Continuous Bayesian Updating

• Continuous Bayesian updating is the recursive revision of probability distributions, using the previous posterior as the new prior to assimilate incoming data.
• It integrates methods like sequential filtering, variational inference, and population-based schemes to manage challenges posed by streaming and non-stationary data.
• Applications span deep neural networks, structural health monitoring, and clinical survival models, ensuring dynamically updated and robust decision-making.

Continuous Bayesian updating is the recursive revision of a probability distribution over latent states, parameters, or models as new information arrives. In its canonical sequential form, the posterior after the $t$-th batch is proportional to the likelihood of the new batch times the previous posterior, so that the previous posterior acts as the new prior: $$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$ (Kochurov et al., 2018). Across the cited literature, the term covers several related regimes: continuous-time filtering, continuous-discrete state-space inference, streaming variational learning, adaptive-memory updating in non-stationary environments, and iterative refinement of posterior quantities such as model evidence when likelihood evaluation is expensive. What unifies these settings is not a single algorithm, but the preservation of Bayesian coherence under repeated incorporation of information.

1. Recursive and continuous-time formulations

The most basic form of continuous Bayesian updating is sequential conditioning. In Bayesian incremental learning for deep neural networks, the update is written exactly as

$$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$

with the posterior approximation from step $t-1$ reused as the prior at step $t$ (Kochurov et al., 2018). This same recursion appears in broader state-space settings, but there the update is usually decomposed into a prediction step and a correction step.

In continuous-discrete filtering, the latent state evolves continuously while observations arrive at discrete times. Between observation times, the predictive density evolves under the Fokker–Planck equation

$\partial_t p_t(x) = \mathcal{L}^\ast p_t(x),$

and at measurement times $t_k$ the exact Bayesian correction is

$p_{t_k}(x) \propto L(y_k\mid x)\, p_{t_k^-}(x).$

Projection-filter methods approximate this infinite-dimensional evolution by restricting the filtering density to a parametric manifold, often an exponential family, and updating parameters rather than arbitrary densities (Emzir, 23 Apr 2025).

A distinct but closely related formulation arises in continuous-time learning problems in which the posterior itself obeys a stochastic differential equation. In a sequential decision model with hidden profitability state $\mu \in \{h,\ell\}$ and noisy cumulative profit $X_t = \mu t + \sigma B_t$, the posterior belief $$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$0 evolves as

$$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$1

Here the signal-to-noise ratio $$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$2 controls the speed of learning, and the optimal decision policy is characterized by posterior thresholds $$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$3 and $$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$4 for exit and expansion (Kwon et al., 2019). This formulation makes explicit that continuous Bayesian updating can be a diffusion on the belief simplex rather than only a sequence of discrete posterior recalculations.

2. Representations and computational schemes

Because exact continuous updating is often intractable, the literature emphasizes structured representations of evolving posteriors. One classical construction represents continuous Bayesian networks by approximating priors, conditionals, messages, and beliefs as finite sums of weighted Gaussians. In that framework, belief propagation becomes repeated multiplication and integration of Gaussian mixtures, with products of Gaussians yielding new Gaussian components and network updates remaining analytic at the level of mixture weights, means, and variances (Driver et al., 2013). This is a density-based representation of continuous updating in graphical models.

For state-space filtering with non-conjugate likelihoods, a recent direction uses information geometry. A 2025 contribution introduces “a simple Riemannian optimization method to be used for the Bayesian update step in the continuous-discrete projection filter for exponential families,” and states that the Bayesian update can be formulated as an optimization problem of $$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$5-Rényi divergence with an easily computed Riemannian gradient; the method is demonstrated on “two highly non-Gaussian Bayesian update problems” (Emzir, 23 Apr 2025). This places the update step itself on a geometric optimization footing rather than requiring analytic conjugacy.

Population-based schemes provide another route. Sequential Tempered MCMC defines a path of intermediate distributions

$$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$6

and alternates importance resampling and MCMC mutation as $$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$7 increases from prior to posterior (Catanach et al., 2018). In that paper, the same framework is used both for Bayesian updating and for reliability assessment, and the Rank-One Modified Metropolis Algorithm is introduced as a mutation kernel that improves performance in high dimensions when the prior strongly constrains the posterior (Catanach et al., 2018).

