides and dividing by ∣d∣∣hi∣|d||h_i|, we get

d1G(hi+1)=d1f(dhi)−ϵ0∣d∣∣hi∣>f(hi)−c1f(chi)−(λ−1)∣d∣d14α∣hi∣−d1ϵ0∣d∣∣hi∣.\begin{aligned} d_1G(h_{i+1}) = d_1\frac{f(dh_i)-\epsilon_0}{|d||h_i|} \\ &> \frac{f(h_i)-c_1f(ch_i)-\frac{(\lambda-1)|d|d_1}{4}\alpha|h_i|-d_1\epsilon_0}{|d||h_i|}. \end{aligned}

Now Case 2 means that

f(chi)<∣c∣f(hi)+(1−∣c∣)ϵ0,f(ch_i) < |c|f(h_i) + (1-|c|)\epsilon_0,

so

−c1f(chi)>−c1∣c∣f(hi)−c1(1−∣c∣)ϵ0.-c_1f(ch_i) > -c_1|c|f(h_i)-c_1(1-|c|)\epsilon_0.

Substituting this gives

d1G(hi+1)>(1−c1∣c∣)f(hi)−(c1(1−∣c∣)+d1)ϵ0−(λ−1)∣d∣d14α∣hi∣∣d∣∣hi∣.d_1G(h_{i+1}) > \frac{(1-c_1|c|)f(h_i)-(c_1(1-|c|)+d_1)\epsilon_0-\frac{(\lambda-1)|d|d_1}{4}\alpha|h_i|}{|d||h_i|}.

Since c1+d1≤1c_1+d_1\leq1, we have

c1(1−∣c∣)+d1≤1−c1∣c∣.c_1(1-|c|)+d_1\leq1-c_1|c|.

Therefore,

d1G(hi+1)>(1−c1∣c∣)(f(hi)−ϵ0)−(λ−1)∣d∣d14α∣hi∣∣d∣∣hi∣.d_1G(h_{i+1}) > \frac{(1-c_1|c|)(f(h_i)-\epsilon_0)-\frac{(\lambda-1)|d|d_1}{4}\alpha|h_i|}{|d||h_i|}.

Hence,

d1G(hi+1)>1−c1∣c∣∣d∣G(hi)−(λ−1)d14α,d_1G(h_{i+1}) > \frac{1-c_1|c|}{|d|}G(h_i)-\frac{(\lambda-1)d_1}{4}\alpha,

and dividing by d1d_1 yields the claim.

Claim 3. In Case 2, if ∣hi∣<β|h_i| < \beta and G(hi)≥α/2G(h_i)\geq\alpha/2 then G(hi+1)>α/2G(h_{i+1})>\alpha/2.

Indeed, if G(hi)≥α/2G(h_i)\geq\alpha/2, then by Claim 2,

G(hi+1)>α2−λ−14α=λ+14α>α2,G(h_{i+1})>\frac{\alpha}{2}-\frac{\lambda-1}{4}\alpha=\frac{\lambda+1}{4}\alpha>\frac{\alpha}{2},

since λ>1\lambda>1.

Claim 4. If max⁡{h0,h1,…,hn}<β\max\{h_0,h_1,\ldots,h_n\}<\beta then min⁡{f(h0),f(h1),…,f(hn)}>ϵ0\min\{f(h_0),f(h_1),\ldots,f(h_n)\}>\epsilon_0.

Proof of Claim 4: By Claim 0,

G(h0)>α2.G(h_0)>\frac{\alpha}{2}.

Suppose for some i≤ni\leq n,

G(hi)>α2.G(h_i)>\frac{\alpha}{2}.

Then, using Claim 1 in Case 1, or Claim 3 in Case 2,

G(hi+1)>α2.G(h_{i+1}) > \frac{\alpha}{2}.

Therefore, by induction,

G(hi)>α2for all i=0,1,…,n.G(h_i) > \frac{\alpha}{2} \qquad\text{for all } i=0,1,\ldots,n.

