betaFeedback

Reading v1View latest version

WorksMathematics

Explicit Measurable Counterexamples to Finite-Difference Equations

Abstract

We study finite-difference operators of the form D(f,h)=∑k=1nakf(bkh), D(f,h)=\sum_{k=1}^n a_k f(b_kh), where the coefficients may be complex and the nodes are real. Ash, Catoiu, and Csörnyei characterized the finite differences for which D(f,h)=o(h)⟹f(h)=o(h) D(f,h)=o(h)\quad\Longrightarrow\quad f(h)=o(h) holds for every measurable real-valued function. We extend their counterexample theorem to complex-valued functions defined on C\mathbb C. Except for the same exceptional four-term configuration, we construct a Lebesgue measurable function ff such that D(f,h)=0for every h∈C, D(f,h)=0\qquad\text{for every }h\in\mathbb C, while f(h)≠o(1)f(h)\not=o(1) as h→0h\to0. The counterexamples are supported on countable multiplicative subgroups. The construction is based on zeros of Laurent polynomials and multiplicative characters, and avoids the infinite-dimensional linear-algebra argument used in the earlier proof.

Let n∈Nn\in\mathbb{N}, let a1,…,an∈Ca_1,\ldots,a_n\in\mathbb{C}, and let b1,…,bn∈Rb_1,\ldots,b_n\in\mathbb{R} be distinct. For a function ff defined on C\mathbb{C}, we consider the finite-difference operator

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

The numbers bkb_k are called the nodes, and the numbers aka_k are the corresponding coefficients. This paper concerns the implication

D(f,h)=o(h)as h→0⟹f(h)=o(h)as h→0.D(f,h)=o(h)\quad\text{as }h\to0\quad\Longrightarrow\quad f(h)=o(h)\quad\text{as }h\to0.

for measurable functions ff. In the real-valued setting, Ash, Catoiu, and Csörnyei [1] completely characterized the finite differences for which this implication holds for every measurable function f:R→Rf:\mathbb{R}\to\mathbb{R}. Surprisingly, apart from multiplication by a nonzero constant and a rescaling of the variable, the only exceptional difference is

R(f,h)=λf(βh)+λf(−βh)+f(h)−f(−h),λβ≠0.R(f,h)=\lambda f(\beta h)+\lambda f(-\beta h)+f(h)-f(-h),\qquad\lambda\beta\ne0.

2020 Mathematics Subject Classification. Primary 39B05; Secondary 26A24, 28A05.

The difficult part of their result is the converse. Namely, if a finite difference is not of this exceptional form, then there exists a measurable function ff such that

D(f,h)=0for every h,D(f,h)=0 \qquad\text{for every }h,

while

f(h)≠o(h)as h→0.f(h)\ne o(h) \qquad\text{as }h\to0.

The proof in [1] is existential in nature: the counterexample is obtained from the solvability of an infinite system of linear equations. In contrast, our proof gives an explicit construction. The counterexamples are supported on countable multiplicative subgroups and are described by multiplicative characters whose parameters are obtained from zeros of finite Laurent polynomials. Thus, the construction not only establishes existence but also illuminates the underlying algebraic mechanism responsible for the failure of the implication.

To illustrate this algebraic mechanism concrete, consider the operator

D(f,h)=f(4h)−3f(6h)+f(9h)=f(22h)−3f(2⋅3h)+f(32h).D(f,h)=f(4h)-3f(6h)+f(9h)=f(2^{2}h)-3f(2\cdot3h)+f(3^{2}h).

The corresponding Laurent polynomial is

P+(x,y)=x2−3xy+y2.P_{+}(x,y)=x^{2}-3xy+y^{2}.

To apply Lemma 1, we choose an integer point that does not belong to any of the hyperplanes determined by the exponent vectors. In this case, the relevant hyperplanes are

2u=u+v,2u=2v,u+v=2v.2u=u+v,\qquad2u=2v,\qquad u+v=2v.

The integer point (1,2)(1,2) avoids all three of them. Consequently,

P+(t,t2)=t2−3t3+t4=t2(t2−3t+1).P_{+}(t,t^{2})=t^{2}-3t^{3}+t^{4}=t^{2}(t^{2}-3t+1).

Setting

t=3+52,t=\frac{3+\sqrt{5}}{2},

we see that tt is a nonzero root of P+(t,t2)P_{+}(t,t^{2}).

Now, choose the countable multiplicative subgroup

G={ε5ℓ2p3q:ε∈{±1}, ℓ,p,q∈Z}.G=\left\{\varepsilon5^{\ell}2^{p}3^{q}:\varepsilon\in\{\pm1\},\ \ell,p,q\in\mathbb{Z}\right\}.