A closely related reliability-based construction is Bayesian Updating with Structural reliability methods. In its revised form, the driving variable is

$$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$8

with $$p(\theta \mid D_{1:t}) \propto p(D_t \mid \theta)\,p(\theta \mid D_{1:t-1})$$9. For any threshold $$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$0, the conditional law satisfies

$$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$1

and the evidence obeys

$$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$2

This removes the need to pre-specify the original BUS multiplier and allows Subset Simulation to produce posterior samples and evidence estimates once the threshold crosses the admissible regime (DiazDelaO et al., 2015).

A different computational frontier arises when likelihood evaluation is expensive. “Streamlined Bayesian Active Learning Cubature” models the log-likelihood $$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$3 with a Gaussian process, uses the posterior mean $$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$4 to define the plug-in evidence estimator

$$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$5

and stops when a bound based only on GP mean and variance is small. The method treats evidence estimation as Bayesian cubature and yields posterior samples “as a by-product” (Li et al., 14 Sep 2025). This suggests that continuous updating can also proceed by repeatedly refining a surrogate for the likelihood surface rather than by streaming observations alone.

3. Streaming data, forgetting, and adaptive memory

In streaming settings, the central difficulty is not merely recursive conditioning, but how much of the past should remain active when the data-generating process changes. Bayesian incremental learning for deep neural networks addresses the fixed-task, sequential-chunk setting with the variational objective

$$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$6

so that the previous approximate posterior becomes the prior for the next chunk (Kochurov et al., 2018). In experiments on MNIST and CIFAR-10, this sequential posterior propagation avoided the collapse observed under naive fine-tuning on new chunks alone (Kochurov et al., 2018).

For explicitly non-stationary data streams, hierarchical power priors discount the previous posterior before combining it with new data. In that framework, the prior natural parameter for time $$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$7 is

$$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$8

and the latent forgetting parameter is driven by

$$p(w \mid D_1,\dots,D_t) \propto p(D_t \mid w)\, p(w \mid D_1,\dots,D_{t-1}),$$9

in the single-drift case, or by the corresponding sum over parameter blocks in the multi-drift case (Masegosa et al., 2017). The interpretation is explicit: old sufficient statistics are retained only to the extent warranted by the inferred $t-1$0, allowing the model to adapt to concept drift while remaining within a Bayesian variational framework (Masegosa et al., 2017).

Bayes with Adaptive Memory generalizes this idea further by allowing the agent to choose which past data to remember and which to forget. At time $t-1$1, with memory buffer $t-1$2 and readout weights $t-1$3, the BAM prior is

$t-1$4

and the posterior update is

$t-1$5

The memory configuration is chosen by Bayesian model selection with a penalized complexity prior over $t-1$6 (Nassar et al., 2022). This suggests that in non-stationary environments, continuous Bayesian updating is inseparable from adaptive control of the effective prior sample.

4. Normative and interpretive perspectives

One strand of the literature asks not how to compute sequential updates, but which update rules are admissible on decision-theoretic grounds. In a finite-state setting, an updating rule that systematically distorts Bayesian posteriors is written as $t-1$7, where $t-1$8 is the Bayesian posterior and $t-1$9 is a distortion on the belief simplex. Within this class, Bayes’ law is the only strictly Blackwell-monotone updating rule, and when the state is non-binary, Bayes’ law and the trivial dogmatic rule are the only continuous Blackwell-monotone updating rules (Whitmeyer, 2023). In other words, continuity together with the requirement that more information is never worse leaves almost no room for non-Bayesian continuous updating (Whitmeyer, 2023).

A complementary interpretation comes from forecasting contests. In the Kelly-betting formulation, each probabilistic forecast model is treated as a canonical Kelly bettor; bankroll serves as a proxy for Bayesian credibility, and market consensus probabilities and implied model credibilities can be updated in real time as each model updates (Beuoy, 10 Feb 2026). In the multinomial case, the wealth update

$t$0

has a direct analogue to Bayes’ theorem, with $t$1 playing the role of likelihood, $t$2 the role of evidence, and marked-to-market wealth the role of prior credibility (Beuoy, 10 Feb 2026). This gives a real-time operational meaning to continuous model updating: posterior model weights are replaced by continuously rebalanced bankrolls.