Consequently,

f(hi)−ϵ0=∣hi∣G(hi)>0,f(h_i)-\epsilon_0=|h_i|G(h_i)>0,

so

f(hi)>ϵ0for all i=0,1,…,n.f(h_i)>\epsilon_0 \qquad\text{for all } i=0,1,\ldots,n.

Claim 5. If max⁡{h0,h1,…,hn}<β\max\{h_0,h_1,\ldots,h_n\}<\beta and Case 2 is used mm times in the sequence {h0,h1,…,hn}\{h_0,h_1,\ldots,h_n\} then

G(hn)>λmα4.G(h_n)>\frac{\lambda^m\alpha}{4}.

Proof of Claim 5: If mm be the number of times Case 2 is used in the construction of

h0,h1,…,hn.h_0,h_1,\ldots,h_n.

Then Case 1 is used exactly n−mn-m times, so

∣hn∣=∣h0∣∣c∣n−m∣d∣m.|h_n|=|h_0||c|^{n-m}|d|^m.

Each time Case 1 occurs, Claim 1 shows that GG does not decrease.

T(x):=λx−λ−14α.T(x):=\lambda x-\frac{\lambda-1}{4}\alpha.

Since Case 2 occurs exactly mm times, we obtain from Claim 2 that

G(hn)>Tm(G(h0)).G(h_n)>T^m(G(h_0)).

Expanding this recursion gives

G(hn)>λmG(h0)−λ−14α(1+λ+⋯+λm−1).G(h_n)>\lambda^mG(h_0)-\frac{\lambda-1}{4}\alpha(1+\lambda+\cdots+\lambda^{m-1}).

Since

1+λ+⋯+λm−1=λm−1λ−1,1+\lambda+\cdots+\lambda^{m-1}=\frac{\lambda^m-1}{\lambda-1},

this becomes

G(hn)>λmG(h0)−λm−14α>λm(G(h0)−α4).G(h_n)>\lambda^mG(h_0)-\frac{\lambda^m-1}{4}\alpha>\lambda^m\left(G(h_0)-\frac{\alpha}{4}\right).

Since G(h0)>α/2G(h_0)>\alpha/2 by Claim 0, we conclude that

G(hn)>λmα4.G(h_n)>\lambda^m\frac{\alpha}{4}.

Claim 6. For all ii,

∣hi∣<β.|h_i|<\beta.

Proof of Claim 6: In fact, we will prove the stronger inequality ∣hi∣<β∣d∣|h_i|<\frac{\beta}{|d|}. Suppose not. Then there exists a smallest i≥0i\geq0 such that

∣hi∣≥β∣d∣.|h_i|\geq\frac{\beta}{|d|}.

Since ∣h0∣<β∣d∣|h_0|<\frac{\beta}{|d|} by construction, we must have i≥1i\geq1. By minimality of ii, we have

∣hi−1∣<β∣d∣|h_{i-1}| < \frac{\beta}{|d|}

Since hih_i is obtained from hi−1h_{i-1} by multiplying either by cc or by dd, we have

∣hi∣=∣c∣∣hi−1∣or∣hi∣=∣d∣∣hi−1∣.|h_i| = |c||h_{i-1}| \quad\text{or} \quad|h_i| = |d||h_{i-1}|.

In either case, using ∣hi−1∣<β∣d∣|h_{i-1}| < \frac{\beta}{|d|} and 0<∣c∣<1<∣d∣0 < |c| < 1 < |d|, we obtain

∣hi∣<β.|h_i| < \beta.

Let mm be the number of times we used Case 2 in the sequence {h0,h1,…,hi}\{h_0,h_1,\ldots,h_i\}. Since

∣hi∣=∣h0∣∣c∣i−m∣d∣m≤∣h0∣∣d∣m,|h_i| = |h_0||c|^{i-m}|d|^m \leq|h_0||d|^m,

we have

∣d∣m≥∣hih0∣.|d|^m \geq\left|\frac{h_i}{h_0}\right|.

