Beta-Bernoulli Bayesian Fusion

• Beta-Bernoulli Bayesian Fusion is a conjugate model that uses a Beta prior with Bernoulli likelihoods to combine binary evidence through aggregated success-failure counts.
• It underpins diverse applications such as distributed posterior combination, occupancy-grid mapping, and soft evidence integration in robotic and state estimation tasks.
• By preserving exact conjugacy, the framework enables robust fusion of heterogeneous data sources while mitigating bias from repeated prior usage.

Beta-Bernoulli Bayesian fusion denotes a family of Bayesian constructions in which binary observations, binary latent states, or Bernoulli-like soft evidence are combined through a Beta prior or a beta-process prior. In its finite conjugate form, it is the simplest hierarchical model for a sequence of binary outcomes with an unknown bias; in broader usage, the same pattern appears in distributed posterior combination, algebraic treatments of exchangeable programs, occupancy-grid mapping, outlier-robust filtering, robotic cost modulation, and Bayesian nonparametric feature allocation. Across these settings, the central mechanism is the same: evidence is accumulated in success–failure form, either literally or through pseudo-counts, and fused into a posterior or posterior-like object whose semantics are preserved under appropriate update, marginalization, or reparameterization rules (Staton et al., 2018).

1. Conjugate core and sufficient statistics

From a Bayesian viewpoint, the Beta-Bernoulli model places a Beta prior on an unknown Bernoulli bias $\theta \in [0,1]$,

$$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$

with density

$$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$

and a Bernoulli likelihood for binary observations $x_i \in \{0,1\}$,

$$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$

Given IID observations $x_1,\dots,x_n$, the posterior is again Beta:

$$\theta \mid x_{1:n} \sim \mathrm{Beta}(\alpha+s,\beta+n-s), \quad s=\sum_i x_i.$$

The Beta prior is therefore conjugate to the Bernoulli likelihood, and the sufficient statistics are the counts of successes and failures, $s$ and $n-s$ (Staton et al., 2018).

This conjugate structure is the canonical finite-dimensional form of Beta-Bernoulli fusion. Distinct data sources, repeated trials, or multiple program fragments all contribute through additive updates to the same pair of hyperparameters. The resulting posterior predictive and posterior mean depend only on aggregate success and failure counts, not on the order in which observations arrived. That invariance under reordering is the probabilistic content behind later algebraic and distributed formulations.

In distributed analysis with a shared prior, the same logic appears as prior-corrected posterior combination. If each local analysis yields a sub-posterior $f_c(\theta)$ built from the same prior $$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$0, then the correct fusion target is

$$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$1

For partitioned Bernoulli data with subset sizes $$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$2 and success counts $$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$3, this yields

$$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$4

so the fused posterior is exactly the centralized Beta posterior obtained from pooled counts (Dai et al., 2021). This provides a precise statement of a recurrent issue in Bayesian fusion: when local posteriors share a prior, naive multiplication overuses prior information.

2. Algebraic formulation and program-level fusion

A distinct line of work treats the Beta-Bernoulli process as an algebraic effect rather than merely a conjugate pair. In this formulation, a process is created by a new operation and queried by a get operation. Two implementations are presented: BetaBern, which samples a latent real $$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$5 and returns Bernoulli draws via flip(r), and Polya, which implements Pólya’s urn with reinforcing hidden state $$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$6. De Finetti’s theorem implies that these implementations have the same observable behavior: sequences of outputs are exchangeable and correspond to mixtures of IID Bernoulli sequences with Beta mixing measure (Staton et al., 2018).

The algebraic signature introduces a binding operation for process creation,

$$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$7

and a parameterized Bernoulli choice,

$$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$8

Its measure-theoretic semantics interprets terms as probability kernels. In particular,

$$\theta \sim \mathrm{Beta}(\alpha,\beta), \qquad \alpha,\beta>0,$$9

and

$$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$0

so new integrates over a latent Beta-distributed bias and get performs Bernoulli choice with that bias (Staton et al., 2018).

Within this calculus, fusion is governed by commutativity, discardability, and conjugacy. The formal commutativity axioms include, for example,

$$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$1

and

$$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$2

Discardability includes

$$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$3

when the sampled parameter is unused, and

$$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$4

The conjugacy axiom is

$$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$5

These equations express exchangeability, elimination of unused latent variables, and one-step Bayesian updating in purely syntactic form (Staton et al., 2018).

Their significance for fusion is explicit. Repeated get p calls can be regrouped so that only the number of true and false outcomes matters; independent new calls can be reordered; unused process samplings can be erased. The paper’s normalization into permutation-invariant trees $$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$6 makes sufficient statistics emerge syntactically: leaves depend only on the number of successes, not on their positions. The theory is complete with respect to the measure-theoretic semantics, and also Post-style syntactically complete, so no additional nontrivial program equalities are consistent with the target finite probabilistic semantics (Staton et al., 2018).