The representation of each element of GG in this form is unique. Define f:C→Rf:\mathbb{C}\to\mathbb{R} by setting

f(z)=0if z∉G,f(z)=0 \qquad\text{if }z\notin G,

and

f(ε5ℓ2p3q)=tp+2q.f\left(\varepsilon5^{\ell}2^{p}3^{q}\right)=t^{p+2q}.

A direct calculation yields

f(4h)=t2f(h),f(6h)=t3f(h),f(9h)=t4f(h)f(4h)=t^{2}f(h),\qquad f(6h)=t^{3}f(h),\qquad f(9h)=t^{4}f(h)

for every h∈Ch\in\mathbb{C}. Hence

D(f,h)=(t2−3t3+t4)f(h)=0D(f,h)=\left(t^{2}-3t^{3}+t^{4}\right)f(h)=0

for every h∈C.Sinceh\in\mathbb{C}. SinceG$ is countable, ff is trivially Lebesgue measurable. Moreover, because

f(5−ℓ)=1(ℓ∈N),f(5^{-\ell})=1 \qquad(\ell\in\mathbb{N}),

it follows that f≠o(1)f\ne o(1) as h→0h\to0, completing the counterexample.

Lemma 1. Let

P(x1,…,xm)=∑k=1nakx1αk1⋯xmαkmP(x_1,\ldots,x_m)=\sum_{k=1}^{n}a_kx_1^{\alpha_{k1}}\cdots x_m^{\alpha_{km}}

be a Laurent polynomial with non-zero coefficients and at least two distinct exponent vectors,

(αi1,…,αim)≠(αj1,…,αjm).(\alpha_{i1},\ldots,\alpha_{im})\ne(\alpha_{j1},\ldots,\alpha_{jm}).

Then there exists

A∈(C∖{0})mA\in(\mathbb{C}\setminus\{0\})^m

such that

P(A)=0.P(A)=0.

Proof. By combining like terms, we may assume that all of the exponent vectors of PP are distinct. For each pair of exponent vectors,

(αi1,…,αim)≠(αj1,…,αjm),(\alpha_{i1},\ldots,\alpha_{im})\ne(\alpha_{j1},\ldots,\alpha_{jm}),

consider the hyperplane in Rm\mathbb{R}^m defined by

αi1x1+⋯+αimxm=αj1x1+⋯+αjmxm.\alpha_{i1}x_1+\cdots+\alpha_{im}x_m=\alpha_{j1}x_1+\cdots+\alpha_{jm}x_m.

Since each hyperplane has measure zero, there exists a point s=(s1,…,sm)∈Rms=(s_1,\ldots,s_m)\in\mathbb{R}^m that does not lie on any of the hyperplanes. Let ϵ>0\epsilon>0 be the distance from ss to the nearest hyperplane. By scaling, if necessary, we may assume ϵ>m\epsilon>\sqrt{m}. Let r=(r1,…,rm)r=(r_1,\ldots,r_m) be the nearest point to ss that is in the lattice Zm\mathbb{Z}^m. Then rr also does not lie on any of the hyperplanes.

Define

Q(t)=P(tr1,…,trm).Q(t)=P(t^{r_1},\ldots,t^{r_m}).

Then QQ is a nonconstant Laurent polynomial in one variable. By choice of rr all of the exponents of QQ are distinct. Multiplying by a suitable power of tt, we obtain a nonconstant ordinary polynomial

R(t)=tNQ(t).R(t)=t^NQ(t).

Since RR has at least two terms, it is not a monomial. Hence, there exists

a∈C∖{0}a\in\mathbb{C}\setminus\{0\}

such that

R(a)=0.R(a)=0.

Therefore

Q(a)=0,Q(a)=0,

and consequently

P(ar1,…,arm)=0.P(a^{r_1},\ldots,a^{r_m})=0.

Since a≠0a\ne0, all coordinates aria^{r_i} are nonzero.

Theorem 2. Let

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

be a finite difference with nonzero coefficients ak∈Ca_k \in\mathbb{C} and distinct nodes bk∈Rb_k \in\mathbb{R}. Suppose that there are at least two nonzero nodes and that DD cannot be scaled to be of the form

R(f,h)=af(bh)+af(−bh)+f(h)−f(−h).R(f,h)=a f(bh)+a f(-bh)+f(h)-f(-h).

