acceptodds
Under review as a conference paper at ICLR 2027

Towards Stabilising Hybrid Filtering via Bounded Residual Corrections

Abstract

Filtering of dynamical systems remains a central concern in science and engineering. Classical model-based algorithms such as the Extended Kalman Filter offer interpretability and stability guarantees under regularity conditions, but can perform poorly under model misspecification. To compensate, hybrid methods combine model-based approaches with neural corrections, preserving some interpretability inherent in classical filters. However, naïve embedding of neural mappings risks driving hybrid filters to instability, a severe vulnerability in safety-critical applications. Aiming to resolve this instability without severely impairing expressivity or interpretability, we propose Bounded Residual Correction (BRC), an operationally filter- and architecture-agnostic framework that augments model-based representations with norm-bounded learned residual corrections. We then prove that to preserve Reif et al. (1999)'s probability-one error boundedness when correcting local linearisation-based estimates with learned residuals, bounding them is necessary and constraining their norms below a threshold is sufficient alongside nominal regularity conditions. Empirically, across extensive evaluations, self-supervised BRC-augmented filters outperform prior works in accuracy and probabilistic calibration, confirming that despite constraining residuals' expressivity, norm-bounding can enhance rather than hinder performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Dynamical Systems, Filtering, Kalman Filter, Model Misspecification, Hybrid Filtering, Stability Guarantees

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7yabOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper adds a neural residual correction to model-based filters such as the EKF, and bounds the norm of that correction. The authors show that if the bound is small enough, the EKF’s stability guarantee still holds. Experiments on a linear system, the Lorenz attractor and NCLT show lower error and better calibration than KalmanNet-style hybrid filters, and BRC does not diverge where some baselines do.

Critical strengths
  1. The method is simple and easy to add to existing filters. The correction is also easy to inspect, since it is just an offset from the nominal estimate.
  2. The formalized neural residual correction has a theoretical guarantee.
Critical weaknesses
  1. I am not sure how much the theory adds. Bounded perturbations keeping a stable EKF bounded feels close to standard ISS-type results. \bar{c} cannot be computed, so in practice c_max is still tuned by hand, and the bound is local and does not go to zero. It would help if the authors said clearly what is new compared with existing EKF perturbation analysis.

  2. The experiments are small. All systems have 3 to 6 states. In the Lorenz in-distribution test, the test and training sets share the same ground-truth trajectory, and only the noise is different. NCLT uses one segment with an 800-step test set. That seems too little to back the claims about safety-critical use.

  3. The related work is fairly old, mostly 2016–2021. Recent work on learned observers with Lyapunov or contraction guarantees is not discussed, and it matters for the claim of being “the first formal stability guarantee for hybrid filters”.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7yab (1/2)

We thank the reviewer for their attention. Our responses to their comments are below; all references to texts within our manuscripts, cited using §/Eq. numbers, are to the original manuscript unless mentioned:

Weakness 1

  • Significance of the theorem. As mentioned in §1 (contribution 2) and as we explain in response to the reviewer's comment about related work, to our knowledge, this is the first formal stability guarantee for hybrid filters. It serves the following purposes: (1) It validates norm-bounding as an approach to stabilising hybrid filters (Common Comment, [D]). (2) Along with the cmax⁡c_{\max} sweep of §4.2, it isolates the conditions at least one of which must have failed whenever the filter diverges, with the remedy of reducing cmax⁡c_{\max} (Common Comment, [E]). (3) It is the more fundamental result, from which the corollary on bounded misspecification follows (Common Comment, [A]). Further, the modifications that learned residuals induce in the Riccati equation and in the cross-terms of the stochastic Lyapunov analysis are, to our knowledge, studied here for the first time (our response to Reviewer mwWf, "Novelty of Method and Theorem").
  • ISS and perturbed EKF analysis. As discussed in the Common Comment (section [C]), ISS results (Sontag, 2008; Jiang and Wang, 2001; Teel et al., 2014) do not yield our theorem, because BRC's corrections are not exogenous perturbations, which these approaches require. The same holds for total-stability results (Khalil, 2002, Chapter 9.2). The most non-trivial steps of extending the EKF stability analysis are in our response to Reviewer mwWf, under "Extension of EKF Stability Analysis".
  • Tuning cmax⁡c_{\max}. We clarify that to learn any hybrid filter, one would have to optimise hyperparameters, and cmax⁡c_{\max} can be optimised in that same process. It does not require any special hand-tuning.
  • Bound being local and not approaching zero. Both are, in general, irreducible limitations stemming from nonlinearity of the dynamical system, where exact inference is intractable (Särkkä, 2013). The EKF's linearisation error (Taylor remainders) can, in general, only be upper-bounded, not eliminated, and the resulting guarantees upper bound the error but do not optimise it. In our work and in Reif et al., locality (i.e. the condition on the initial error) serves to bound this linearisation error, as well as the Jacobians and covariances, as in all EKF stability results (our response to Reviewer pSuV, "Strength of Drift Assumption").

