Superresolution in the maximum entropy approach to invert Laplace transforms

Abstract: The method of maximum entropy has proven to be a rather powerful way to solve the inverse problem consisting of determining a probability density $f_S(s)$ on $[0,\infty)$ from the knowledge of the expected value of a few generalized moments, that is, of functions $g_i(S)$ of the variable $S.$ A version of this problem, of utmost relevance for banking, insurance, engineering and the physical sciences, corresponds to the case in which $S \geq 0$ and $g_i(s)=\exp(-α_i s),$ th expected values $E[\exp-α_i S)]$ are the values of the Laplace transform of $S$ the points $α_i$ on the real line. Since inverting the Laplace transform is an ill-posed problem, to devise numerical tecniques that are efficient is of importance for many applications, specially in cases where all we know is the value of the transform at a few points along the real axis. A simple change of variables transforms the Laplace inversion problem into a fractional moment problem on $[0,1].$ It is remarkable that the maximum entropy procedure allows us to determine the density on $[0,1]$ with high accuracy. In this note, we examine why this might be so.

• The paper presents a maximum entropy solution for Laplace inversion that guarantees convergence through monotonic entropy stabilization and L1 convergence.
• It reformulates the problem as a fractional moment reconstruction, effectively recovering density functions from sparse data.
• Empirical results confirm superresolution, enabling high-quality density recovery even with a limited number of moments.

Superresolution in the Maximum Entropy Framework for Laplace Transform Inversion

Problem Formulation and Background

The paper addresses a classical inverse problem: reconstructing a probability density $f_S(s)$ on $[0, \infty)$ from knowledge of a finite collection of Laplace transform values at points $\{a_i\}_{i=1}^K$. This challenge is highly relevant in applications ranging from exit time distributions in diffusion processes, risk quantification in insurance and finance, and signal processing. The general formulation is as follows: Given $\mathbb{E}[\exp(-a_i S)] = M_i$ for $i = 1, \dots, K$, recover $f_S(s)$. The inversion of the Laplace transform is notoriously ill-posed, especially when only a limited set of transform values is available.

A well-known approach is to recast the problem as a fractional moment problem on $[0,1]$ under the change of variables $y = e^{-s}$, thus translating the Laplace inversion into the task of reconstructing a density $f_Y(y)$ with constraints $\mathbb{E}[Y^{a_i}] = M_i$. Determination of $f_Y$ from its fractional moments is subject to uniqueness conditions, notably characterized by Lin’s theorem, which asserts unique determination under specific conditions on $\{a_i\}$.

Maximum Entropy Solution

The maximum entropy (MaxEnt) principle offers a variational characterization for the solution of such underdetermined inverse problems, seeking the density that maximizes entropy subject to supplied moment constraints. Explicitly, the MaxEnt density has the form:

$$f_K(y) = \frac{1}{Z(\lambda^*)} \exp\left(-\langle \lambda^*, y^a \rangle\right),$$

where $Z(\lambda^*)$ is the normalization factor, $y^a$ denotes the vector of $y^{a_i}$ terms, and $\lambda^*$ is optimized to match the given moment constraints. The formulation leverages convex duality, as elaborated by Borwein and Lewis.

The paper also rigorously discusses the properties of entropy and Kullback-Leibler divergence within this setting, following the formal framework in convex analysis. The entropy function is shown to be concave, upper semi-continuous, and compact on super-level sets, ensuring well-posedness of the variational problem.

Superresolution Phenomenon and Convergence Results

A key empirical observation—supported theoretically in the paper—is the phenomenon of superresolution in the MaxEnt framework: accurate recovery of the underlying density from a surprisingly small number of moments (and hence Laplace transform values). Specifically, numerical evidence shows stabilization of the entropy and small $L^1$ variation in the reconstructed density when the number of moments increases from 4 to 8; in practical tasks, this translates to high-quality density reconstruction from severely undersampled Laplace data.

The main theoretical contribution is a convergence result (Theorem 3.1): Given finite entropy of the true solution, the sequence of MaxEnt solutions $f_K$ constructed from the first $K$ moments satisfies

• Monotonic decrease of entropy $S(f_K)$ to $S(f)$ as $K \to \infty$
• $L^1$ convergence: $\|f_K - f\|_1 \to 0$ The proof exploits properties of the Kullback-Leibler divergence and duality, closing prior gaps in the argument noted by Frontini and Tagliani, and extending results in entropy estimation for truncated moment problems. Moreover, the result demonstrates that entropy-stabilization is both an indicator and a guarantee of solution accuracy.

Implications and Future Directions

The analysis confirms that the maximum entropy approach is highly effective for the Laplace inversion from a minimal set of data, outperforming classical regularization-based and algebraic moment inversion strategies, especially in regimes of sparse information. The capability to achieve “superresolution” has far-reaching implications in practical settings where only partial or noisy data is available, such as risk aggregation, time-to-failure estimation, and reconstruction tasks in physics and engineering.

Theoretical implications include a strengthening of the connection between entropy-based regularization and strong consistency properties in ill-posed inverse problems, suggesting avenues for further research:

• Examination of entropy-constrained inversion in more general families of integral equations, including those arising in machine learning kernels and non-parametric Bayesian inference
• Extension to multidimensional and function-valued moment constraints, leveraging recent advances in high-dimensional convex analysis
• Analysis of robustness under noise and model mismatch, drawing from developments in compressed sensing and uncertainty quantification

Conclusion

The paper presents a thorough analytical and empirical study of superresolution via the maximum entropy method for Laplace transform inversion. It establishes precise convergence guarantees, elucidates the connection between entropy stabilization and density recovery, and demonstrates that MaxEnt-based inversion achieves accurate reconstruction from minimal data. These advances substantiate the practical superiority of the maximum entropy framework for a central class of inverse problems and highlight its utility for applications in risk, finance, and beyond.