Then there exists a Lebesgue measurable function

f:C⟶Cf:\mathbb{C}\longrightarrow\mathbb{C}

such that

f(h)≠o(1)as h→0f(h)\ne o(1) \qquad\text{as } h\to0

and

D(f,h)=0for every h∈C.D(f,h)=0 \qquad\text{for every } h\in\mathbb{C}.

Proof. If one of the nodes is zero, say b1=0b_1=0, define

f(z)={−1a1∑k=2nak,z=0,1,z≠0.f(z)= \begin{cases} -\dfrac{1}{a_1}\sum_{k=2}^{n}a_k, & z=0,\\ 1, & z\ne0. \end{cases}

Then f(h)≠o(1)f(h)\ne o(1) as h→0h\to0, and

D(f,h)=a1f(0)+∑k=2nak=0D(f,h)=a_1f(0)+\sum_{k=2}^{n}a_k=0

for every h∈Ch\in\mathbb{C}. We may therefore assume that none of the nodes bkb_k is zero.

Let HH be the multiplicative subgroup of R∖{0}\mathbb{R}\setminus\{0\} generated by

{−1,b1,…,bn}.\{-1,b_1,\ldots,b_n\}.

Since HH is finitely generated, there exist positive real numbers q1,…,qmq_1,\ldots,q_m such that every element of HH can be written uniquely in the form

εq1n1⋯qmnm,ε∈{±1},ni∈Z.\varepsilon q_1^{n_1}\cdots q_m^{n_m}, \qquad\varepsilon\in\{\pm1\},\quad n_i\in\mathbb{Z}.

For each node bkb_k, write

bk=σkq1αk1⋯qmαkm,σk∈{±1}.b_k=\sigma_kq_1^{\alpha_{k1}}\cdots q_m^{\alpha_{km}}, \qquad\sigma_k\in\{\pm1\}.

Define the Laurent polynomials

P+(x1,…,xm)=∑k=1nakx1αk1⋯xmαkmP_+(x_1,\ldots,x_m)=\sum_{k=1}^{n}a_kx_1^{\alpha_{k1}}\cdots x_m^{\alpha_{km}}

and

P−(x1,…,xm)=∑k=1nakσkx1αk1⋯xmαkm.(1)P_-(x_1,\ldots,x_m)=\sum_{k=1}^{n}a_k\sigma_kx_1^{\alpha_{k1}}\cdots x_m^{\alpha_{km}}. \tag*{(1)}

Suppose first that, after combining like terms and eliminating terms with zero coefficients, P+P_{+} is not a monomial. By Lemma 1, there exists

A=(A1,…,Am)∈(C∖{0})mA=(A_{1},\ldots,A_{m})\in(\mathbb{C}\setminus\{0\})^{m}

such that

P+(A)=0.P_{+}(A)=0.

Choose s>1s>1 such that

log⁡s,log⁡q1,…,log⁡qm\log s,\log q_{1},\ldots,\log q_{m}

are linearly independent over Q\mathbb{Q}, and let

G={εsℓq1n1⋯qmnm:ε∈{±1}, ℓ,n1,…,nm∈Z}.G=\left\{\varepsilon s^{\ell}q_{1}^{n_{1}}\cdots q_{m}^{n_{m}}:\varepsilon\in\{\pm1\},\ \ell,n_{1},\ldots,n_{m}\in\mathbb{Z}\right\}.

For each element of GG all the integers ε,ℓ,n1,…,nm\varepsilon,\ell,n_{1},\ldots,n_{m} are unique. Define f:C→Cf:\mathbb{C}\to\mathbb{C} by

f(z)=0if z∉G,f(z)=0\qquad\text{if }z\notin G,

and

f(εsℓq1n1⋯qmnm)=εA1n1⋯Amnm.f\left(\varepsilon s^{\ell}q_{1}^{n_{1}}\cdots q_{m}^{n_{m}}\right)=\varepsilon A_{1}^{n_{1}}\cdots A_{m}^{n_{m}}.

For every z∈Cz\in\mathbb{C} and every node bkb_{k}, we have

f(bkz)=A1αk1⋯Amαkmf(z).f(b_{k}z)=A_{1}^{\alpha_{k1}}\cdots A_{m}^{\alpha_{km}}f(z).

Indeed, this follows directly from the definition when z∈Gz\in G; when z∉Gz\notin G, both sides vanish because bk∈Gb_{k}\in G and hence bkz∉Gb_{k}z\notin G. Consequently,

D(f,z)=P+(A)f(z)=0D(f,z)=P_{+}(A)f(z)=0

for every z∈Cz\in\mathbb{C}.

