Pith. sign in

REVIEW 3 major objections 4 minor 38 references

State-Dependent Observation Noise Reintroduces Epistemic Value in Linear-Gaussian Active Inference

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Letting observation noise depend on the state restores the epistemic drive in linear-Gaussian active inference.

desk verdict Clean pinning lemma and honest H3' caveat; a genuinely new observation-side boundary for the EFE collapse, though the general epistemic-value claim is conditional and the exact-inference question is left open. read the letter →

arxiv 2607.20306 v1 pith:SO2MXHKQ submitted 2026-07-22 q-bio.NC cs.SYeess.SY

Daniel Corva This is my paper · ORCID
classification q-bio.NCcs.SYeess.SY
keywords activeinferenceexpectedfreeenergyepistemicvaluestate-dependentobservationnoisedualeffectKalmanfilterlinear-Gaussianstate-spacemodelcertaintyequivalence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Active inference is a framework in which agents act to minimize expected free energy, a quantity that balances goal-seeking against expected information gain. Previous work showed that in linear-Gaussian state-space models this information term is constant across policies, so the agent degenerates into a pure goal-seeker with a fixed filter. This paper tries to establish the minimal departure on the observation side that breaks that collapse: let the observation-noise covariance depend on the state, R(x), with the control moving the predicted mean at which R is evaluated. Then the posterior covariance and filter gain become action-dependent, no fixed linear-Gaussian filter can reproduce the agent's beliefs, and under a mild visibility condition—automatic for scalar observations—the per-step information gain is non-constant across policies. If right, it locates curiosity in the sensor model: an agent explores exactly when its actions change how well it expects to see.

What carries the argument

The load-bearing object is the state-dependent observation noise covariance R(x) evaluated at the predicted belief mean μ⁻, the first-order Gaussian filter's evaluation rule. The action-dependence enters through μ⁻: a policy moves the mean, the mean selects the noise level, and the noise level sets the innovation covariance S = CΣ⁻Cᵀ + R(μ⁻), the posterior covariance Σ⁺ = Σ⁻ − Σ⁻CᵀS⁻¹CΣ⁻, and the gain. The per-step epistemic value has the closed form ε = ½ ln det(CΣ⁻Cᵀ + R(μ⁻)) − ½ ln det R(μ⁻), which depends on policy precisely through μ⁻; this identity is what the constancy collapse exploits when R is fixed and what the paper exploits when R varies. The supporting mechanism is a pinning/in

What would settle it

Run the paper's scalar example (A = B = C = Q = 1, prior variance 1, R(x) = 1 + x²) and compare one-step actions u = 0 and u = 2: the claim predicts posterior variances 2/3 and 2/7 and one-step epistemic values of ½ ln 3 ≈ 0.55 and ½ ln(7/5) ≈ 0.17 nats. Alternatively, exhibit any fixed linear-Gaussian filter—a policy-independent schedule of noise covariances—that reproduces the agent's posterior means and covariances on every observation sequence; the theorem asserts none exists, and the paper's own witness treats such an exhibit as a refutation.

Watch

Extended reading notes

Core claim

The paper's central claim is that state-dependent observation noise is a genuine, minimal escape from the linear-Gaussian flattening result. In the standard model—linear dynamics, linear observation map, Gaussian noises, additive control—replace the fixed observation covariance R with a continuous state-dependent R(x), and let the agent run the usual first-order Gaussian filter in which R is evaluated at the predicted mean μ⁻. Because actions shift μ⁻, they shift the innovation covariance S = CΣ⁻Cᵀ + R(μ⁻), and therefore the posterior covariance and the gain. A pinning lemma shows that any candidate linear-Gaussian filter that reproduces the agent's posterior covariance at one step must be u

Load-bearing premise

The load-bearing premise is that the agent's beliefs are correctly described by the standard first-order Gaussian filter—R evaluated at the predicted mean and a Gaussian posterior covariance maintained throughout—rather than by the exact non-Gaussian posterior under state-dependent noise; if exact Bayesian inference is demanded, the flattening notion, the pinning lemma, and the theorem would all need to be re-derived, a step the paper explicitly leaves open.