The opinion-dynamics literature supplies another interpretation. In a Bayesian framework for opinion updates, agents maintain subjective probability distributions over either a discrete state or a continuous variable, observe neighbors’ communicated actions or probabilities, and update by Bayes’ rule. For continuous verbalization of discrete choice, the update

$t$3

becomes additive in log-odds,

$t$4

while bounded-confidence-type rules emerge as approximations of continuous Bayesian trust weights (0811.0113). This literature makes explicit that continuous Bayesian updating can be continuous in the opinion space even when interactions occur at discrete times.

5. Major application domains

Structural health monitoring provides one of the clearest long-horizon applications. In a deteriorating bridge system, dynamic response data are obtained sequentially from accelerometers, processed by output-only operational modal analysis, and used in a classical Bayesian model updating procedure to learn deterioration parameters over time. The resulting posterior is then propagated into structural reliability, yielding sequentially updated hazard rates and accumulated failure probabilities; these feed a preposterior Bayesian decision analysis for maintenance and repair (Kamariotis et al., 2021). The same paper defines the value of information of SHM as the difference between the expected total life-cycle cost under the optimal prior-based strategy and the expected total cost under the optimal monitoring-based strategy (Kamariotis et al., 2021).

Clinical survival prediction is another application in which dynamic updating is essential. In rapidly changing environments, the paper on COVID-19 mortality prediction distinguishes one-time updating from repeated dynamic updating and compares refitting, recalibration, and Bayesian updating for survival models. The simulation study found that “a dynamic updating process outperformed one-time discrete updating,” and the authors state that “Bayesian dynamic updating has the advantages of making use of knowledge from previous updates and requiring less data compared to refitting” (Tanner et al., 2023). In that work, Bayesian updating is implemented by carrying coefficient estimates and uncertainties forward as priors for the next survival model fit (Tanner et al., 2023).

Engineering reliability problems provide another domain in which continuous updating is not only about parameters but also about rare-event probabilities. Sequential Tempered MCMC is used first to transform the prior over model parameters into a posterior given data and then, treating the posterior as a new prior, to estimate posterior failure probabilities in a water distribution system with unknown leaks and demands (Catanach et al., 2018). This unifies posterior inference and reliability assessment within a single sequential Monte Carlo architecture (Catanach et al., 2018).

These applications show that continuous Bayesian updating is often valuable precisely when model outputs feed downstream decisions. This suggests that the importance of updating accuracy cannot be separated from the propagation of posterior uncertainty into reliability, treatment choice, control, or model selection.

6. Prior sensitivity, robustness, and unresolved issues

Continuous Bayesian updating is not automatically path-independent. When objective or noninformative priors are used, and different data sets depend on the same continuous parameter in different ways, standard sequential updating can become order dependent because the noninformative prior appropriate to one experiment need not match that for another. In the revised method proposed for this setting, one updates not only the cumulative likelihood but also the cumulative Fisher information, using

$t$5

and then constructs the noninformative prior from the combined information rather than from whichever experiment happened to be processed first (Lewis, 2013). This is a direct challenge to the common intuition that sequential Bayes is automatically order invariant under objective priors.

Robustness to prior misspecification is treated explicitly in robust Bayesian inference via interacting Wasserstein gradient flows. There, an ambiguity set of admissible priors is defined by a Wasserstein ball around a nominal prior,

$t$6

and the method computes optimal and worst-case priors within that set together with their corresponding posteriors (Igea et al., 2024). The posterior approximation itself is evolved as a Wasserstein gradient flow toward the Bayesian posterior for the current prior, while the prior is simultaneously deformed within the ambiguity set (Igea et al., 2024). This suggests that robustness in continuous updating may require simultaneous control of posterior dynamics and prior transport.

Finally, several computational and conceptual issues remain recurrent across the literature. Non-conjugate updates motivate geometric optimization and divergence minimization (Emzir, 23 Apr 2025). Expensive likelihoods motivate GP-based active learning and Bayesian cubature (Li et al., 14 Sep 2025). Non-stationary environments require memory management rather than indiscriminate accumulation of all past data (Masegosa et al., 2017, Nassar et al., 2022). And normative analyses show that once continuity and information monotonicity are imposed, Bayes’ rule is unusually rigid (Whitmeyer, 2023). Taken together, these results indicate that continuous Bayesian updating is best understood not as a single procedure, but as a family of recursively coherent inference mechanisms whose feasibility depends on representation, computational approximation, prior design, and the structure of temporal change.