Linearity-Inducing Priors for Poisson Parameter Estimation Under $L^{1}$ Loss

Alex Dytso; H. Vincent Poor; Leighton P. Barnes

REVIEW 3 major objections 4 minor 1 cited by

Linearity-Inducing Priors for Poisson Parameter Estimation Under $L^{1}$ Loss

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Under Poisson noise, any sufficiently regular increasing function can be realized as the conditional median by a suitably chosen prior, and affine medians arise from priors beyond the gamma family.

desk verdict Novel construction, correct finite-dimensional piece, but Lemma 3's tail bound uses a moment equality that doesn't hold on the support, so the main theorem is unproven. read the letter →

arxiv 2505.21102 v1 pith:ICIN5PLH submitted 2025-05-27 math.ST cs.ITmath.ITstat.TH

Leighton P. Barnes , Alex Dytso , H. Vincent Poor This is my paper

classification math.STcs.ITmath.ITstat.TH MSC 62C1062F1562E10

keywords PoissonnoiseconditionalmedianL1lossBayesianestimationpriordistributionsconjugatepriorsgammadistributionmomentconditions

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper studies Bayesian estimation of a Poisson parameter under $L^1$ loss, where the optimal estimator is the conditional median. It claims that any increasing function $f$ that is constant after some point and satisfies a summability bound can be made the exact conditional median $\mathrm{med}(X|Y=y)=f(y)$ by an appropriate prior supported on $\{f(i-1)\}$. For affine estimators $f(y)=ay+b$ with $00$, the construction yields a family of priors distinct from the gamma conjugate priors that still produce exactly linear conditional medians. The existence argument matches moment equalities for an auxiliary distribution, then reweights by $e^w$ and passes to a total-variation limit. The authors present this as the first explicit non-gamma family with the affine conditional-median property under Poisson noise, and as the first general realizability theorem for arbitrary increasing conditional medians.

What carries the argument

The load-bearing object is the moment system $E[W^y 1_{W\le f(y)}]=E[W^y 1_{W> f(y)}]$ for all $y\in\mathbb{N}_0$, together with the exponential reweighting $dP_X(w)=e^w dP_W(w)$ that turns those equalities into the conditional-median condition. The construction solves truncated systems using an invertible matrix whose rows encode the split of the support at $f(y)$, obtains a tail bound of the form $P[W\ge w_{i+1}]\le (f(\lfloor i/\kappa \rfloor)/f(i))^{\lfloor i/\kappa \rfloor}$, and then uses $\ell^1$ compactness to extract a limiting probability vector. The summability condition (9) is exactly what makes the reweighted prior $P_X$ have finite total mass.

What would settle it

Solve the truncated linear system $A_M p = e_1$ for $f(y)=ay+b$ with $a=b=0.3$ and several values of $M$, then compare the exact tail probability $P[W\ge w_{i+1}]$ with the claimed bound $(f(\lfloor i/\kappa \rfloor)/f(i))^{\lfloor i/\kappa \rfloor}$; if the bound fails in exact arithmetic for some $i$, the equality step behind Lemma 3 is false, while if it holds across many $M$ and $i$, the missing atom term is likely controllable.

Watch

Extended reading notes

Core claim

The central claim is that prescribing a conditional median under Poisson noise is equivalent to solving the moment system $E[W^y 1_{W\le f(y)}]=E[W^y 1_{W> f(y)}]$ on an auxiliary random variable $W$ supported on $\{w_i=f(i-1)\}$. Once a probability law $P_W$ satisfies these equalities for every $y$, the reweighted measure $dP_X(w)=e^w dP_W(w)$ is a prior whose Poisson posterior median is exactly $f(y)$. The paper constructs such a $P_W$ by solving finite truncated linear systems in the atom probabilities, using an induction lemma for the existence of solutions, bounding the tail mass with a Markov-type inequality, and then taking a subsequential $\ell^1$ limit that preserves all moment equalities. In the linear case the construction gives a non-gamma prior with $\mathrm{med}(X|Y=y)=ay+b$ for $0<a<1/e$ and $b>0$, and finite-$M$ approximations are shown numerically to converge to the target median.

Load-bearing premise

The construction depends on treating 'W is at least f(y)' as interchangeable with 'W is strictly greater than f(y)' in the moment equalities, but the two events differ by the probability mass placed exactly at f(y), and the proof does not show that this mass term is negligible.