Because

ρ=log⁡λlog⁡∣d∣,\rho= \frac{\log\lambda}{\log|d|},

we have

λ=∣d∣ρ,\lambda= |d|^\rho,

and therefore

λm=∣d∣ρm≥(∣hih0∣)ρ.\lambda^m = |d|^{\rho m} \geq\left(\left|\frac{h_i}{h_0}\right|\right)^\rho.

Using Claim 5, we obtain

f(hi)−ϵ0=∣hi∣G(hi)>∣hi∣λmα4≥∣hi∣(∣hih0∣)ρα4.f(h_i)-\epsilon_0=|h_i|G(h_i)>|h_i|\lambda^m\frac{\alpha}{4}\geq|h_i|\left(\left|\frac{h_i}{h_0}\right|\right)^\rho\frac{\alpha}{4}.

Since ∣hi∣≥β/∣d∣|h_i|\geq\beta/|d|, this yields

f(hi)−ϵ0≥α4(β∣d∣)1+ρ∣h0∣−ρ.f(h_i)-\epsilon_0\geq\frac{\alpha}{4}\left(\frac{\beta}{|d|}\right)^{1+\rho}|h_0|^{-\rho}.

By the choice of h0h_0,

∣h0∣ρ<α4(β∣d∣)1+ρ,|h_0|^\rho<\frac{\alpha}{4}\left(\frac{\beta}{|d|}\right)^{1+\rho},

and therefore

α4(β∣d∣)1+ρ∣h0∣−ρ>1.\frac{\alpha}{4}\left(\frac{\beta}{|d|}\right)^{1+\rho}|h_0|^{-\rho}>1.

Hence,

f(hi)−ϵ0>1.f(h_i)-\epsilon_0>1.

Since ϵ0>0\epsilon_0>0, we obtain

f(hi)>1.f(h_i)>1.

But this contradicts the choice of β\beta, because ∣hi∣<β|h_i|<\beta and because ∣f(h)∣<1|f(h)|<1 whenever ∣h∣<β|h|<\beta. This contradiction establishes Claim 6.

Claim 7. ϵ0<f(hi)<1\epsilon_0<f(h_i)<1 for all ii.

Proof of Claim 7: By Claim 6, ∣hi∣<β|h_i|<\beta for all ii. Therefore, by choice of β\beta, f(hi)<1f(h_i)<1 for all ii, and by Claim 4, f(hi)>ϵ0f(h_i)>\epsilon_0 for all ii.

Claim 8. ∣hi∣→0|h_i|\to0.

Proof of Claim 8: Suppose first that Case 2 occurs infinitely many times. Then Claims 1 and 5 imply G(hi)→∞G(h_i) \to\infty. By definition, G(hi)=f(hi)−ϵ0∣hi∣G(h_i) = \frac{f(h_i)-\epsilon_0}{|h_i|} and by Claim 7, ∣f(hi)∣<1|f(h_i)| < 1. So, this is possible only if ∣hi∣→0|h_i| \to0.

Now suppose that Case 2 occurs only finitely many times. Then from some point onward only Case 1 occurs, so

hi+1=chih_{i+1}=ch_i

for all sufficiently large ii. Hence

∣hi+1∣=∣c∣ ∣hi∣|h_{i+1}|=|c|\,|h_i|

for all sufficiently large ii, and since 0<∣c∣<10 < |c| < 1, it follows that hi→0h_i \to0. This finishes the proof of Claim 8.

To finish the proof of the theorem, note that Claims 7 and 8 directly contradict that f(h)=o(1)f(h) = o(1). Therefore our assumption was false, and hence

f(h)=o(h).f(h)=o(h).

We are now ready to handle the irrational case. This next theorem together with Theorem 25 establishes Theorem 24.

