Uncertain Maximum Entropy (uMaxEnt)

• Uncertain Maximum Entropy (uMaxEnt) is an inference methodology that generalizes classical MaxEnt by incorporating uncertain, incomplete, and noisy observations via an observation model.
• It reformulates feature constraints into model-dependent terms, resulting in non-convex optimization that is often solved using a generalized expectation-maximization procedure.
• uMaxEnt finds applications in inverse problems, inverse reinforcement learning, and robust uncertainty quantification, balancing minimal bias with real-world observation challenges.

Searching arXiv for recent and foundational papers on uncertain maximum entropy and related formulations. Uncertain Maximum Entropy (uMaxEnt) denotes a family of maximum-entropy inference procedures for settings in which the available information is uncertain, indirect, noisy, partially observed, or aggregated rather than directly expressed as exact empirical feature expectations. In the most explicit contemporary formulation, uMaxEnt retains the standard MaxEnt objective of selecting a distribution with the least amount of information possible while constrained to match available information, but it replaces direct feature-matching by constraints mediated through an observation model $Pr(\omega \mid X)$. This makes uMaxEnt a generalization of both classical MaxEnt and latent maximum entropy, while also introducing model-dependent constraints and loss of convexity (Bogert et al., 2022, Bogert et al., 2023). The literature further suggests that the label has been used in adjacent ways for inverse problems in deterministic systems with uncertain initial conditions, underdetermined distribution reconstruction, and even uncertainty about which entropy functional should be maximized (Montecinos et al., 2017, Das et al., 2015, Gorban, 2012).

1. Conceptual foundation

Maximum entropy inference is commonly justified as a dynamic method for revising degrees of belief in the light of new information, whereas probability theory supplies the static consistency conditions that degrees of belief must satisfy at a given time. In that division of labor, MaxEnt is not a rival to probability theory but a complementary updating rule. Daniel Hunter argued that a complete theory of uncertain reasoning can be obtained by combining maximum entropy inference with probability theory, the former as a dynamic theory and the latter as a static theory (Hunter, 2013).

uMaxEnt arises precisely where classical MaxEnt becomes operationally difficult. Standard MaxEnt assumes that empirical feature expectations can be computed from observations of the model variables themselves. The motivating difficulty in the uMaxEnt literature is that, in many real-world applications that use noisy sensors, computing the feature expectations may be challenging due to partial observation of the relevant model variables. The same issue appears when observations are missing, corrupted, or only indirectly related to the latent variables of interest, and when observational error makes naive empirical constraints physically impossible to satisfy (Bogert, 2021, Bogert et al., 2023).

A recurring theme is that uMaxEnt is intended to preserve the MaxEnt commitment to minimum-bias inference while explicitly accounting for the observation process. This is the point at which the observation model becomes part of the inferential principle rather than an external preprocessing step. A plausible implication is that uMaxEnt should be understood less as a single algorithm than as a general pattern: entropy maximization subject to constraints that themselves encode uncertainty.

2. Formal definition and relation to standard MaxEnt

The classical maximum entropy program over a discrete space $\mathbb{X}$ is

$$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$

subject to normalization and empirical feature constraints

$\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$

Its solution has the familiar exponential-family form

$$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$

In relative-entropy form, the corresponding update maximizes

$$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$

subject to the new constraints (Hunter, 2013, Bogert et al., 2022).

uMaxEnt modifies the feature constraints when $X$ is not directly observed. Let $\omega \in \Omega$ denote observations with empirical frequencies $\tilde{Pr}(\omega)$, and let $Pr(\omega \mid X)$ be a known observation model. Then the uMaxEnt program is

$\mathbb{X}$0

subject to

$\mathbb{X}$1

and

$\mathbb{X}$2

where

$\mathbb{X}$3

The right-hand side is therefore the expectation of the feature under the posterior over $\mathbb{X}$4 given empirical observations $\mathbb{X}$5. The crucial complication is that these empirical feature expectations now depend on the learned model itself, which makes the optimization non-convex (Bogert et al., 2022, Bogert et al., 2023).

The relation among the main variants is compactly summarized as follows.