Editorial extensions

If this is right

  • A fixed observation noise R is effectively a commitment to a certainty-equivalent agent: the epistemic term is constant, so any active-inference model that keeps R constant has an information drive no policy can actually exercise.
  • No fixed linear-Gaussian filter, even one permitted a policy-independent schedule of noise covariances chosen in advance, can reproduce the beliefs of an agent with reachable state-dependent observation noise: the model class cannot be flattened.
  • For scalar observations, reachable non-constancy of R alone is sufficient for non-constant epistemic value, because the epistemic map is strictly monotone in the noise variance; no extra visibility condition is needed.
  • The existence of a planning reduction, the stepwise constancy of epistemic value, and the absence of a second-order dual effect are equivalent in this model class, so reintroducing epistemic value and reintroducing the dual effect are the same act.
  • The boundary of the linear-Gaussian collapse sits on the observation model: only the fact that R depends on a controllable state is needed to make information gain action-sensitive—not nonlinear dynamics, not multiplicative control, not a modified objective.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: if this theorem generalizes to the hierarchical, nonlinear settings where state-dependent sensory precision is already used as a model of attention, then attention and information-seeking would be the same mechanism whenever precision sits downstream of controllable states—something the paper raises as a conjecture.
  • Inference: a natural companion result would treat nonlinear observation maps h(x) with fixed noise—the other classic source of the dual effect; an analogous pinning argument should show that it too blocks flattening, completing the observation-side anatomy.
  • Inference: if an agent has to learn R(x) from experience rather than receive it, curiosity could emerge as a by-product of fitting the noise model, since estimating where the sensor is reliable automatically makes some actions more informative than others—a testable prediction about learning dynamics.
  • Inference: the paper's horizon-1 numbers (an information pull of about 1.7 nats toward the informative cue versus a pragmatic gradient of about 4.5 nats) make a concrete quantitative prediction that a two-step planner should eventually take the detour, turning the claimed local epistemic force into a global one.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper studies a linear-Gaussian state-space model in which the observation noise covariance is a state-dependent function R(x) and the agent uses a first-order Gaussian filter that evaluates R at the predicted mean (Section 3.2). Its main result (Theorem 1, Section 3.3) states that under positive definiteness, full observation rank, and reachable non-constancy of R, the posterior covariance and Kalman gain depend on the policy; no fixed linear-Gaussian reduction (Definition 1) exists; and under an additional informational non-constancy assumption H3' the per-step epistemic value of the expected free energy is non-constant across policies. Corollary 1 removes H3' for scalar observations. Corollary 2 relates planning reductions to the absence of a second-order dual effect. The paper also presents cpomdp, an archived software witness that raises an IncompatibleLinearizationError for such models, and two demonstrations.

Significance. If established, the paper provides a clean, observation-side counterexample to the Koudahl et al. collapse: a model that remains linear-Gaussian in dynamics, observation map, and noise family, yet whose agent has action-dependent information gain. The proof is concise and transparent; Lemma 1's pinning argument is the right tool, and the executable witness with check gates is a genuine reproducibility asset. The paper is also careful to flag its own limitations (Remarks 2–3, Discussion), which is commendable. The main caveats are the scope of the formal result to the first-order Gaussian filter and the conditional nature of part (iii) for n_o>1.