Theorem 28. Let D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh) be an irrational difference with 1<∣r∣<∣s∣1<|r|<|s|. Suppose that

∣a∣+1≤∣b∣and∣s∣+∣a∣<∣br∣.|a|+1\leq|b| \qquad\text{and} \qquad|s|+|a|<|br|.

Then D(f,h)D(f,h) has the MZ-property.

Proof. Let

c=1r,d=sr,c1=∣ab∣,and d1=∣1b∣.c=\frac{1}{r}, \qquad d=\frac{s}{r}, \qquad c_1=\left|\frac{a}{b}\right|, \qquad\text{and } d_1=\left|\frac{1}{b}\right|.

Then

1<∣r∣<∣s∣ implies 0<∣c∣<1<∣d∣,∣a∣+1≤∣b∣ implies c1+d1≤1,∣s∣+∣a∣<∣br∣ implies ∣sbr∣+∣abr∣<1,i.e. c1∣c∣+d1∣d∣<1.\begin{aligned} 1<|r|<|s| &\text{ implies } 0<|c|<1<|d|,\\ |a|+1\leq|b| &\text{ implies } c_1+d_1\leq1,\\ |s|+|a|<|br| &\text{ implies } \left|\frac{s}{br}\right|+\left|\frac{a}{br}\right|<1,\qquad\text{i.e. } c_1|c|+d_1|d|<1. \end{aligned}

Therefore, by Theorem 27,

f(h)−c1f(ch)−d1f(dh)=−c1f(ch)+f(h)−d1f(dh)f(h)-c_1f(ch)-d_1f(dh)=-c_1f(ch)+f(h)-d_1f(dh)

has the MZ-property. Recall that the MZ-property is invariant under scaling the nodes and coefficients. Dividing the nodes by cc shows that

−c1f(h)+f(1ch)−d1f(dch)=−c1f(h)+f(rh)−d1f(sh)-c_1f(h)+f\left(\frac{1}{c}h\right)-d_1f\left(\frac{d}{c}h\right)=-c_1f(h)+f(rh)-d_1f(sh)

has the MZ-property. Multiplying the coefficients by ∣b∣|b| shows that

−∣b∣c1f(h)+∣b∣f(rh)−∣b∣d1f(sh)=−∣a∣f(h)+∣b∣f(rh)−f(sh)-|b|c_1f(h)+|b|f(rh)-|b|d_1f(sh)=-|a|f(h)+|b|f(rh)-f(sh)

has the MZ-property. Then, applying Lemma 26, we get that

D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh)

has the MZ-property. □

The cases where MZ-property fails

Theorem 29. Let r,sr,s be nonzero real numbers with log⁡∣r∣log⁡∣s∣∉Q\frac{\log|r|}{\log|s|} \notin\mathbb{Q}, and let a,b∈Ca,b \in\mathbb{C}. Assume that

∣a∣+∣b∣>1and∣∣a∣−∣br∣∣≤∣s∣.|a|+|b|>1 \qquad\text{and}\qquad\bigl||a|-|br|\bigr|\leq|s|.

Then for the difference

D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh)

there exists a complex-valued function ff such that

f(h)=o(1),f(h)≠o(h),andD(f,h)≡0 for all h∈R.f(h)=o(1),\qquad f(h)\ne o(h),\qquad\text{and}\qquad D(f,h)\equiv0\text{ for all }h\in\mathbb{R}.

Proof. Since

max⁡θ∈R∣−a+∣b∣eiθ∣=∣a∣+∣b∣>1,\max_{\theta\in\mathbb{R}}|-a+|b|e^{i\theta}|=|a|+|b|>1,

there exists τ∈R\tau\in\mathbb{R} such that

∣−a+∣b∣eiτ∣>1.|-a+|b|e^{i\tau}|>1.

