Affine invariance conjecture for the Poisson-series function class

About 1 year old · traced to

Let C\mathscr C denote the class of functions for which the Poisson series under consideration has the required convergence properties, and let an affine transformation of the variable mean x↦ax+bx\mapsto ax+b with a≠0a\neq0. Affine-invariance conjecture. The space C\mathscr C is invariant under affine transformations of the variable. The authors note that the corresponding impaired-integrability space is scale- and shift-invariant, whereas this property is not immediate for C\mathscr C; the claim is therefore formulated as a weaker conjecture.

References

Primary source

Jerzy Szulga, “Improper Poisson integral of the sinc function”, arXiv:2504.05574 (2025).

Progress summary

Refreshed
Claimed solved

A posted attempt says the conjecture fails for reflections and proves the corrected positive-direction version, but nobody has independently checked the argument.

The conjecture asks whether the Poisson-series function class is preserved by rescaling and shifting the variable. The primary paper studies this class through the sinc function and leaves the general invariance question as a conjecture.

Posted attempt

An undated attempt claims a counterexample for negative slopes, using f(x)=1/(1+ex)f(x)=1/(1+e^x), and claims a complete convergence criterion that proves invariance for a>0a>0 and shifts within the domain. The argument has not been independently verified.

Current status (as of August 2026): The unrestricted affine-invariance conjecture has an unverified counterexample claim, while the positive-slope version has an unverified claimed proof; no independent verification is recorded.

Sources

Solutions 1

CounterexampleThis solution needs a summarySee full solutionHide full solution

The unrestricted condition a≠0a\ne0 is false: negative slopes do not preserve the Poisson-series class. The corrected positive-slope assertion holds, with a complete convergence criterion.

Let NN be a rate-λ\lambda Poisson process on [0,∞)[0,\infty), with arrival times SjS_j. Consider the smooth bounded function on the whole real line

f(x)=11+ex.f(x)=\frac1{1+e^x}.

Campbell's formula gives

E∑j≥1f(Sj)=λ∫0∞dx1+ex=λlog⁡2<∞.\mathbb E\sum_{j\ge1}f(S_j) =\lambda\int_0^\infty\frac{dx}{1+e^x} =\lambda\log2<\infty.

Hence

∑j≥1f(Sj)<∞almost surely.\sum_{j\ge1}f(S_j)<\infty \qquad\text{almost surely}.

For the permitted affine reflection x↦−xx\mapsto-x, however,

f(−Sj)=11+e−Sj⟶1almost surely.f(-S_j)=\frac1{1+e^{-S_j}}\longrightarrow1 \qquad\text{almost surely}.

Thus the transformed series diverges, since its terms do not even tend to zero. Therefore invariance under all a≠0a\ne0 fails.

In the primary source, functions are defined on [0,∞)[0,\infty), so negative slopes are not affine self-maps of the stated domain. The intended assertion for a>0a>0 is true. More precisely, for any measurable ff, put

B={x:∣f(x)∣>1},g=f1{∣f∣≤1}.B=\{x:|f(x)|>1\}, \qquad g=f\mathbf1_{\{|f|\le1\}}.

Then

∑j≥1f(Sj) converges almost surely\sum_{j\ge1}f(S_j)\text{ converges almost surely}

if and only if

∣B∣<∞,∫0∞∣g(x)∣2 dx<∞,lim⁡T→∞∫0Tg(x) dx exists.|B|<\infty, \qquad \int_0^\infty|g(x)|^2\,dx<\infty, \qquad \lim_{T\to\infty}\int_0^Tg(x)\,dx \text{ exists}.

For sufficiency, the contribution from BB has finitely many terms, while

MT=∫0Tg(x)(dNx−λ dx)M_T=\int_0^Tg(x)(dN_x-\lambda\,dx)

is an L2L^2-bounded martingale because

E∣MT−MS∣2=λ∫ST∣g(x)∣2 dx.\mathbb E|M_T-M_S|^2 =\lambda\int_S^T|g(x)|^2\,dx.

Adding the convergent deterministic drift proves convergence.

Conversely, convergence forces finitely many large jumps, hence ∣B∣<∞|B|<\infty. After removing them, the independent-increment characteristic function gives

∣Eei∫STg dN∣=exp⁡(−λ∫ST(1−cos⁡g(x)) dx)⟶1.\left| \mathbb E e^{i\int_S^Tg\,dN} \right| = \exp\left( -\lambda\int_S^T(1-\cos g(x))\,dx \right) \longrightarrow1.

Since ∣g∣≤1|g|\le1 and 1−cos⁡g≥c∣g∣21-\cos g\ge c|g|^2, this forces g∈L2g\in L^2. The compensated martingale therefore converges, and subtracting it from the convergent Poisson integral forces convergence of the deterministic drift.

For a>0a>0, b≥0b\ge0, the substitution u=ax+bu=ax+b preserves all three conditions:

∣{∣f(a ⋅+b)∣>1}∣=1a∣B∩[b,∞)∣,|\{|f(a\,\cdot+b)|>1\}| =\frac1a|B\cap[b,\infty)|, ∫0∞∣g(ax+b)∣2 dx=1a∫b∞∣g(u)∣2 du,\int_0^\infty|g(ax+b)|^2\,dx =\frac1a\int_b^\infty|g(u)|^2\,du, ∫0Tg(ax+b) dx=1a∫baT+bg(u) du.\int_0^Tg(ax+b)\,dx =\frac1a\int_b^{aT+b}g(u)\,du.

Thus the source's well-defined positive-orientation conjecture is true, while its unrestricted negative-slope reformulation is false.