Editorial extensions

If this is right

For any increasing $f$ satisfying condition (9), a Poisson-noise experiment can be given a prior whose $L^1$-optimal estimator is exactly $f$, so estimator design becomes a direct choice of the estimator rather than a search over priors.
Affine conditional medians with slope $0<a<1/e$ and intercept $b>0$ are attainable by non-gamma priors, giving the first explicit counterpart to the gamma-prior characterization for conditional means.
The proof gives a concrete numerical recipe: solve the finite linear system $A_M p = e_1$ and reweight by $e^{w_i}$ to approximate the desired prior to any chosen accuracy.
Because Poisson posterior medians are always nondecreasing, the monotonicity assumption on $f$ is not an artificial restriction, and the theorem covers every estimator shape that the noise model can in principle produce.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct numerical test of the tail bound on the exact truncated solutions, say for $f(y)=0.3y+0.3$, would show whether the atom of mass placed exactly at $f(y)$ is negligible; this is the step the proof leaves implicit.
The same moment-matching construction should transfer to other integer-valued observation models, such as binomial or negative-binomial noise, where conjugate priors also fail to produce affine conditional medians; only the summability condition would need reworking.
The slope restriction $0<a<1/e$ is presented as sufficient, so the true admissible range may be wider; the finite-truncation solver could be used to probe larger slopes numerically.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Read the Barnes-Dytso-Poor paper on linearity-inducing priors for Poisson L1 estimation. The main idea is good: they try to show that you can build a prior supported on {f(i-1)} so that the conditional median of X given Y is any prescribed increasing f, with an affine corollary for a < 1/e. The finite-dimensional part is genuinely new and works: Lemma 2's construction of probability vectors satisfying the truncated moment equalities is correct, and Lemma 1's equivalence between the moment conditions and the desired median is sound. The a < 1/e threshold in Corollary 1 also matches the summability condition.

The problem is Lemma 3. Equation (40) replaces the indicator {W >= f(floor(i/kappa))} with {W <= f(floor(i/kappa))} by invoking the moment condition (24). But (24) equates expectations over {W <= f(y)} and {W > f(y)}, not over {W >= f(y)}. The support contains the atom w_{y+1}=f(y), so the two expectations differ by p_{y+1} w_{y+1}^y. That term is positive in general, e.g., the affine case with M=3, y=1 gives p_2=3/13. So the tail bound (36) is not proven. Everything downstream in Section III-C—the l1 Cauchy argument, the total mass lower bound, the error term Delta, and the properness of PX in Section III-D—uses that tail bound. Without it, the subsequential limit may not be a probability measure or may not satisfy the moment identities. So Theorem 1 and Corollary 1 do not follow from the submitted proof.

Is this fatal? It's a load-bearing error, but not obviously a false result. The construction is plausible, and the gap might be repairable with a sharper argument that handles the atom or a different tightness route. As written, though, the main theorems are unproven.

The writing is clear, the literature coverage is appropriate (the gamma-prior conditional-mean result and the Gaussian linearity results are properly cited), and I don't see any circular reasoning or data fitting. The paper is worth a serious referee, but that referee should be asked to focus on whether Lemma 3 can be fixed.

For you: if you work on L1 estimation or Poisson channels, it's worth a quick look to see if the gap suggests a fix. But don't cite it as a proven result yet. I'd send it to peer review, with a clear request for a major revision.

Best.

Referee Report

3 major / 4 minor

Summary. This paper considers Bayesian Poisson parameter estimation under L^1 loss, where the optimal estimator is the conditional median. The authors claim that for any increasing function f satisfying a summability condition, there exists a prior P_X supported on {f(i-1)} such that med(X|Y=y)=f(y) for all y; specializing to f(y)=ay+b with 0<a<1/e yields a family of priors distinct from gamma that achieve an affine conditional median. The proof proceeds by finite-dimensional moment matching (Theorem 3), a tail bound (Lemma 3), and a compactness/limiting argument to pass to infinite support.