Also,

min⁡θ∈R∣−a+∣b∣∣r∣eiθ∣=∣∣a∣−∣b∣∣r∣∣≤∣s∣,\min_{\theta\in\mathbb{R}}|-a+|b||r|e^{i\theta}|=\bigl||a|-|b||r|\bigr|\leq|s|,

so there exists σ∈R\sigma\in\mathbb{R} such that

∣−a+∣b∣∣r∣eiσ∣≤∣s∣.|-a+|b||r|e^{i\sigma}|\leq|s|.

Define

Ψ(t,θ)=∣s∣t−∣−a+∣b∣∣r∣teiθ∣,(t,θ)∈[0,1]×R.\Psi(t,\theta)=|s|^t-|-a+|b||r|^t e^{i\theta}|,\qquad(t,\theta)\in[0,1]\times\mathbb{R}.

Then

Ψ(0,τ)<0andΨ(1,σ)≥0.\Psi(0,\tau)<0\qquad\text{and}\qquad\Psi(1,\sigma)\geq0.

Let

θ(t)=(1−t)τ+tσandF(t)=Ψ(t,θ(t)).\theta(t)=(1-t)\tau+t\sigma\qquad\text{and}\qquad F(t)=\Psi(t,\theta(t)).

Then FF is continuous on [0,1][0,1], with

F(0)<0andF(1)≥0.F(0)<0\qquad\text{and}\qquad F(1)\geq0.

Hence, by the intermediate value theorem, there exists t0∈(0,1]t_0\in(0,1] such that

F(t0)=0.F(t_0)=0.

Thus

∣s∣t0=∣−a+∣b∣∣r∣t0eiθ(t0)∣.|s|^{t_0}=|-a+|b||r|^{t_0}e^{i\theta(t_0)}|.

Therefore there exists θ1∈R\theta_1\in\mathbb{R} such that

∣s∣t0eiθ1=−a+∣b∣∣r∣t0eiθ(t0),|s|^{t_0}e^{i\theta_1}=-a+|b||r|^{t_0}e^{i\theta(t_0)},

and θ2∈R\theta_2\in\mathbb{R} such that

a+b∣r∣t0eiθ2+∣s∣t0eiθ1=0.(19)a+b|r|^{t_0}e^{i\theta_2}+|s|^{t_0}e^{i\theta_1}=0. \tag*{(19)}

Let

H={∣r∣n∣s∣m:n,m∈Z}.H=\{|r|^n|s|^m:n,m\in\mathbb{Z}\}.

Since log⁡(∣r∣)/log⁡(∣s∣)∉Q\log(|r|)/\log(|s|)\notin\mathbb{Q}, each h∈Hh\in H has a unique representation

h=∣r∣n∣s∣m.h=|r|^n|s|^m.

Define

f(h)={ei(nθ2+mθ1)∣h∣t0,h∈H, h=∣r∣n∣s∣m,0,h∉H.f(h)= \begin{cases} e^{i(n\theta_2+m\theta_1)}|h|^{t_0}, & h\in H,\ h=|r|^n|s|^m,\\ 0, & h\notin H. \end{cases}

Since 0<t0≤10 < t_0 \le1, f(h)=o(1)f(h) = o(1) and f(h)≠o(h)f(h) \ne o(h). For h∈Hh \in H, and hence both rh∈Hrh \in H and sh∈Hsh \in H,

D(f,h)=(a+b∣r∣t0eiθ2+∣s∣t0eiθ1)f(h)=0D(f,h) = \left(a+b|r|^{t_0}e^{i\theta_2}+|s|^{t_0}e^{i\theta_1}\right)f(h)=0

by (1). For h∉Hh \notin H, and hence both rh∉Hrh \notin H and sh∉Hsh \notin H,

D(f,h)=0.D(f,h)=0.

Thus

D(f,h)≡0for all h∈R.D(f,h)\equiv0 \qquad\text{for all } h\in\mathbb{R}.

The following examples show that Theorem 29 can go either way in the rational case.