Principle	Observation regime	Typical solution
MaxEnt	Direct observation of $\mathbb{X}$6	Closed-form (Boltzmann)
Latent MaxEnt	Perfectly observed part $\mathbb{X}$7, hidden $\mathbb{X}$8	EM/approximate
uMaxEnt	Noisy, partial, or indirect observations $\mathbb{X}$9 or $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$0 with $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$1	Generalized EM

The literature is explicit that standard MaxEnt and latent MaxEnt are special cases of the uMaxEnt constraint. When $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$2 and $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$3, uMaxEnt reduces to standard MaxEnt. When $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$4 and $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$5, the same formulation reduces to latent maximum entropy (Bogert et al., 2022).

3. Algorithmic treatment

Because the uMaxEnt constraints depend nonlinearly on $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$6, there is no closed-form maximizer analogous to the ordinary log-linear solution. The standard computational response is an expectation-maximization procedure generalized from latent maximum entropy. In the E-step, one computes expected sufficient statistics under the current model: $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$7 In the M-step, one solves a standard MaxEnt problem with these expected sufficient statistics as the new feature constraints and updates the Lagrange multipliers. Repeating the two steps until convergence yields a fixed point in which the inferred distribution is consistent with both the observed data and the uncertainty induced by the observation process (Bogert, 2021, Bogert et al., 2022).

A second line of work proposes a convex approximation. The approximation is motivated by the desire for a solution independent of any particular observation realization and assumes that the best fit remains log-linear in the features: $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$8 Within that approximation, one alternates between reconstructing feature expectations from the current posterior $$\max_{Pr(X)} \left(-\sum_{X \in \mathbb{X}} Pr(X)\log Pr(X)\right)$$9 and solving a convex MaxEnt problem for the updated parameters. The reported procedure uses random restarts and returns the solution with the highest entropy in order to mitigate local maxima (Bogert et al., 2023).

The same formalism has also been extended to settings where the observation model is supplied by a classifier rather than by direct analytic specification. In that variant, a classifier maps raw observations $\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$0 to labels $\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$1, and one uses estimates of $\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$2 together with $\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$3 to write constraints of the form

$\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$4

This was proposed specifically to simplify the use of sparse, large data sets and to propagate classifier uncertainty into the entropy-maximization stage rather than collapsing observations to hard labels (Bogert, 2021).

4. Deterministic inverse problems and uncertain causes

One influential use of the broader uMaxEnt idea appears in deterministic physical systems with uncertain initial conditions. For classical projectile motion, the direct dynamics are

$\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$5

The inverse problem considered in this setting is to infer the most unbiased distribution over $\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$6 when the average horizontal range $\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$7 is known over many realizations. Using MaxEnt, one maximizes relative entropy

$\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$8

subject to $\sum_X Pr(X)\phi_k(X)=\hat{\phi}_k.$9, which yields

$$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$0

For projectile motion, the geometrically natural prior induced by uniformity in the Cartesian components of velocity gives $$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$1, and the posterior becomes

$$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$2

with $$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$3 after imposing the average-range constraint (Montecinos et al., 2017).

This analysis produces explicit joint and marginal distributions for initial speed and angle, as well as a range distribution that is exponential in $$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$4, decaying over the scale $$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$5. The same study reports that analytical predictions agree excellently with Monte Carlo simulations using a Metropolis-Hastings algorithm, and that the average trajectory over the inferred ensemble is no longer a parabola but exhibits a distinctive long tail. In that setting, outcome statistics are used to invert deterministic laws and estimate the distribution of causes rather than the other way around (Montecinos et al., 2017).

A closely related underdetermined inverse problem appears when an unknown high-dimensional distribution $$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$6 must be inferred from a known lower-dimensional distribution $$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$7 connected by a deterministic map $$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$8. When $$Pr(X)=\frac{e^{\sum_k \lambda_k \phi_k(X)}}{Z(\lambda)}, \qquad Z(\lambda)=\sum_{X'} e^{\sum_k \lambda_k \phi_k(X')}.$$9, many values of $$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$0 map to the same $$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$1, so there is no unique solution. In the discrete case, maximizing Shannon entropy under the consistency constraints yields

$$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$2

which distributes $$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$3 equally among all $$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$4 values mapping to $$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$5. In the continuous case, minimizing relative entropy against a uniform reference yields

$$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$6

This is explicitly described as “spreading out the uncertainty” over each inverse image or fiber of the map $$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$7 (Das et al., 2015).