Significance. If correct, the main theorem would be a substantive contribution to Bayesian L^1 estimation: it would provide the first explicit non-conjugate family with the affine conditional-median property under Poisson noise and a general realizability result for arbitrary increasing conditional medians. The paper also gives an algorithm for truncated approximation and numerical illustrations. However, the proof as presented contains a central error in Lemma 3 that invalidates the limiting argument; the claimed results are therefore not established by this manuscript.

major comments (3)

[Section III-B, Lemma 3, Eq. (40)] The equality in Eq. (40) is invalid. The moment condition (24) (or (35)) states E[W^y 1_{W≤f(y)}] = E[W^y 1_{W>f(y)}] for y=⌊i/κ⌋, but step (40) replaces the indicator 1_{W≥f(y)} with 1_{W≤f(y)}. Because the support contains w_{y+1}=f(y), the two events differ by the atom {W=f(y)}. For the affine example f(i)=0.3i+0.3, M=3, y=1, i=3, κ=2, the probability vector from Theorem 3 has p_2=3/13, and the two sides of (40) are approximately 0.356 and 0.240, so the equality is false. Consequently the tail bound (36) is not proven.
[Section III-C and Section III-D] The unproven tail bound (36) is the load-bearing component of the infinite extension. It is invoked at (42) to make the truncated vectors tight, at (47) and (50) to show the limiting total mass is one, at (58)-(60) to control the residual Δ in the moment transfer, and at (63) to prove that P_X is a proper distribution. Since (36) is not established, the ℓ1 compactness argument and the conclusion of Theorem 2 do not follow; consequently Theorems 1 and Corollary 1 are unsupported.
[Section III-C, Eq. (49)] With the indexing w_i=f(i-1), the identity sum_{i=1}^{j+1} P_W(w_i) = 1 - P(W≥w_{j+1}) is off by one; the correct relation is sum_{i=1}^{j+1} P_W(w_i) = 1 - P(W≥w_{j+2}). The subsequent bound (50) should use a tail bound for w_{j+2}. This is secondary to the failure of Lemma 3 but suggests that the indexing in the limit arguments needs careful revision.

minor comments (4)

[Theorem 2] Theorem 2 states f:N0→R0; this should be f:N0→R+.
[Lemma 1] In Lemma 1, the measure dP_X(w)=e^w dP_W(w) is defined up to normalization; the paper should explicitly state that P_X is the normalized version and that the normalization constant cancels in the posterior ratio.
[Corollary 1 proof] The proof of Corollary 1 writes the sum in (14) with an inf over κ≥1 without explaining why the inf appears; the convergence argument should be expanded.
[Figure 3] Figure 3 plots conditional medians for truncated approximations; the caption should clarify that these are finite-M approximations and not the limiting prior distribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the prior is constructed by solving moment equations equivalent to the stated median, and self-citations are contextual only.

full rationale

Reviewing the derivation chain, I find no circular step. The target function f enters as a theorem input: the support W is deliberately chosen as {f(i-1)}, but existence of a probability vector solving the moment equations (24)/(35) is a genuine linear-algebraic result (Lemma 2/Theorem 3), not an assumption of the conclusion. Lemma 1 reduces the median condition to equivalent moment equations; the paper then solves those equations and converts back via (16). This is a standard inverse-problem construction, not a fitted input renamed as prediction. The summability condition (9) is exactly what is needed to make the limiting measure and the reweighted PX proper (Eqs. (62)-(63)), so its use is a sufficient-condition check, not circularity. Self-citations [9] and [12]-[14] are contextual (gamma-mean characterization and prior linear-median work) and do not carry the Poisson median construction. The skeptical issue with Lemma 3, Eq. (40), is a potential proof error: condition (24) equates the indicators {W≤f(⌊i/κ⌋)} and {W>f(⌊i/κ⌋)}, not {W≥f(⌊i/κ⌋)}, so the atom at w_{⌊i/κ⌋+1}=f(⌊i/κ⌋) may make the displayed equality false. That is a correctness gap in the submitted argument, not a circularity: the paper does not assume the tail bound, it attempts to derive it from the moment system. Accordingly, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters are fitted to data; a and b are user-chosen targets, and kappa is a technical parameter in a sufficient condition. The paper introduces no new physical or mathematical entities beyond a family of prior distributions. The main burden is the proof gap in Lemma 3, not hidden fitting or circularity.

assumptions (4)

standard math The conditional median, defined through the quantile function, is the minimizer of expected absolute error (Eq. 1).
This is the decision-theoretic starting point of the paper; it is standard but unproved here.
domain assumption The Poisson likelihood with the 0^0=1 convention (Eq. 2) is the noise model.
All results are specific to this likelihood; the proof uses the factorial structure of x^y e^{-x}.
domain assumption Bayes estimators under Poisson noise are nondecreasing in y (cited to [17]), which motivates restricting f to increasing functions.
Used in the discussion after Theorem 1; not proved in this paper.
standard math Sequential compactness and completeness of probability vectors in l1, used to extract the limiting distribution in Section III-C.
Standard functional analysis fact; no proof given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Linearity-Inducing Priors for Poisson Parameter Estimation Under $L^{1}$ Loss." pith.science (2026). https://pith.science/paper/ICIN5PLH