major comments (3)
  1. [Theorem 1(iii), Eq. (10), H3' in §3.3] Theorem 1(iii) is, for n_o>1, a restatement of the assumption H3' rather than a derivation. H3' is defined as the existence of policies whose per-step epistemic values differ at the first separating time, and the proof says only that 'H3' says precisely that these two quantities differ'. The scalar Corollary 1 gives a genuine sufficient condition, and Remark 6 gives a Loewner-comparability condition, but neither is the theorem's general hypothesis. Since Example 1 shows H3 does not imply H3', the abstract's phrase 'under a mild non-degeneracy condition' is accurate only if H3' is itself accepted as that condition — but H3' is the claimed phenomenon. I recommend stating part (iii) as an explicit conditional, or proving a nontrivial sufficient condition on R and C that implies H3'.
  2. [§3.2 and Remark 2; Discussion limitations] The central claim is proved for the first-order Gaussian filter that evaluates R at the predicted mean μ^-, not for the exact non-Gaussian posterior of the generative model (6). Remark 2 explicitly leaves open whether the exact conditional covariance carries a strict Bar-Shalom–Tse dual effect, and the Discussion's fourth limitation says the distance to the exact filter is unquantified. The title and abstract nevertheless state unqualifiedly that state-dependent observation noise itself reintroduces epistemic value in linear-Gaussian active inference. This matters because a skeptic could attribute the non-constancy to the evaluation rule. My own reading is that the mechanism is probably robust — for R(x)=1+x^2 the exact predictive entropy includes E log R(x), which depends on the controllable mean — but the paper should either prove that or restrict the title/abstract claims to the first
  3. [Definition 1 vs abstract, §1 and §3.2] Definition 1's 'fixed linear-Gaussian reduction' is specifically a Kalman filter with a policy-independent noise schedule (A,B,C,Q,R̄_k). The abstract and Section 1 say 'no fixed linear-Gaussian filter reproduces the agent'. A general linear-Gaussian filter with arbitrary policy-independent gains is not ruled out by the theorem as stated; the theorem rules out the Kalman-filter-with-named-covariance-schedule object. The proof is correct for the stated definition, but the unqualified phrase in the abstract is stronger. Please either broaden the theorem to arbitrary policy-independent linear filters or add the Definition 1 qualifier to the abstract and introduction.
minor comments (4)
  1. [§2.1, §2.2] Minor wording: 'readers from active inference will find Section 2.1 familiar, whilst control theorists Section 2.2' is missing a verb after 'control theorists'.
  2. [Remark 3] The claim that Gaussian smoothing is 'injective, so the smoothed landscape is non-constant exactly when R is' is asserted without proof or reference. It is plausible, but not immediate for arbitrary continuous R; please add a short argument or citation.
  3. [§5, T-maze] The T-maze demonstration is not formally covered by Theorem 1, as the paper acknowledges. Consider labeling it explicitly as a numerical illustration rather than an executable version of the theorem, to avoid confusion with the 'witness' language used for the single-chain model.
  4. [Notation, Table 1] The symbol ε_k(π) is used in Theorem 1 before its definition in Table 1; a forward pointer in §3.3 would help.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 1(iii) assumes H3', which is defined as the very non-constancy it 'proves'; the general epistemic-value claim is circular-by-definition, while the scalar corollary and no-flattening result remain genuine.

  1. self definitional [Section 3.3, H3' and Theorem 1(iii); proof of Theorem 1(iii) in Section 4]
    "H3' Informational non-constancy. Grant H3 and let k∗ be the first time at which R differs across the predicted means of two policies. There exist policies π, π′ whose per-step epistemic values differ there: ε_k∗(π) ≠ ε_k∗(π′). ... H3' says precisely that these two quantities differ between π and π′ at k = k∗, so ε_k∗(π) ≠ ε_k∗(π′)."

    H3' is not derived; it is defined as the existence of two policies with differing per-step epistemic values. Theorem 1(iii) then assumes H3' and concludes exactly that existence, with the proof stating 'H3' says precisely that...'. For n_o > 1, the epistemic-value claim is thus a restatement of the assumption, not a derivation from the state-dependent noise mechanism. The paper's own Remark 4 acknowledges the H3/H3' gap, but acknowledgment does not remove the definitional character. The non-circular content is Corollary 1 (scalar monotonicity supplies H3' from H3) and parts (i)-(ii); the abstract's general claim of restored epistemic value under a 'non-degeneracy condition' is the target repackaged as an assumption.