5. Learning performance, robustness, and limitations

In inverse reinforcement learning with partial observations, uMaxEnt has been instantiated as an extension of maximum causal entropy IRL. The motivating example is apprenticeship learning under environmental occlusion, where the learner observes only noisy or ambiguous traces of the demonstrator. In the reported “fugitive tracking” domain, the uMaxCausalEntIRL algorithm is compared with baselines labeled TRUE, ML, and WOERR/cHiddenDataEM. The paper reports that uMaxCausalEntIRL attains low inverse learning error (ILE) even as noise increases, continues to improve as more trajectories are observed, and approaches the performance of the “infinite observation” control as data grows; the most-likely heuristic does not (Bogert et al., 2022).

A second experimental program studies noisy observation models more generally. On synthetic random models, uMaxEnt is reported to achieve the lowest error to ground truth relative to Most-Likely-x and MaxEnt-MaxEnt, with error decreasing as the number of samples increases. In a large, sparse observation setting based on image recognition with occlusion and large unobservable spaces, the same work reports that uMaxEnt with the true observation function nearly matches an oracle and that a black-box observational model still improves over naive averaging or MaxEnt on noisy features (Bogert et al., 2023).

The literature also contains an important negative result. When a target distribution over $$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$8 binary variables is constrained only by the expectation values of $$H(post,prior)=-\sum_i post(z_i)\log\frac{post(z_i)}{prior(z_i)}$$9 random observables, one may define a biased measure over the compatible “version space” with an entropic bias $X$0, interpolating between the unbiased measure at $X$1 and the pointwise measure concentrated at the maximum entropy distribution as $X$2. For this random, unstructured setting, the mean square distance $X$3 between the target distribution and the center-of-mass distribution over the version space does not vary with $X$4. The conclusion is explicit: the maximum entropy distribution is not closer to the target distribution than any other distribution compatible with the observable values (Obuchi et al., 2015). This is a direct caution against treating MaxEnt as universally superior in every feasible set.

6. Broader interpretations and related entropy-based uncertainty formalisms

The uMaxEnt label has also been used in a different, more foundational sense: uncertainty not in the observations, but in the entropy functional itself. The Maxallent program begins from the observation that there now exists a rich choice of entropies for fitting needs, which creates an “uncertainty of uncertainty.” Rather than selecting one entropy, Maxallent characterizes the set of all conditionally “most random” distributions under a given constraint set. Each element of that set is a maximizer of its own entropy, and the set is described through the Markov order, the partial order under which all continuous-time Markov processes are monotonic (Gorban, 2012).

A related variational approach defines an uncertainty measurement $X$5 by

$X$6

with generalized expectation $X$7 depending on a function $X$8. This construction is used to recover the Shannon entropy, the Tsallis entropy, the entropy for incomplete statistics, and additional nonstandard entropy forms. In that paper, uMaxEnt refers to approaches that seek to infer the “maximizable entropy” consistent with an observed or assigned probability distribution, rather than to infer a distribution from uncertain observations (0810.4345).

Entropy-based treatments of uncertainty also appear in robust stochastic systems and control. One framework measures uncertainty in the noise of a discrete-time stochastic system by the conditional relative entropy of the actual noise law with respect to a nominal white-noise product measure, leading to a balance equation, a dissipation inequality, a superadditivity property, and a minimum-supply Schrödinger bridge problem. Another formulates trajectory optimization under uncertain dynamics through alternating relative-entropy trust-region problems that find the worst-case maximum entropy Gaussian posterior over dynamics parameters and the corresponding robust policy (Vladimirov et al., 2012, Abdulsamad et al., 2021). These works are not identical to the observation-model formulation of uMaxEnt, but they belong to the same broader movement in which entropy and relative entropy quantify admissible uncertainty.

Taken together, these strands show that uMaxEnt has become an umbrella for several closely related but technically distinct projects: generalizing MaxEnt to noisy or partial observations; inverting deterministic systems under uncertain causes; reconstructing high-dimensional distributions from lower-dimensional data; and addressing uncertainty about the entropy concept itself. The common invariant is the MaxEnt ambition to encode all and only the information supplied by constraints while preserving residual uncertainty as entropy.