@misc{pith2026250521102,
  author       = {Pith},
  title        = {Pith review of: Linearity-Inducing Priors for Poisson Parameter Estimation Under $L^1$ Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ICIN5PLH}},
  note         = {Machine review of arXiv:2505.21102}
}

abstract

We study prior distributions for Poisson parameter estimation under $L^1$ loss. Specifically, we construct a new family of prior distributions whose optimal Bayesian estimators (the conditional medians) can be any prescribed increasing function that satisfies certain regularity conditions. In the case of affine estimators, this family is distinct from the usual conjugate priors, which are gamma distributions. Our prior distributions are constructed through a limiting process that matches certain moment conditions. These results provide the first explicit description of a family of distributions, beyond the conjugate priors, that satisfy the affine conditional median property; and more broadly for the Poisson noise model they can give any arbitrarily prescribed conditional median.

Figures

Figures reproduced from arXiv: 2505.21102 by the authors.

**Figure 2.** Comparison of the gamma cdf (red), which induces a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Conditional medians induced by the distribution in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Functional uniqueness and stability of Gaussian priors in optimal L1 estimation
cs.IT 2025-11 reject novelty 6.0 of 10

Near-linearity of the optimal L1 median is claimed to force the prior close to Gaussian in Hermite coefficients, under strong decay assumptions; the L2 stability rate is re-derived.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · cited by 1 Pith paper

[1]

The practical limits of photon communication,

R. McEliece, E. Rodemich, and A. Rubin, “The practical limits of photon communication,” Jet Propulsion Laboratory Deep Space Network Progress Reports, vol. 42, pp. 63–67, 1979

work page 1979
[2]

Capacity of a pulse amplitude modulated direct detection photon channel,

S. Shamai, “Capacity of a pulse amplitude modulated direct detection photon channel,” IEE Proceedings I (Communications, Speech and Vision), vol. 137, no. 6, pp. 424–430, 1990

work page 1990
[3]

Poisson communication theory,

S. Verdú, “Poisson communication theory,” International Technion Com- munication Day in Honor of Israel Bar-David , vol. 66, 1999

work page 1999
[4]

On the capacity of the discrete-time poisson channel,

A. Lapidoth and S. M. Moser, “On the capacity of the discrete-time poisson channel,” IEEE Transactions on Information Theory , vol. 55, no. 1, pp. 303–322, 2008

work page 2008
[5]

Properties of the support of the capacity-achieving distribution of the amplitude-constrained Pois- son noise channel,

A. Dytso, L. Barletta, and S. Shamai (Shitz), “Properties of the support of the capacity-achieving distribution of the amplitude-constrained Pois- son noise channel,” IEEE Transactions on Information Theory , vol. 67, no. 11, pp. 7050–7066, 2021

work page 2021
[6]

Capacities and optimal input distributions for particle-intensity channels,

N. Farsad, W. Chuang, A. Goldsmith, C. Komninakis, M. Médard, C. Rose, L. Vandenberghe, E. E. Wesel, and R. D. Wesel, “Capacities and optimal input distributions for particle-intensity channels,” IEEE Transactions on Molecular, Biological and Multi-Scale Communications, vol. 6, no. 3, pp. 220–232, 2020

work page 2020
[7]

A theoretical analysis of neuronal variability,

R. B. Stein, “A theoretical analysis of neuronal variability,” Biophysical Journal, vol. 5, no. 2, pp. 173–194, 1965

work page 1965
[8]

The variable discharge of cortical neurons: Implications for connectivity, computation, and information coding,

M. N. Shadlen and W. T. Newsome, “The variable discharge of cortical neurons: Implications for connectivity, computation, and information coding,” Journal of Neuroscience, vol. 18, no. 10, pp. 3870–3896, 1998

work page 1998

Show all 17 references

[9]

Estimation in Poisson noise: Properties of the conditional mean estimator,

A. Dytso and H. V . Poor, “Estimation in Poisson noise: Properties of the conditional mean estimator,” IEEE Transactions on Information Theory, vol. 66, no. 7, pp. 4304–4323, 2020

2020
[10]

Conjugate priors for exponential fami- lies,

P. Diaconis and D. Ylvisaker, “Conjugate priors for exponential fami- lies,” The Annals of Statistics , vol. 7, no. 2, pp. 269–281, 1979