3. Distributed fusion, shared priors, and exact posterior combination

In distributed settings, data may be available only through local posteriors or local Monte Carlo output. The abstract problem is to unify sub-posteriors into a single coherent inference. In Bayesian settings with shared priors, the difficulty is that prior information may be present in several local estimators and can be overused unless the fusion rule corrects for it (Wu et al., 2022).

For Beta-Bernoulli models, direct count aggregation solves this problem when sufficient statistics are available. When they are not, the more general Bayesian Fusion framework constructs an exact fusion target on an extended path space. Under a proposal measure $$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$7, one simulates $$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$8 interacting diffusions,

$$p(\theta \mid \alpha, \beta) = \frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\,\Gamma(\beta)} \,\theta^{\alpha-1}(1-\theta)^{\beta-1},$$9

with common coalescence at time $x_i \in \{0,1\}$0, and then corrects by sequential importance weighting in a sequential Monte Carlo algorithm. The endpoint marginal is the exact fusion density $x_i \in \{0,1\}$1 rather than an approximation (Dai et al., 2021).

This construction is especially relevant beyond conjugacy. In the Beta-Bernoulli case, if the local sub-posteriors are already Beta, the fused posterior is analytically available from summed counts. But the framework remains applicable when only pointwise evaluations or samples from $x_i \in \{0,1\}$2 are available, when privacy constraints prevent sharing raw Bernoulli data, or when the local analyses are non-conjugate extensions of a Bernoulli observation model. The paper’s stated motivation is precisely to avoid the degradation that can occur in approximate distributed fusion as the number of analyses increases or the sub-posteriors do not concur (Dai et al., 2021).

A recurring implication is that Beta-Bernoulli fusion separates two operations that are often conflated. One operation is evidence accumulation, which is additive in success–failure statistics. The other is prior accounting, which must be performed exactly once. When those operations are not distinguished, distributed fusion can become systematically biased toward the shared prior.

4. Soft evidence, pseudo-counts, and sequential semantic updates

A more recent usage of Beta-Bernoulli Bayesian fusion treats continuous scores in $x_i \in \{0,1\}$3 as Bernoulli-like evidence through pseudo-counts. In digital-twin-guided robot path planning, each obstacle’s danger state is modeled by a Bernoulli random variable with unknown probability $x_i \in \{0,1\}$4, and a uniform prior is assigned:

$x_i \in \{0,1\}$5

An LLM is used as a semantic sensor that returns a normalized danger score $x_i \in \{0,1\}$6. Rather than binarizing that score, the method interprets it as the mean of $x_i \in \{0,1\}$7 virtual Bernoulli trials and forms pseudo-counts

$x_i \in \{0,1\}$8

followed by the updates

$x_i \in \{0,1\}$9

The posterior mean

$$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$0

then scales a context-aware repulsive gain,

$$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$1

The parameter $$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$2 acts as a trust knob: large $$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$3 pulls the posterior more strongly toward the LLM score, while small $$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$4 leaves the prior more influential (Amani et al., 25 Sep 2025).

This formulation is a direct generalization of ordinary Beta-Bernoulli updating from hard observations to soft semantic evidence. The danger score is not itself a Bernoulli sample; it is encoded as fractional success and failure counts. Because $$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$5 remain positive, the posterior mean remains between $$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$6 and $$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$7, which the paper identifies as a source of numerical stability in chained prompt settings. Sequential fusion is immediate:

$$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$8

so multiple prompts from different workers or at different times accumulate as additional pseudo-observations (Amani et al., 25 Sep 2025).

The same pattern supports a broader interpretation of Beta-Bernoulli fusion as a stable interface between symbolic or semantic scores and downstream numerical decision systems. In the planning example, the planner never interacts directly with the LLM; it only sees the Beta-Bernoulli-updated numeric cost. This suggests a modular role for the model: heterogeneous evidence sources can be reduced to success–failure updates and fused before entering an optimizer, controller, or heuristic search routine.

5. Operational forms in mapping and robust information fusion

In occupancy-grid mapping, each cell is binary—occupied or free—and Bayesian fusion is typically implemented in log-odds form. If $$x_i \mid \theta \sim \mathrm{Bernoulli}(\theta),\qquad p(x_i\mid \theta) = \theta^{x_i}(1-\theta)^{1-x_i}.$$9, then

$x_1,\dots,x_n$0

and for independent observations with likelihood ratio $x_1,\dots,x_n$1,

$x_1,\dots,x_n$2

This additive log-odds rule is the standard Bayesian arm in occupancy mapping and can be read as a reparameterized Beta-Bernoulli update, where counts of “occupied” and “free” evidence are folded into a single scalar state variable (Berlenko et al., 21 Feb 2026).