Suppose next that, after combining like terms and eliminating terms with zero coefficients, P−P_{-} is not a monomial. Choose

A∈(C∖{0})mA\in(\mathbb{C}\setminus\{0\})^{m}

such that

P−(A)=0.P_{-}(A)=0.

Using the same group GG, define f(z)=0f(z)=0 for z∉Gz\notin G, and define

f(εsℓq1n1⋯qmnm)=εA1n1⋯Amnm.f\left(\varepsilon s^{\ell}q_{1}^{n_{1}}\cdots q_{m}^{n_{m}}\right)=\varepsilon A_{1}^{n_{1}}\cdots A_{m}^{n_{m}}.

Then

f(bkz)=σkA1αk1⋯Amαkmf(z)f(b_{k}z)=\sigma_{k}A_{1}^{\alpha_{k1}}\cdots A_{m}^{\alpha_{km}}f(z)

for every z∈Cz\in\mathbb{C}. It follows that

D(f,z)=P−(A)f(z)=0D(f,z)=P_{-}(A)f(z)=0

for every z∈Cz\in\mathbb{C}.

In either case, along the sequence

hℓ=s−ℓ,h_{\ell}=s^{-\ell},

we have hℓ→0h_\ell\to0 and

∣f(hℓ)∣=1.|f(h_\ell)| = 1.

Thus f(h)≠o(1)f(h) \ne o(1) as h→0h \to0. Moreover, GG is countable, so ff, being zero outside the countable set GG, is Lebesgue measurable on C\mathbb{C}.

It remains to consider the case in which both P+P_{+} and P−P_{-} reduce to monomials or to the zero polynomial.

Suppose first that P+P_{+} is a monomial. Then there is a node bb such that, if −b-b is also a node, the sum of the corresponding coefficients is nonzero. Moreover, for every node c∉{b,−b}c \notin\{b,-b\}, the node −c-c is also present, and the coefficients of cc and −c-c are negatives of one another.

If P+P_{+} is the zero polynomial, then every node occurs together with its opposite, and the coefficients of cc and −c-c are negatives of one another.

Similarly, if P−P_{-} is a monomial, then there is a node bb such that, if −b-b is also a node, the difference of the corresponding coefficients is nonzero. Moreover, for every node c∉{b,−b}c \notin\{b,-b\}, the node −c-c is also present, and the coefficients of cc and −c-c are the same.

The two polynomials cannot both be zero, since this would force every coefficient to be zero.

If P−P_{-} is the zero polynomial, then

D(f,h)=af(bh)+af(−bh).D(f,h) = af(bh) + af(-bh).

Define

f(z)={1,z>0,−1,z<0,0,otherwise.f(z) = \begin{cases} 1, & z > 0, \\ -1, & z < 0, \\ 0, & \text{otherwise}. \end{cases}

Then D(f,h)=0D(f,h) = 0 for every hh, and f(h)≠o(1)f(h) \ne o(1) as h→0h \to0.

If P+P_{+} is the zero polynomial, then

D(f,h)=af(bh)−af(−bh).D(f,h) = af(bh) - af(-bh).

In this case the constant function f≡1f \equiv1 satisfies D(f,h)=0D(f,h) = 0 for every hh.

Finally, suppose that both P+P_{+} and P−P_{-} are monomials. Then, after relabeling the two pairs if necessary,

D(f,h)=αbf(bh)+αbf(−bh)+αcf(ch)−αcf(−ch).D(f,h) = \alpha_b f(bh) + \alpha_b f(-bh) + \alpha_c f(ch) - \alpha_c f(-ch).

After multiplying the difference by a nonzero constant and replacing hh by chch, if necessary, this is of the exceptional form

R(f,h)=af(bh)+af(−bh)+f(h)−f(−h),R(f,h) = af(bh) + af(-bh) + f(h) - f(-h),

contrary to the hypothesis.

Finally, if the coefficients are real, then either the real part or the imaginary part of the constructed counterexample provides a real-valued function with the same properties.

The characterization of measurable finite differences obtained by Ash, Catoiu, and Csörnyei becomes substantially richer under the additional assumption that f(h)=o(1)f(h) = o(1) as h→0h \to0. For example, the finite difference

D(f,h)=f(2h)−3f(h)D(f,h) = f(2h) - 3f(h)

has the property that

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).

This and related applications to Riemann derivatives are established in [2].

References

  1. [1]J. M. Ash, S. Catoiu, and M. Csörnyei, Generalized vs. ordinary differentiation, Proc. Amer. Math. Soc. 145 (2017), 1553–1565.DOI
  2. [2]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

Paper details

vlnc:W_xttvrzpe
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

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