full rationale

The paper's main derivation chain is otherwise self-contained. Lemma 1's pinning argument is proved from the covariance update (8)-(9), and Theorem 1(i)-(ii) genuinely follow from H1-H3 without importing the conclusion. Corollary 1 is a real result: for scalar observations, strict monotonicity makes H3 imply H3', so epistemic non-constancy is established rather than assumed. The cpomdp witness is software built for this paper, but it is not used to prove the theorem, so the self-citation is not load-bearing. The paper also transparently limits the result to the agent's maintained Gaussian recursion rather than the exact non-Gaussian posterior (Remark 2; Section 6 limitations), and I do not count that scope limitation as circularity. The one clear reduction-by-construction is H3': the central general epistemic-value claim of Theorem 1(iii) is true by definition because H3' already asserts the existence of policies with differing per-step epistemic values. This warrants a score of 6 rather than higher because the non-flattening theorem and the scalar case are independent, substantive results.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

No numbers are fitted to data: A,B,C,Q,R(x) are inputs, and H3/H3' are qualitative conditions rather than fitted constants. No new physical or theoretical entities are introduced; cpomdp and IncompatibleLinearizationError are software artifacts, not theoretical entities.

assumptions (6)
  • domain assumption H1: R(x) is continuous and positive definite for every x.
    Guarantees innovation covariances are invertible so the Gaussian updates are defined (Section 3.3).
  • domain assumption H2: the observation map C has full row rank.
    Makes CC^T invertible, so Lemma 1 can conclude Rbar = R(mu^-) from equality of precision updates (Section 3.3; proof of Lemma 1).
  • domain assumption H3: R is non-constant on the reachable set of predicted means.
    Excludes the degenerate case where the control cannot move the noise operating point; used in Theorem 1(i)–(ii).
  • domain assumption H3': there exist policies with differing per-step epistemic values at the first separating time.
    This is the load-bearing premise for Theorem 1(iii); it is essentially the conclusion restated, though the paper labels it a non-degeneracy condition (Section 3.3).
  • ad hoc to paper The agent's inference is the first-order Gaussian filter: R_k = R(mu_k^-), and beliefs are the Gaussian posterior mean/covariance from the standard Kalman update.
    Central modeling choice. The exact posterior under R(x) is non-Gaussian; the theorem characterizes this maintained Gaussian recursion, not exact inference (Section 3.2, Remark 2).
  • standard math Standard closed forms for Gaussian mutual information / EFE and the information-form Kalman update.
    Used in Eq. (4), Eq. (9), and Theorem 1(iii) derivation (Sections 2.1 and 4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of State-Dependent Observation Noise Reintroduces Epistemic Value in Linear-Gaussian Active Inference." pith.science (2026). https://pith.science/paper/SO2MXHKQ

@misc{pith2026260720306,
  author       = {Pith},
  title        = {Pith review of: State-Dependent Observation Noise Reintroduces Epistemic Value in Linear-Gaussian Active Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SO2MXHKQ}},
  note         = {Machine review of arXiv:2607.20306}
}
read the original abstract

Recent work established that under active inference, linear-Gaussian state-space models lose their epistemic drive (any incentive to act so as to gain information) "under any circumstances". The epistemic term of the Expected Free Energy becomes constant: the agent flattens to a Kalman filter whose gain sequence is fixed in advance, regardless of action. The minimal departure that restores the drive is unknown; the only established route is control entering the dynamics multiplicatively; the observation side of this boundary is unexplored. We show that state-dependent observation noise is such a departure: a covariance R(x) that varies with the state x, representing a sensor's accuracy degrading with range. The agent runs the standard first-order Gaussian filter of this literature, R evaluated at the predicted mean. Coupling R(x) to a controllable latent mean makes the posterior covariance, and hence the effective Kalman gain, depend on the action. Consequently, no fixed linear-Gaussian filter reproduces the agent and, under a mild rank condition on the observation map and a non-degeneracy condition on R(x), epistemic value is no longer constant; for scalar observations, reachable non-constancy alone is needed. This is a minimal constructive instance of the Bar-Shalom-Tse dual effect in the agent's maintained covariance: actions now influence the quality of future estimates, not merely the state. Our library cpomdp detects the incompatibility from model specification alone and raises a typed IncompatibleLinearizationError. The theorem ships with an executable witness: exhibiting any fixed filter that reproduced the agent's beliefs would refute both theorem and witness at once. Together this offers a precise, observation-side characterisation of curiosity in a Gaussian agent, bridging dual control and active inference.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 9 canonical work pages