Weakness 2

  • Smallness of System Dimension. Low-dimensional state estimation problems are central to fundamental science and applied engineering, from chaotic dynamics to robot localisation. Model misspecification is as much, if not more, of a concern in them than in high-dimensional applications. Nothing specific to BRC's neural backbones, gated addition (Eq. (1)), or our theorem precludes applicability in higher dimensions.
  • Empirical Evaluation Protocol.
    • In the linear benchmark experiment, true states are stochastic, and therefore differ across the train, validate and test folds.
    • The in-distribution protocol of the Lorenz experiment was, as we had mentioned, in line with Revach et al., 2022. More importantly, since ResRORF (our filter) outperformed all other filters in-distribution, there is no reason other than, possibly, overfitting to its training distribution for ResRORF to perform worse than the other filters out-of-distribution at zero-shot. Finally, we rule out this possibility via the two out-of-distribution evaluations at zero-shot.
    • The NCLT dataset, well-studied in robotics, contains robot trajectories on various different days of the year with significant external changes (temperatures in summer, winter etc.) that impact the robot's dynamics. Each trajectory is, therefore, usually studied separately (Garcia Satorras et al., 2019; Revach et al., 2022). We picked the longest possible contiguous trajectory that satisfied quality requirements and split it chronologically, so that the test fold is held out from training and validation in time (§E.4.1).
  • Safety. Safety-critical use is our motivation (Abstract, §1). We define safety as stability or non-divergence (§1). Towards safety so defined, we contribute a formal stability guarantee (§3), and our experiments additionally provide empirical evidence: BRC-augmented filters remained stable under misspecification where several baselines diverged across diverse settings (§4).
AuthorsOpenReview ↗

Response to Reviewer 7yab (2/2)

Weakness 3

  • Recency of related work.
    • We respectfully note that we discuss several works published beyond 2021. We also evaluate a number of these works as baselines. Specifically, the recent filters from 2024 to 2026 we cover include: model-based (Duran-Martin et al., 2024), data-driven (Dowling et al., 2024; Ghosh et al., 2024; Noren et al., 2026). Our coverage of hybrid filters---most related to our work---includes recent representatives of both families of survey of Shlezinger et al. (2025): task-oriented (Mortada et al., 2025; Shen et al., 2025) and state-space-model-oriented (Straka et al., 2025).
    • The Related Work section (§5), in particular, is organised by estimation objective, separating smoothing, system identification and filtering, for which the works that introduced them are the clearest instances, whereas recent and a more comprehensive coverage of hybrid filters is provided by the survey of Shlezinger et al. (2025) and our introduction (§1) and evaluation (§4, §F) sections.
  • Learned observers with Lyapunov or contraction guarantees, and the claim of a first guarantee.
    • In this paper, we distinguish a formal stability guarantee from a theoretical stability property. A property holds under conditions, some of which concern the filter itself and yet cannot be ensured by the user, since the filter's design gives the user no control over them. A guarantee is one whose conditions concerning the filter, rather than the dynamical system or the nominal model, can all be ensured by the user. We clarify this in §1 of the revised manuscript. This distinction matters in practice, because filter-side conditions that cannot be ensured give a practitioner no assurance in deployment, and no remedy when the filter diverges.
    • For instance, consider Hu et al., 2021, who establish a mean-square error bound (Theorem 1) conditional on a Lyapunov decrease condition in expectation (Eq. (16)) and on ergodicity of the filter-induced error chain (Assumption 1). However, nothing in the filter or its training process is able to ensure these conditions. This is a property and not a guarantee in the above sense.
    • BRC, on the other hand, provides a guarantee. Its two filter-side conditions are that its covariance residuals be positive semi-definite, which can be ensured by construction (e.g. per Algorithm 1), and that its norm-bound remain below a threshold, which the user can achieve by reducing cmax⁡c_{\max} if needed. Even when the threshold cannot be computed, our theoretical results guarantee that such a threshold exists. Once the conditions on the dynamical system and the nominal model---the exogenous conditions---are satisfied, stabilising BRC therefore comes down to aspects the user controls.
    • Among all the works we cite and those in the survey of Shlezinger et al. (2025), none provides a guarantee in this sense. Thus, to our knowledge, ours is the first formal stability guarantee for hybrid filters. We would welcome any pointers to works we may have missed.

