Vector Poisson Channel Analysis

• Vector Poisson Channel is a probabilistic model that transforms nonnegative input vectors into integer count outputs via independent Poisson processes modulated by a mixing matrix and dark current.
• It enables joint measurement analysis and system design optimization for applications such as X-ray imaging, document classification, and molecular communication.
• Key advances include closed-form gradients of mutual information and conditions for linear conditional mean estimators, improving sensor scheduling and experimental design.

A vector Poisson channel is a probabilistic model in which a nonnegative input vector is mapped to an integer-valued output vector via independent but not necessarily identical Poisson processes, often modulated by a mixing matrix and an additive dark-current vector. The channel underlies a broad array of sensing, communication, and inference problems where the observed data comprise counts, such as X-ray imaging, document classification, or molecular communication. Unlike scalar Poisson channels, the vector structure admits rich interaction between input components, enabling not only the modeling of joint measurement effects but also analysis and optimization of system design via information-theoretic or detection-theoretic metrics.

1. Canonical Model and Mathematical Structure

Let $X \in \mathbb{R}_+^n$ be a nonnegative input vector representing, for example, underlying signal intensities of $n$ sources. The output $Y \in \mathbb{Z}_+^m$ is an $m$-dimensional vector of integer-valued counts, observed as the realization of $m$ independent Poisson variables, given by

$$P_{Y|X}(y|x) = \prod_{i=1}^m \mathrm{Pois}\left(y_i; (\Phi x)_i + \lambda_i \right)$$

where $\Phi \in \mathbb{R}_+^{m \times n}$ is a nonnegative mixing or projection matrix encoding the influence of source $j$ on counter $i$, and $\lambda \in \mathbb{R}_+^m$ is the vector of dark-current (background count rates). No restrictions are imposed a priori on the law $p_X(x)$ except those required for integrability and differentiability.

The vector Poisson channel encompasses several specializations, most notably:

• The case $m=n$ and $\Phi$ diagonal yields independent scalar Poisson channels.
• $\Phi$ with nontrivial row structure models aggregate (e.g., joint) sensing or multiplexed measurements.

The observation model encapsulates the essential property that, conditionally on $x$, counts in disjoint time intervals or across disjoint source sets are independent and Poisson-distributed.

2. Information-Theoretic and Detection-Theoretic Metrics

Two principal performance metrics are used to quantify and optimize vector Poisson channels: mutual information and the probability of correct detection in hypothesis testing.

Mutual Information

The mutual information $I(X;Y)$ quantifies the reduction in uncertainty about $X$ given observation $Y$,

$$I(X;Y) = \mathbb{E}[\ln p_{Y|X}(Y|X)] - \mathbb{E}[\ln p_Y(Y)]$$

This criterion is central in experimental design and compressive sensing, as it admits analytic gradients with respect to system parameters and reflects the informativeness of the measurement scheme. For instance, in a two-source, one-counter scenario under a total-time constraint, optimizing time allocations to individual and joint source measurements reveals a paradigm in which aggregate observations (joint) enhance SNR, whereas individual observations enable disambiguation, with a phase-transition in optimal scheduling as prior parameters are varied (Fahad et al., 2022).

MAP Probability of Detection

In discrete settings, such as Bernoulli $X$ and two Poisson sources, optimal detection corresponds to choosing the most probable hypothesis given observed counts,

$\hat{x}(y) = \arg\max_{x} P[X = x] f(y|x)$

$P_d = \sum_{y} \max_{x} P[X=x] f(y|x)$

This detection metric, while closely related to mutual information, can differ quantitatively in the optimal allocation strategy — particularly in regimes where the MAP detector exploits the properties of the Poisson distribution's tail (Fahad et al., 2022).

3. Conditional Mean Estimation and Linearity Criteria

The Bayesian minimum mean-square error (MMSE) estimator for $X$ given $Y$, $E[X|Y]$, plays a central role in estimation theory and sensor design. The paper "The Vector Poisson Channel: On the Linearity of the Conditional Mean Estimator" fully characterizes priors that achieve linear conditional means in the vector Poisson model.

Given $Y \sim \text{Pois}(\Phi X + \lambda)$,

$$E[X|Y=y] = \frac{\int x p_X(x) p_{Y|X}(y|x) dx}{\int p_X(x) p_{Y|X}(y|x) dx}$$

The main result establishes:

• $E[X|Y]$ is linear in $Y$ if and only if $X$ has a product-gamma prior and the dark current $\lambda = 0$.
• Any nonzero $\lambda$ breaks exact linearity for all nondegenerate priors.
• Approximate linearity (in MSE sense) forces the input distribution to be close, in characteristic function norm, to a product-gamma law (Dytso et al., 2020).

This structural property distinguishes the Poisson model sharply from the Gaussian case, where any jointly Gaussian prior leads to an affine estimator irrespective of additive noise mean.