Example 30. Let D(f,h)=f(h)+2f(4h)+f(8h)D(f,h)=f(h)+2f(4h)+f(8h), with a=1,b=2,r=4,s=8a=1,b=2,r=4,s=8. Then D(f,h)D(f,h) is rational with ∣a∣+∣b∣>1|a|+|b|>1 and ∣∣a∣−∣br∣∣=7<∣s∣\lvert|a|-|br|\rvert=7<|s|. The inequalities of Theorem 29 are satisfied. Since the nodes are positive, there is only one characteristic polynomial, x3+2x2+1x^3+2x^2+1. The roots are approximately −2.21-2.21, 0.1+0.67i0.1+0.67i, and 0.1−0.67i0.1-0.67i. The absolute values of its roots are 2.212.21 and 0.670.67, which are outside of (1,2](1,2] so by Theorem 12, D(f,h)D(f,h) has the MZ-property.

Example 31. Let D(f,h)=f(h)−3f(2h)+f(8h)D(f,h)=f(h)-3f(2h)+f(8h). Then D(f,h)D(f,h) is rational with a=1,b=−3,r=2,s=8a=1,b=-3,r=2,s=8. Then ∣a∣+∣b∣>1|a|+|b|>1 and ∣∣a∣−∣br∣∣=5<∣s∣\lvert|a|-|br|\rvert=5<|s|. The inequalities of Theorem 29 are satisfied. Since the nodes are positive, there is only one characteristic polynomial, x3−3x+1x^3-3x+1. One of its roots, 1.531.53, is in the region (1,2](1,2]. So, by Theorem 12, D(f,h)D(f,h) does not have the MZ-property.

Theorem 32. Suppose that

1+∣a∣>∣b∣and∣s∣<∣b∣∣r∣−∣a∣.1+|a|>|b| \qquad\text{and} \qquad|s|<|b||r|-|a|.

Then for the difference

D(f,h)=∣a∣f(h)−∣b∣f(rh)+f(sh)D(f,h)=|a|f(h)-|b|f(rh)+f(sh)

there exists a complex-valued function ff such that

f(h)=o(1),f(h)∉o(h),andD(f,h)≡0 for all h∈R.f(h)=o(1), \qquad f(h)\notin o(h), \qquad\text{and} \qquad D(f,h)\equiv0 \text{ for all } h\in\mathbb{R}.

Proof. Define

g(t)=∣s∣log⁡∣r∣t−∣b∣t+∣a∣.g(t)=|s|^{\log|r|t}-|b|t+|a|.

Then

g(1)=1−∣b∣+∣a∣>0g(1)=1-|b|+|a|>0

and

g(∣r∣)=∣s∣−∣b∣∣r∣+∣a∣<0.g(|r|)=|s|-|b||r|+|a|<0.

Hence, by continuity, there exists

1<t0<∣r∣1<t_0<|r|

such that

g(t0)=0.g(t_0)=0.

Define

f(h)=∣h∣log⁡∣r∣t0.f(h)=|h|^{\log|r|t_0}.

Then

f(sh)=∣s∣log⁡∣r∣t0f(h)f(sh)=|s|^{\log|r|t_0}f(h)

and

f(rh)=t0f(h).f(rh)=t_0f(h).

Therefore

D(f,h)=(∣a∣−∣b∣t0+∣s∣log⁡∣r∣t0)f(h)=g(t0)f(h)=0.D(f,h)=\left(|a|-|b|t_0+|s|^{\log|r|t_0}\right)f(h)=g(t_0)f(h)=0.

Since

1<t0<∣r∣,1 < t_0 < \lvert r\rvert,

we have

0<log⁡∣r∣t0<1,0 < \log_{\lvert r\rvert} t_0 < 1,

and consequently

f(h)=o(1)andf(h)≠o(h).f(h) = o(1) \qquad\text{and} \qquad f(h) \ne o(h).

We now complete the proof of Theorem 2.