References: Attached as a supplemental file ('Review_References.pdf').

Reviewer_7yabOpenReview ↗

Reply to author's responses

Thank you for the response. However, no new experiments were added, and the authors acknowledge that the theoretical conditions serve to isolate failure modes rather than to certify stability, with c_max still tuned as a hyperparameter. I will therefore keep my score.

AuthorsOpenReview ↗

Response to Reviewer 7yab’s Post-Rebuttal Comment: Practical Value of the Theorem and c_max

We thank the reviewer for their response. We respond below on the theorem's practical value and on cmax⁡c_{\max}.

Practical value of the theorem

  • Irreducible limitation: stability cannot, in general, be certified in advance

    • Locality. We have already explained why locality is, in general, necessary for local-linearisation filters such as the EKF; it is an irreducible limitation across the literature (please see our earlier response, "Bound being local and not approaching zero"). Accordingly, stability results for these filters require, in general, a sufficiently small initial error (Reif et al., 1999; La Scala et al., 1995; Boutayeb et al., 1997; Song and Grizzle, 1992; Kluge et al., 2010; Barrau and Bonnabel, 2017). Certifying that the initial error is small enough, however, requires knowledge of the true initial state, which in general is not available.
    • Remaining conditions. Even if the initial state were known, certifying stability in advance, via our theorem as well as the results in the literature (see our response to Reviewer pSuV, "Strength of Drift Assumption"), requires bounds on the Jacobians, Taylor remainders, covariances and noise, which in general cannot be verified in advance. Karvonen et al., 2020 make the same point for the majority of stability analyses of the EKF, UKF and related filters, whose assumptions they find "effectively impossible to verify".
    • Misspecification. Bounded misspecification is necessary for any stability guarantee (§M.4, scalar linear counterexample; Common Comment A), and it is consistent with the premise of hybrid filtering that the nominal model and the data must be sufficiently informative (§1). Without knowledge of the true system, however, the bound cannot be verified either.
    • In advance, at every step. Checking conditions during the run does not change this. The error evolves recursively (Eq. (18)), so at any time tt the guarantee for the steps to come is the same guarantee as at the start, with the current error in place of the initial error. Reif et al., 1999 note in Remark 1 that the bounds on the Jacobians and covariances can be verified "during the state estimation process", since the filter computes them, but this covers only what the filter has computed so far, not the steps the guarantee concerns. The remaining conditions concern the true system (the current error, the linearisation remainders, the noise and the misspecification), and are no more available at step tt than at step 00.
  • What our theorem provides

    • A stability guarantee for hybrid filters, where none existed. To our knowledge, none of the hybrid filters in the survey by Shlezinger et al., 2025, nor those we cite, has a stability guarantee, even under model fidelity (§1, contribution 2).
    • Isolating failure modes. As the reviewer notes, our conditions isolate the failure modes of a BRC-augmented filter: whenever it diverges, at least one of them has failed. What distinguishes this for hybrid filters is that the failure modes become actionable. The EKF's are well studied (Bar-Shalom et al., 2001, Chapter 5.5), and BRC's are remedied by keeping covariance residuals positive semi-definite (e.g. by construction per Algorithm 1) and, if that is ensured, reducing cmax⁡c_{\max}, which recovers the EKF at 00 (Common Comment E; §4.2).

cmax⁡c_{\max} as a hyperparameter

  • Necessity. Norm-bounding is, in general, necessary for a guarantee of bounded errors (§M.3; Common Comment D); i.e. such a guarantee requires a fail-safe bound on learned residuals.
  • Regularisation. Regularisation can encourage small residuals but, in general, cannot enforce a bound.
  • Learning cmax⁡c_{\max}. cmax⁡c_{\max} could itself be learned, but it would then need an upper bound, so that a fail-safe exists, and that upper bound would again be a hyperparameter.
  • Conclusion. Some hyperparameter is always necessary for the fail-safe that the guarantee requires. Ours is cmax⁡c_{\max}.