  1. [1]

    On epistemics in expected free energy for linear Gaussian state space models

    Koudahl, M.T.; Kouw, W.M.; de Vries, B. On epistemics in expected free energy for linear Gaussian state space models. Entropy 2021, 23, 1565. https://doi.org/10.3390/e23121565 41

    work page doi:10.3390/e23121565 2021

  2. [2]

    Realising synthetic active inference agents, Part I: Epistemic objectives and graphical specification language

    Koudahl, M.T.; van de Laar, T.; de Vries, B. Realising synthetic active inference agents, Part I: Epistemic objectives and graphical specification language. arXiv 2023, arXiv:2306.08014

    arXiv 2023

  3. [3]

    Realizing synthetic active inference agents, Part II: Variational message updates

    van de Laar, T.; Koudahl, M.; de Vries, B. Realizing synthetic active inference agents, Part II: Variational message updates. Neural Comput. 2024, 37, 38–75. https://doi.org/10.1162/neco_a_01713

    work page doi:10.1162/neco_a_01713 2024

  4. [4]

    Dual effect, certainty equivalence, and separation in stochastic control

    Bar-Shalom, Y.; Tse, E. Dual effect, certainty equivalence, and separation in stochastic control. IEEE Trans. Autom. Control 1974, 19, 494–500. https://doi.org/10.1109/TAC.1974.1100635

    arXiv 1974

  5. [5]

    Nonlinear estimation with state-dependent Gaussian observation noise

    Spinello, D.; Stilwell, D.J. Nonlinear estimation with state-dependent Gaussian observation noise. IEEE Trans. Autom. Control 2010, 55, 1358–1366. https://doi.org/10.1109/TAC.2010.2042006

    arXiv 2010

  6. [6]

    Dual control theory

    Feldbaum, A.A. Dual control theory. I. Avtom. Telemekh. 1960, 21, 1240–1249

    1960

  7. [7]

    Dual control theory problems

    Feldbaum, A.A. Dual control theory problems. IFAC Proc. Vol. 1963, 1, 541–550. https://doi.org/10.1016/S1474-6670(17)69687-3

    work page doi:10.1016/s1474-6670(17)69687-3 1963

  8. [8]

    Active inference on discrete state- spaces: A synthesis

    Da Costa, L.; Parr, T.; Sajid, N.; Veselic, S.; Neacsu, V.; Friston, K. Active inference on discrete state- spaces: A synthesis. J. Math. Psychol. 2020, 99, 102447. https://doi.org/10.1016/j.jmp.2020.102447

    arXiv 2020

Show all 38 references
  1. [9]

    Whence the expected free energy? Neural Comput

    Millidge, B.; Tschantz, A.; Buckley, C.L. Whence the expected free energy? Neural Comput. 2021, 33, 447–482. https://doi.org/10.1162/neco_a_01354

    2021 doi

  2. [10]

    Active inference and epistemic value in graphical models

    van de Laar, T.; Koudahl, M.; van Erp, B.; de Vries, B. Active inference and epistemic value in graphical models. Front. Robot. AI 2022, 9, 794464. https://doi.org/10.3389/frobt.2022.794464

    2022

  3. [11]

    Dual effect, certainty equivalence, and separation revisited: A counterexample and a relaxed characterization for optimality

    Derpich, M.S.; Yüksel, S. Dual effect, certainty equivalence, and separation revisited: A counterexample and a relaxed characterization for optimality. IEEE Trans. Autom. Control 2023, 68, 1259–1266. https://doi.org/10.1109/TAC.2022.3151189

    2023

  4. [12]