1979
[11]

Characterization of conjugate priors for discrete exponential families,

J.-P. Chou, “Characterization of conjugate priors for discrete exponential families,” Statistica Sinica, vol. 11, pp. 409–418, 2001

2001
[12]

L1 estimation in Gaussian noise: On the optimality of linear estimators,

L. P. Barnes, A. Dytso, and H. V . Poor, “ L1 estimation in Gaussian noise: On the optimality of linear estimators,” in Proceedings of the 2023 IEEE International Symposium on Information Theory (ISIT), June 2023, pp. 1872–1877

2023
[13]

L1 estimation: On the optimality of linear estimators,

L. P. Barnes, A. Dytso, J. Liu, and H. V . Poor, “ L1 estimation: On the optimality of linear estimators,” IEEE Transactions on Information Theory, vol. 70, no. 11, pp. 8026–8039, 2024

2024
[14]

Multivariate priors and the linearity of optimal bayesian es- timators under Gaussian noise,

——, “Multivariate priors and the linearity of optimal bayesian es- timators under Gaussian noise,” in Proceedings of the 2024 IEEE International Symposium on Information Theory (ISIT) , July 2024, to appear

2024
[15]

On conditions for linearity of optimal estimation,

E. Akyol, K. Viswanatha, and K. Rose, “On conditions for linearity of optimal estimation,” IEEE Transactions on Information Theory , vol. 58, no. 6, pp. 3497–3508, 2012

2012
[16]

Bounds for the difference between median and mean of gamma and Poisson distributions,

J. Chen and H. Rubin, “Bounds for the difference between median and mean of gamma and Poisson distributions,” Statistics & Probability letters, vol. 4, no. 6, pp. 281–283, 1986

1986
[17]

Monotonicity of Bayes estimators,

P. Nowak, “Monotonicity of Bayes estimators,” Applicationes Mathe- maticae, vol. 4, no. 40, pp. 393–404, 2013

2013

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

The practical limits of photon communication,

R. McEliece, E. Rodemich, and A. Rubin, “The practical limits of photon communication,” Jet Propulsion Laboratory Deep Space Network Progress Reports, vol. 42, pp. 63–67, 1979

work page 1979

[2] [2]

Capacity of a pulse amplitude modulated direct detection photon channel,

S. Shamai, “Capacity of a pulse amplitude modulated direct detection photon channel,” IEE Proceedings I (Communications, Speech and Vision), vol. 137, no. 6, pp. 424–430, 1990

work page 1990

[3] [3]

Poisson communication theory,

S. Verdú, “Poisson communication theory,” International Technion Com- munication Day in Honor of Israel Bar-David , vol. 66, 1999

work page 1999

[4] [4]

On the capacity of the discrete-time poisson channel,

A. Lapidoth and S. M. Moser, “On the capacity of the discrete-time poisson channel,” IEEE Transactions on Information Theory , vol. 55, no. 1, pp. 303–322, 2008

work page 2008

[5] [5]

Properties of the support of the capacity-achieving distribution of the amplitude-constrained Pois- son noise channel,

A. Dytso, L. Barletta, and S. Shamai (Shitz), “Properties of the support of the capacity-achieving distribution of the amplitude-constrained Pois- son noise channel,” IEEE Transactions on Information Theory , vol. 67, no. 11, pp. 7050–7066, 2021

work page 2021

[6] [6]

Capacities and optimal input distributions for particle-intensity channels,

N. Farsad, W. Chuang, A. Goldsmith, C. Komninakis, M. Médard, C. Rose, L. Vandenberghe, E. E. Wesel, and R. D. Wesel, “Capacities and optimal input distributions for particle-intensity channels,” IEEE Transactions on Molecular, Biological and Multi-Scale Communications, vol. 6, no. 3, pp. 220–232, 2020

work page 2020

[7] [7]

A theoretical analysis of neuronal variability,

R. B. Stein, “A theoretical analysis of neuronal variability,” Biophysical Journal, vol. 5, no. 2, pp. 173–194, 1965

work page 1965

[8] [8]

The variable discharge of cortical neurons: Implications for connectivity, computation, and information coding,

M. N. Shadlen and W. T. Newsome, “The variable discharge of cortical neurons: Implications for connectivity, computation, and information coding,” Journal of Neuroscience, vol. 18, no. 10, pp. 3870–3896, 1998

work page 1998