Additional Reference Karvonen, T., Bonnabel, S., Moulines, E., & Särkkä, S. (2020). On Stability of a Class of Filters for Nonlinear Stochastic Systems. SIAM Journal on Control and Optimization, 58(4), 2023–2049. https://doi.org/10.1137/19M1285974

Reviewer_mwWfOpenReview ↗
Rating
2
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes to add a neural network-based bounded residual correction to the filtering algorithm to improve the stability. The proof is given to show that the estimation error can be bounded under sufficient conditions. Many baseline methods are compared in some selected systems.

Critical strengths

The theory is clear, and the idea is easy to follow. Many baeline methods are compared to demonstrate the proposed method's effectiveness.

Critical weaknesses

To obtain the result, the nominal model's fidelity is still assumed. This seems to deviate from the paper's motivation for the model mis-specification. Moreover, the technical contribution seems incremental since there are other methods that add the residual term.

Questions for authors
  1. The author is suggested to give a preliminary section for those who are unfamiliar with state estimation. Some notations like \mathcal{I}_{t|t-1}, s, a, b, PARTITION(z), should be further explained.
  2. My biggest concern is that the paper is motivated by model misspecification, but the main theorem assumes model fidelity (Assumption 6). Adding a bounded learned residual to a nominal filter is a relatively straightforward design, and the theorem establishes conditional non-destabilization rather than a guarantee for state estimation under model mismatch. The authors should clarify the methodological novelty and explain how the mismatch magnitude, correction bound, and resulting estimation error are related.
  3. The contribution seems to be incremental, as the core design adds norm-bounded learned residuals to a nominal filter. While the stability proof involves substantial technical detail, it appears to extend an existing EKF stability analysis to accommodate these corrections.
  4. It would be helpful to clarify what new methodological or theoretical insight this provides beyond treating the learned residuals as bounded perturbations.
  5. The experimental evaluation seems somewhat limited in terms of system diversity and real-world coverage. Given the paper’s focus on stability, the authors should also systematically vary the degree of model misspecification and examine its impact on estimation error and divergence rates.
  6. The drawback is not discussed.
5 comments
AuthorsOpenReview ↗

Response to Reviewer mwWf (1/3)

We thank the reviewer for their attention to our paper. Our responses to their comments are below; all references to texts within our manuscripts, cited using §/Eq. numbers, are to the original manuscript unless mentioned:

Novelty of Method and Theorem (Weakness, Questions 2, 3, 4)

Method

  • Residual Hybrid Filters: We do not claim that adding a learned residual to a model-based estimate is new in itself. However, hybrid filters of this residual kind are uncommon, and the two we are aware of and reference in our paper do not bound their residuals:
    • Ko et al., 2007 (§2): They present a variant that learns a (nonparametric) Gaussian process residual that is unconstrained.
    • APBM (Straka et al., 2025; §4): APBM carries neural parameters in an augmented state, stacking the specified model state and raw neural weights and biases, which are jointly learned with state estimates via Cubature Kalman filtering (CKF) updates. Because these parameters are not directly observable, CKF's stability condition of covariance boundedness is put at risk (§4.1). Moreover, APBM constrains (shrinks) neural parameters rather than residuals themselves, and does so conditionally. Even then, small parameters do not guarantee small residuals---a large measurement outlier may produce a large network input that small parameters fail to attenuate. APBM's learned residuals are thus not, in general, bounded, and its parameter shrinking was designed for interpretability and yields no stability result.
  • Gain-learning Hybrid Filters: These filters (e.g. KNet and RKNet, §4) represent the more predominant line of hybrid filters (see survey by Shlezinger et al., 2025). They replace analytical model-based gains with unbounded learned gains; a gain does not act additively on the estimate, so these are not residual learning filters.
  • Stability Guarantees: None of the abovementioned filters (including those in the survey by Shlezinger et al., 2025) has a stability guarantee, even under model fidelity. To our knowledge, BRC-augmented local linearisation filters are the only hybrid filters with one (§1, contribution 2). We addressed a specific, open and practically relevant problem with a parsimonious intervention, and thus do not view BRC's simplicity as a shortcoming.
  • Flexibility: BRC is operationally filter- and architecture-agnostic (contribution 1)---it applies to mean-covariance propagating, mixture-of-density and sequential Monte Carlo filters, and to MLPs, RNNs and causal transformers (§4, §A). In contrast, to our knowledge, other hybrid state estimation works are tied to specified model-based estimators. BRC's stability guarantees, empirical performance and flexibility together represent an advance for hybrid filtering as a whole and open a clear path for further research (§6).