A significant comparative result concerns the relation between Bayesian fusion and Dempster’s combination rule. Using a pignistic-transform-based fair comparison, the paper reports that under BetP matching, Bayesian fusion is consistently favored, with 15/15 directional consistency and $x_1,\dots,x_n$3, though the absolute differences are small, $x_1,\dots,x_n$4–$x_1,\dots,x_n$5. Under normalized plausibility matching, the direction reverses. The result is therefore explicitly matching-criterion-specific, not a universal dominance statement (Berlenko et al., 21 Feb 2026). This addresses a common misconception: differences between Bayesian and belief-function fusion may reflect parameter matching as much as the fusion rule itself.

A second operational form appears in robust state estimation for networked systems with contaminated measurements. There, each sensor-time pair receives a binary indicator $x_1,\dots,x_n$6 denoting whether a measurement is nominal or an outlier, with

$x_1,\dots,x_n$7

and

$x_1,\dots,x_n$8

The measurement likelihood is switched off when $x_1,\dots,x_n$9:

$$\theta \mid x_{1:n} \sim \mathrm{Beta}(\alpha+s,\beta+n-s), \quad s=\sum_i x_i.$$0

Mean-field variational Bayesian inference then yields posterior expectations $$\theta \mid x_{1:n} \sim \mathrm{Beta}(\alpha+s,\beta+n-s), \quad s=\sum_i x_i.$$1 that directly scale the information contribution of each measurement, so outliers contribute almost no information while nominal measurements contribute normally. Centralized and decentralized information fusion filters are developed from this hierarchical model (Wang et al., 2019).

In this robust-filtering setting, Beta-Bernoulli fusion no longer describes fusion of the physical state directly. Instead, it governs fusion of measurement reliability. The state posterior is conditioned on a second fused posterior over binary trust variables. This two-level construction makes explicit that Beta-Bernoulli fusion is often a gating mechanism for information inclusion, not merely a model for event frequency.

6. Nonparametric extensions, analogues, and scope conditions

The finite Beta-Bernoulli model extends to the beta-Bernoulli process, a Bayesian nonparametric prior for collections of binary-valued features. A draw from the beta process yields an infinite collection of probabilities in the unit interval, and a draw from the Bernoulli process turns these into binary-valued features. In the three-parameter generalization,

$$\theta \mid x_{1:n} \sim \mathrm{Beta}(\alpha+s,\beta+n-s), \quad s=\sum_i x_i.$$2

the discount parameter $$\theta \mid x_{1:n} \sim \mathrm{Beta}(\alpha+s,\beta+n-s), \quad s=\sum_i x_i.$$3 induces Type I and Type II power laws in feature usage, while individual observations remain sparse. Posterior inference can exploit a stick-breaking representation, and the resulting model is used for discrete factor analysis (Broderick et al., 2011).

Scalable inference for these models is itself a fusion problem: global feature probabilities and dictionaries must aggregate evidence from many local sparse activation patterns. In beta process factor analysis, stochastic variational inference confirms that using Gibbs sampling within SVI to maintain certain posterior dependencies is extremely effective, and that approximations able to model intra-local variable dependence perform best. This is a notable contrast with topic-model settings, where other dependency structures are often emphasized (Shah et al., 2015). The implication is that in feature-allocation models, correct fusion is not only about global count sharing but also about preserving local combinatorial dependence among binary indicators.

At the same time, several neighboring literatures use Bernoulli-like fusion without an explicit Beta-Bernoulli conjugate layer. In crowd-sourced HD mapping with B-spline fusion, the authors state that there is no explicit Beta-Bernoulli model, although PMB Bernoulli existence probabilities and Gaussian B-spline fusion are conceptually analogous. In distributed multi-target tracking with multi-Bernoulli filters, generalized covariance intersection fuses Bernoulli existence probabilities and spatial densities, again creating an analogue rather than a literal Beta-Bernoulli update (Xie et al., 2 Mar 2026). These cases are related by structure—binary existence variables and Bayesian combination rules—but they should not be conflated with finite Beta-conjugate Bernoulli fusion.

The formal scope of the algebraic theory is also deliberately limited. In the algebraic-effects treatment, hyperparameters are natural numbers, rational weights are used, terms are first-order, and zero hyperparameters are excluded to avoid degenerate Beta distributions. Real-world probabilistic programming systems may use non-rational weights, arbitrary continuous distributions, higher-order functions, approximate inference, or inspect latent representations directly; any of these can invalidate the exact algebraic equivalences that hold for the abstract process type (Staton et al., 2018). A related practical caveat arises in pseudo-count methods: interpreting a continuous score as $$\theta \mid x_{1:n} \sim \mathrm{Beta}(\alpha+s,\beta+n-s), \quad s=\sum_i x_i.$$4 virtual Bernoulli trials is a modeling decision, not an observed-data identity. This suggests that “Beta-Bernoulli Bayesian fusion” names a precise conjugate mechanism, but its broader applications depend on how faithfully a domain-specific signal can be cast as Bernoulli-like evidence.