    A counterexample in stochastic optimum control

    Witsenhausen, H.S. A counterexample in stochastic optimum control. SIAM J. Control 1968, 6, 131–

    1968

  5. [13]

    A generalization of the Kalman filter for models with state-dependent observation variance

    Zehnwirth, B. A generalization of the Kalman filter for models with state-dependent observation variance. J. Am. Stat. Assoc. 1988, 83, 164–167. https://doi.org/10.1080/01621459.1988.10478582

    1988

  6. [14]

    State-dependent Kalman filters for robust engine control

    Dutka, A.S.; Javaherian, H.; Grimble, M.J. State-dependent Kalman filters for robust engine control. In Proceedings of the 2006 American Control Conference, Minneapolis, MN, USA, 14–16 June 2006; pp. 1185–1190. https://doi.org/10.1109/ACC.2006.1656378

    2006 arXiv

  7. [15]

    Revisiting active perception

    Bajcsy, R.; Aloimonos, Y.; Tsotsos, J.K. Revisiting active perception. Auton. Robots 2018, 42, 177–196. https://doi.org/10.1007/s10514-017-9615-3

    2018 doi

  8. [16]

    The belief roadmap: Efficient planning in belief space by factoring the covariance

    Prentice, S.; Roy, N. The belief roadmap: Efficient planning in belief space by factoring the covariance. Int. J. Robot. Res. 2009, 28, 1448–1465. https://doi.org/10.1177/0278364909341659

    2009 doi

  9. [17]

    Belief space planning assuming maximum likelihood observations

    Platt, R.; Tedrake, R.; Kaelbling, L.P.; Lozano-Pérez, T. Belief space planning assuming maximum likelihood observations. In Proceedings of Robotics: Science and Systems VI, Zaragoza, Spain, 27–30 June 2010. https://doi.org/10.15607/RSS.2010.VI.037

    2010 doi

  10. [18]

    Motion planning under uncertainty using iterative local optimization in belief space

    van den Berg, J.; Patil, S.; Alterovitz, R. Motion planning under uncertainty using iterative local optimization in belief space. Int. J. Robot. Res. 2012, 31, 1263–1278. https://doi.org/10.1177/0278364912456319

    2012 doi

  11. [19]

    Attention, uncertainty, and free-energy

    Feldman, H.; Friston, K.J. Attention, uncertainty, and free-energy. Front. Hum. Neurosci. 2010, 4, 215. https://doi.org/10.3389/fnhum.2010.00215

    2010 arXiv

  12. [20]

    Perceptions as hypotheses: Saccades as experiments

    Friston, K.; Adams, R.A.; Perrinet, L.U.; Breakspear, M. Perceptions as hypotheses: Saccades as experiments. Front. Psychol. 2012, 3, 151. https://doi.org/10.3389/fpsyg.2012.00151

    2012 arXiv

  13. [21]

    Working memory, attention, and salience in active inference

    Parr, T.; Friston, K.J. Working memory, attention, and salience in active inference. Sci. Rep. 2017, 7, 14678. https://doi.org/10.1038/s41598-017-15249-0 43

    2017 doi

  14. [22]

    Introducing a Bayesian model of selective attention based on active inference

    Mirza, M.B.; Adams, R.A.; Friston, K.; Parr, T. Introducing a Bayesian model of selective attention based on active inference. Sci. Rep. 2019, 9, 13915. https://doi.org/10.1038/s41598-019-50138- 8

    2019 doi

  15. [23]

    Codes on graphs: Normal realizations

    Forney, G.D. Codes on graphs: Normal realizations. IEEE Trans. Inf. Theory 2001, 47, 520–548. https://doi.org/10.1109/18.910573

    2001 doi

  16. [24]

    Factor graphs and the sum-product algorithm

    Kschischang, F.R.; Frey, B.J.; Loeliger, H.-A. Factor graphs and the sum-product algorithm. IEEE Trans. Inf. Theory 2001, 47, 498–519. https://doi.org/10.1109/18.910572