4. Gradient of Mutual Information and Generalized Bregman Divergence

A key advance in the analysis and optimization of vector Poisson channels is the closed-form, matrix calculus-based expression for the gradient of mutual information with respect to system parameters (especially the mixing matrix $\Phi$ and dark current $\lambda$):

$$\frac{\partial}{\partial \Phi_{ij}} I(X;Y) = \mathbb{E} \left[ X_j \ln((\Phi X)_i + \lambda_i) \right] - \mathbb{E} \left[ \mathbb{E}[X_j|Y] \ln( \mathbb{E}[ (\Phi X)_i + \lambda_i | Y] ) \right]$$

This result, presented in "Generalized Bregman Divergence and Gradient of Mutual Information for Vector Poisson Channels," establishes that the gradient can be written as the expectation of a generalized (matrix-valued) Bregman divergence between $X$ and its posterior mean,

$$\nabla_\Phi I(X;Y) = \mathbb{E} \left[ D_{F_P}( X, \mathbb{E}[X|Y] ) \right ]$$

where $D_{F_P}$ is associated to $$F_P(x) = x ( \ln(\Phi x + \lambda) )^T - x \mathbf{1}^T + \mathbf{1}\mathbf{1}^T$$ (Wang et al., 2013).

This framework provides a unifying perspective, as analogous results hold for scalar/vector Gaussian channels with their respective Bregman divergences. The Bregman formalism underpins mirror-descent optimization and characterizes the error landscape exposed by linear versus nonlinear estimation.

5. Experimental Design and Sensor Scheduling

Practical sensing problems, such as two-target detection under a total time constraint, leverage the vector Poisson channel model to address optimal experimental design (Fahad et al., 2022). Key findings include:

• When two sources with known active (rate $\lambda_1$) and inactive (rate $\lambda_0$) states are observed by a time-shared counter, the total observation time is optimally partitioned into individual and joint measurement intervals $(t_1, t_2, t_{12})$.
• The optimal allocation maximizing mutual information or MAP detection probability typically aligns along the symmetry axis $t_1 = t_2 = (T - t_{12}) / 2$, reducing the problem to one-dimensional concave maximization over $t_{12}$.
• The solution exhibits a continuous phase transition: for small prior probability $p$ of source activation, joint sensing dominates; for $p$ near or greater than $1/2$, individual sensing is optimal; high dynamic range (large $\lambda_1/\lambda_0$) broadens the region where hybrid (joint plus individual) schedules are optimal.

This balance captures a fundamental trade-off: joint measurements increase aggregate SNR but suppress source identification, while individual measurements better disambiguate at reduced SNR.

6. Extensions and Algorithmic Applications

The vector Poisson channel framework extends to higher-dimensional and more general settings:

• For $k > 2$ sources, the number of measurement configuration parameters increases combinatorially, but symmetry and concavity properties can often be exploited for optimization, particularly when maximizing mutual information or detection probability (Fahad et al., 2022).
• Unknown rates or non-Bernoulli priors introduce active experiment design challenges, linking to Fedorov's theory of optimal experiments.
• Sequential or adaptive sensor scheduling maps to POMDP formulations, where myopic policies derived from the static analysis offer tractable approximations.

Algorithmically, the closed-form gradient enables gradient-ascent or block-coordinate procedures to optimize the sensing matrix $\Phi$ for compressive-sensing, X-ray imaging, or text analysis applications. Monte Carlo or variational Bayes methods can be used to approximate posterior means and facilitate practical implementation (Wang et al., 2013).

7. Comparative Analysis: Poisson vs. Gaussian Channels

The structural differences between vector Poisson and Gaussian channels are highlighted by the conditions under which linear estimation is optimal:

Channel Type	Linear Estimator Condition	Admissible Priors	Effect of Offset (Dark Current)
Gaussian	$X$ jointly Gaussian	Any mean, any covariance	Affine for any nonzero mean
Poisson	$X$ product gamma, $\lambda=0$	Product gamma	Strictly nonlinear if $\lambda>0$

The Poisson channel exhibits greater structural rigidity: dark current (nonzero $\lambda$) precludes any nontrivial prior from yielding linear estimators (Dytso et al., 2020). In contrast, the Gaussian model's flexibility extends to any mean or covariance structure.

Significance: This distinction informs both statistical inference and experimental design. For Poisson systems, the imposition of product-gamma priors is effectively required for linear-estimator viability, and approximate linearity inexorably draws the underlying law toward the gamma family, as measured by characteristic function proximity.

---

References:

• (Fahad et al., 2022) "Sensing Method for Two-Target Detection in Time-Constrained Vector Poisson Channel"
• (Dytso et al., 2020) "The Vector Poisson Channel: On the Linearity of the Conditional Mean Estimator"
• (Wang et al., 2013) "Generalized Bregman Divergence and Gradient of Mutual Information for Vector Poisson Channels"