Theorem

  • Residuals as Bounded Perturbations The reviewer does not clarify exactly what they mean by 'bounded perturbations'. If this relates to input-to-state stability (ISS) arguments (Sontag, 2008; Jiang and Wang, 2001; Teel et al., 2014), we show that ISS approaches do not apply towards guaranteeing BRC's stability given EKF stability (cf. common comment, [C]). The same arguments also apply to 'Total Stability' results (Khalil, 2002, Chapter 9.2). We therefore do not believe that these approaches trivialise our theorem.

  • Extension of EKF Stability Analysis We do not deny that our work extends EKF stability analysis. Given the principle of nominal stability necessary for perturbed stability (cf. common comment, [D]), it was natural for us to extend Reif et al., 1999's analysis. Two particularly non-trivial steps in establishing BRC stability lay in:

    • Stabilising the BRC-perturbed Riccati equation (Lemma L.11), which we did by applying Weyl's theorem, and
    • Decoupling the cross-terms between the error and the disturbances, including BRC's neural feedback, via Young's inequality (Proposition P.3, Lemma L.15), and keeping the effect of cmax⁡c_{\max} separate from the other terms the theorem is conditioned on, via Jensen's inequality (Lemma L.25).

The first step is used by the second, which is what lays the foundation for isolating conditions for stability, and preserving the EKF's stability conditions (i.e. the smallness of the abovementioned shared terms) as they were in Reif et al.'s theorem.

Thus, methodologically, to our knowledge, there is no other hybrid filter with a bounded, learned residual correction formulation (§1, contribution 1). The modifications that these corrections make to the EKF's Riccati equation and to the cross-terms in its stochastic Lyapunov analysis are, we believe, studied theoretically for the first time in our work.

AuthorsOpenReview ↗

Response to Reviewer mwWf (2/3)

Conditional Non-Destabilisation (Question 2) We discuss the conditional (sufficient-conditions-based) and conservative character our results in the context of existing stability results in section [E.] of the common comment. The fundamental difficulty in establishing necessary conditions or optimality stem from the intractability of exact inference in nonlinear dynamical systems (Särkkä, 2013) and persists across the literature as we mention in section [E.]. Finally, the non-destabilisation nature of the theorem is consistent with the principle of nominal stability necessary if not sufficient for stability under perturbation (please see section [D.] of the common comment).

Relating the Mismatch, Norm-Bound cmax⁡c_{\max} and Error (Question 2)

  • Theorem 1 shows the filter's mean square error to be upper bounded by an asymptotic 'floor' consisting of the function γ(cmax⁡)\gamma(c_{\max}) under model fidelity; the proof of the theorem shows γ(0)=0\gamma(0) = 0 and to be strictly increasing in cmax⁡c_{\max}. To be clear, the original manuscript showed δ\delta as a function of cmax⁡c_{\max}, but we show it to be a constant independent of cmax⁡c_{\max} (please see our 'Change Log' comment "Theorem 1: δ\delta is now a constant" for details). But the relationship of the abovementioned error bound and cmax⁡c_{\max} is observable via γ\gamma. For clarity, we now mention γ\gamma's properties in the theorem's statement as well.
  • This relationship between γ\gamma and the mean-square error upper bound persists under bounded misspecification, as our corollary now shows. It also shows the asymptotic floor to include a function ω(mhmax⁡)\omega(m_h^{\max}), where mhmax⁡m_h^{\max} is a (non-negative) bound on the misspecification. Like γ\gamma, ω\omega satisfies ω(0)=0\omega(0) = 0 and strictly increases with mhmax⁡m_h^{\max}. It is unsurprising that the MSE would increase with misspecification in the worst-case. It is similarly unsurprising that the increasing relationship of cmax⁡c_{\max} with the worst-case MSE would persist, no matter the source of error (noise, nonlinearity or misspecification): in the worst case, neural residuals displace the estimate from the true state by as much as the norm-bound allows. Since misspecification is exogenous (see section [B.] of the common comment), its contribution ω(mhmax⁡)\omega(m_h^{\max}) enters the MSE upper-bound additively, without cross-terms with the cmax⁡c_{\max} contribution γ(cmax⁡)\gamma(c_{\max}).
  • These are worst-case results, and formal results for lower-bounding or optimising filter errors in nonlinear systems remain intractable as mentioned in point # 2 above. We therefore study whether norm-bounding for preserving stability costs performance empirically.
  • The cmax⁡c_{\max} sweep study (§4.2), run under a setting with a fixed degree of misspecification, in particular, shows a non-monotonic profile of the MSE vs. cmax⁡c_{\max} curve on the linear benchmark. We clarify that this is not in contradiction with our theoretical result, as the latter only gives the monotonic relationship in the worst-case. Beyond a point (cmax⁡=100c_{\max} = 100), we observe divergence.
  • Given a nominal model, how much the neural residuals can reduce estimation error under varying degrees of misspecification depends on the capacity of the networks used and on how informative the data is, which is a premise of hybrid filtering (§1) and is studied empirically in the broader Machine Learning literature.
AuthorsOpenReview ↗