Theorem 33. Let D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh) be an irrational difference with distinct nodes, 1,r,s1,r,s where 1≤∣r∣≤∣s∣1 \le\lvert r\rvert\le\lvert s\rvert. If DD has the MZ-property then it satisfies at least one of the following conditions:

  • ∣a∣+∣b∣≤1\lvert a\rvert+\lvert b\rvert\le1

  • ∣br∣+∣s∣<∣a∣\lvert br\rvert+\lvert s\rvert< \lvert a\rvert

  • ∣a∣+1≤∣b∣\lvert a\rvert+1 \le\lvert b\rvert and ∣s∣+∣a∣<∣br∣\lvert s\rvert+\lvert a\rvert< \lvert br\rvert.

Proof. Suppose D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh) is irrational and does not satisfy any of the three conditions. We will show it does not have the MZ-property. Since D(f,h)D(f,h) is irrational, we have 1<∣r∣<∣s∣1<\lvert r\rvert<\lvert s\rvert.

Case 1: ∣a∣+∣b∣>1\lvert a\rvert+\lvert b\rvert>1, ∣br∣+∣s∣≥∣a∣\lvert br\rvert+\lvert s\rvert\ge\lvert a\rvert, and ∣s∣+∣a∣≥∣br∣\lvert s\rvert+\lvert a\rvert\ge\lvert br\rvert. Then ∣s∣≥∣br∣−∣a∣\lvert s\rvert\ge\lvert br\rvert-\lvert a\rvert and ∣s∣≥∣a∣−∣br∣\lvert s\rvert\ge\lvert a\rvert-\lvert br\rvert. So, ∣s∣≥∣∣a∣−∣br∣∣\lvert s\rvert\ge\bigl\lvert\lvert a\rvert-\lvert br\rvert\bigr\rvert. By Theorem 29, D(f,h)D(f,h) does not have the MZ-property.

Case 2: ∣a∣+∣b∣>1\lvert a\rvert+\lvert b\rvert>1, ∣br∣+∣s∣≥∣a∣\lvert br\rvert+\lvert s\rvert\ge\lvert a\rvert, and ∣s∣+∣a∣<∣br∣\lvert s\rvert+\lvert a\rvert<\lvert br\rvert. Then ∣a∣+1>∣b∣\lvert a\rvert+1>\lvert b\rvert, so by Theorem 32, the difference ∣a∣f(h)−∣b∣f(rh)+f(sh)\lvert a\rvert f(h)-\lvert b\rvert f(rh)+f(sh) fails the MZ-property. Then by Lemma 26, D(f,h)=af(h)+bf(rh)+f(sh)D(f,h)=af(h)+bf(rh)+f(sh) also fails the MZ-property.

References

  1. [1]J. M. Ash, S. Catoui, and M. Csörnyei, Generalized vs. ordinary differentiation, Proc. Amer. Math. Soc. 145 (2017), 1553–1565.DOI
  2. [2]J. M. Ash, S. Catoui, and H. Fejzić, Gaussian Riemann derivatives, Israel J. Math. 255 (2023) (1), 177–199.arxiv.org/abs/2211.09209
  3. [3]J. M. Ash, S. Catoui, and H. Fejzić, Two pointwise characterizations of the Peano derivative, Results Math. 79 (2024) (7), Article No. 251, 17pp.arxiv.org/abs/2209.04088
  4. [4]S. Catoui, and H. Fejzić, NonEquivalence Between Peano and Riemann derivatives, Israel J. Math. 274 (2026), no. 1, 407–419.
  5. [5]S. Catoui and H. Fejzić, Problem for MZ differentiable functions. 1-14. (Pre-print)
  6. [6]H. Fejzić, The Marcinkiewicz–Zygmund Property for Riemann Differences with Geometric Nodes, Real Anal. Exchange 50(2) (2025), 445–468.arxiv.org/abs/2507.11463
  7. [7]J. Marcin kiewicz and A. Zygmund, On the differentiability of functions and summability of trigonometric series, Fund. Math. 26 (1936), 1–43.DOI

Paper details

vlnc:W_cn54546n
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

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