    2001 doi

  17. [25]

    An introduction to factor graphs

    Loeliger, H.-A. An introduction to factor graphs. IEEE Signal Process. Mag. 2004, 21, 28–41. https://doi.org/10.1109/MSP.2004.1267047

    2004 arXiv

  18. [26]

    The factor graph approach to model-based signal processing

    Loeliger, H.-A.; Dauwels, J.; Hu, J.; Korl, S.; Ping, L.; Kschischang, F.R. The factor graph approach to model-based signal processing. Proc. IEEE 2007, 95, 1295–1322. https://doi.org/10.1109/JPROC.2007.896497

    2007

  19. [27]

    Simulating active inference processes by message passing

    van de Laar, T.W.; de Vries, B. Simulating active inference processes by message passing. Front. Robot. AI 2019, 6, 20. https://doi.org/10.3389/frobt.2019.00020

    2019

  20. [28]

    Variational message passing and local constraint manipulation in factor graphs

    Şenöz, İ.; van de Laar, T.; Bagaev, D.; de Vries, B. Variational message passing and local constraint manipulation in factor graphs. Entropy 2021, 23, 807. https://doi.org/10.3390/e23070807

    2021 doi

  21. [29]

    RxInfer: A Julia package for reactive real-time Bayesian inference

    Bagaev, D.; Podusenko, A.; de Vries, B. RxInfer: A Julia package for reactive real-time Bayesian inference. J. Open Source Softw. 2023, 8, 5161. https://doi.org/10.21105/joss.05161

    2023 doi

  22. [30]

    Generalised filtering

    Friston, K.; Stephan, K.; Li, B.; Daunizeau, J. Generalised filtering. Math. Probl. Eng. 2010, 2010, 621670. https://doi.org/10.1155/2010/621670

    2010 doi

  23. [31]

    Kalman filters as the steady-state solution of gradient descent on variational free energy

    Baltieri, M.; Isomura, T. Kalman filters as the steady-state solution of gradient descent on variational free energy. arXiv 2021, arXiv:2111.10530

    2021 arXiv

  24. [32]

    Optimal Filtering; Prentice-Hall: Englewood Cliffs, NJ, USA, 1979

    Anderson, B.D.O.; Moore, J.B. Optimal Filtering; Prentice-Hall: Englewood Cliffs, NJ, USA, 1979. 44

    1979

  25. [33]

    Linear Estimation; Prentice-Hall: Upper Saddle River, NJ, USA, 2000

    Kailath, T.; Sayed, A.H.; Hassibi, B. Linear Estimation; Prentice-Hall: Upper Saddle River, NJ, USA, 2000

    2000

  26. [34]

    Bayesian Filtering and Smoothing, 2nd ed.; Cambridge University Press: Cambridge, UK, 2023

    Särkkä, S.; Svensson, L. Bayesian Filtering and Smoothing, 2nd ed.; Cambridge University Press: Cambridge, UK, 2023. https://doi.org/10.1017/9781108917407

    2023 doi

  27. [35]

    cpomdp, version 0.4.2; Zenodo, 2026

    Corva, D. cpomdp, version 0.4.2; Zenodo, 2026. https://doi.org/10.5281/zenodo.21429863

    2026 doi

  28. [36]

    pymdp: A Python library for active inference in discrete state spaces

    Heins, C.; Millidge, B.; Demekas, D.; Klein, B.; Friston, K.; Couzin, I.D.; Tschantz, A. pymdp: A Python library for active inference in discrete state spaces. J. Open Source Softw. 2022, 7, 4098. https://doi.org/10.21105/joss.04098

    2022 doi

  29. [37]

    Active inference and epistemic value

    Friston, K.; Rigoli, F.; Ognibene, D.; Mathys, C.; Fitzgerald, T.; Pezzulo, G. Active inference and epistemic value. Cogn. Neurosci. 2015, 6, 187–214. https://doi.org/10.1080/17588928.2015.1020053

    2015

  30. [147]

    https://doi.org/10.1137/0306011 42

    doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.