Abstract
Compressed sensing (CS) is a signal processing technique that enables the efficient recovery of a sparse high-dimensional signal from low-dimensional measurements. In the multiple measurement vector (MMV) framework, a set of signals with the same support must be recovered from their corresponding measurements. This paper presents the first exploration of the MMV problem where signals are independently drawn from a sparse, multivariate Poisson distribution, a scenario common in microfluidic biosensing. We introduce the Sparse Poisson Recovery (SPoRe) algorithm, which recovers the sparse parameter vector of Poisson rates through maximum likelihood estimation using batch stochastic gradient ascent and Monte Carlo approximations. SPoRe substantially outperforms existing CS algorithms, exhibiting high performance even with one-dimensional measurements and high noise levels. We prove the identifiability of the Poisson model under these lax conditions and validate our findings through simulated experiments.
Introduction
In traditional Compressed Sensing (CS), a fundamental rule of thumb is that the number of measurements ($M$) must exceed the number of non-zero entries ($k$) in the signal ($M > k$). However, in fields like microfluidics, researchers face a physical bottleneck where the number of sensors is severely limited by hardware complexity (often $M < 5$), while the number of analytes can be much higher. Existing algorithms for the Multiple Measurement Vector (MMV) problem fail in this "under-determined" regime because they do not account for the specific statistical structure of the signal. In microfluidics, analytes are captured in partitions according to a Poisson distribution. By leveraging this unique statistical structure, we can achieve "extreme" compressed sensing where recovery is possible even when $M < k$.
Methodology: The MMVP Framework and SPoRe
The authors propose a new framework called Multiple Measurement Vector with Poisson signals (MMVP). Instead of treating the signal as just "sparse," each signal vector is modeled as an independent draw from a multivariate Poisson distribution governed by a sparse rate vector $\lambda^*$.
The core innovation is the SPoRe (Sparse Poisson Recovery) algorithm. SPoRe uses Maximum Likelihood Estimation (MLE) to find the best estimate for $\lambda$. Because the gradient of the log-likelihood involves an intractable sum over all possible integer signal combinations, the algorithm implements batch stochastic gradient ascent enabled by Monte Carlo (MC) approximations. The key insight is that the discrete, statistical nature of Poisson signals is a much stronger constraint than simple sparsity. While standard CS tries to minimize the error between measurement and prediction ($Y - \Phi X$), SPoRe matches the distribution of observed measurements to the expected mixture of Poisson outcomes.
Results and Performance
SPoRe was evaluated against standard CS baselines (such as SOMP and PROMP) in simulated experiments mimicking microfluidic conditions. The results demonstrate unprecedented resource efficiency and robustness.
As shown in Figure 5(a), SPoRe achieves perfect recovery even with a single-dimensional measurement ($M=1$), provided there are enough observations ($D$). This "extreme" recovery is impossible for traditional baseline algorithms, which fail completely when $M < k$. Furthermore, SPoRe exhibits significantly higher tolerance to Additive White Gaussian Noise (AWGN) compared to greedy and convex algorithms (Figure 5b). As the number of observations $D$ increases, the estimation variance of SPoRe approaches the ideal Cramér–Rao bound, proving it is an efficient estimator.
Limitations and Future Work
Despite its performance, SPoRe has computational limitations. The Monte Carlo sampling used to estimate gradients is computationally expensive, which may hinder scaling to very high-dimensional signals ($N > 10^3$). Additionally, SPoRe's performance degrades when the total Poisson rate ($\sum \lambda_n^*$) is very high (Figure 5c), as "good" samples become rare in the Monte Carlo step. Future work will focus on more efficient sampling strategies and real-world validation in microbial diagnostics.
Conclusion
SPoRe represents a significant advancement in compressed sensing by uniquely leveraging Poisson statistics as a prior. By enabling signal recovery in the sensor-constrained $M < k$ regime, this approach opens new possibilities for high-dimensional biosensing where hardware complexity is the primary limiting factor.
Summary updated