Response to Reviewer mwWf (3/3)

Experiments (Question 5)

  • Our evaluation is restricted to established benchmarks previously used in hybrid state estimation (§4.1), so that our baselines, spanning the model-based, hybrid and data-driven paradigms, are compared on their established setups.
  • As we state in §4.1, hybrid filtering is already established as a means of addressing a variety of misspecifications (survey by Shlezinger et al., 2025), including their degree (Revach et al., 2022). We sought to evaluate our contribution within the paradigm: our empirical question is whether norm-bounding learned residuals---our proposed intervention---trades performance for stability.
  • Since misspecification is exogenous (c.f. common comment, [C]), the upper bound on the filter's error increases with cmax⁡c_{\max} at any given degree of misspecification, so this trade-off in cmax⁡c_{\max} can be examined at any given misspecification.
  • We therefore evaluate it across different types of misspecification, rather than degrees: a linear benchmark with dimensional mismatch between the nominal and true state spaces (§4.2), the chaotic Lorenz attractor with transition function mismatch and heavy-tailed measurement outliers (§4.3) and real-world NCLT data with unmodelled road elevation, control input biases and multipath-corrupted GPS (§4.4).
  • In all three experiments, BRC-augmented filters are compared against their nominal filters and against unbounded hybrid filters. In §4.2 and §4.3, we also ablate the neural backbone to isolate its contribution, whereas in §4.4, ablating BRC returns the nominal KF. The number of diverging seeds, where observed, is reported in every table.
  • The Lorenz zero-shot evaluation additionally varies the degree of distribution shift (unseen outlier scales and trajectory shifts, §4.3), and the cmax⁡c_{\max} sweep (§4.2) additionally isolates the effect of the norm-bound, with divergence beyond the threshold.
  • How much a network can reduce error as misspecification grows depends on its capacity and the data, as it does for hybrid filters generally; a systematic account of this is a worthwhile subject for a dedicated study.

Drawbacks (Question 6) We mention several limitations in the manuscript:

  • The assumptions underpinning our theory, those of the theorem (§G) and now also of the corollary (§K.2 of the new manuscript), are by nature limitations of it.
  • Our remarks (§I.4) following the proof of our theorem---e.g. on the conservativeness of our conditions and prospective vs. diagnostic relevance---were also explicit mentions of limitations. We note that these are limitations across the literature and not particular to our work (c.f. [E.] of the common comment) and now move the necessity/sufficiency remark (Remark 3) and conservativeness remark (Remark 4) into §3 in the revised manuscript.
  • Our stability results are for the local-linearisation family of filters only (§3); BRC applies operationally to other families (§A), but their stability remains open (§6).
  • Latent-state non-identifiability (§2) under self-supervision is a limitation for BRC as it is for any learned filter. We provide a mitigative validation procedure (§C), but it requires some ground-truth trajectories, which can be expensive to obtain in real-world applications.

Presentation (Question 1) We thank the reviewer for the suggestion, which we agree would aid self-sufficiency. We rely on Anderson and Moore (1979) as our reference for state estimation throughout the paper (§1, §5 and the proofs), and a preliminary section of our own would add little beyond that standard text. We now explain the notation and steps of Algorithms 1 and 2 more clearly in Figure 4 in §A of the revised manuscript (linked to §2 of the main text). The notations It∣t−1\mathcal I_{t|t-1} and It∣t\mathcal I_{t|t} (neural network inputs) are covered in the paragraph immediately preceding Algorithm 1, and our implementations (§E) give specific examples of some options that may be considered for practical application.

References: Attached as a supplemental file ('Review_References.pdf').

Reviewer_pSuVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The article proposes a stabilization approach for the data driven-model based hybrid state estimation problem. In the hybrid model, mean and covariance of state estimates are decomposed as model based system's (such as Extended Kalman Filter) estimates and data driven (neural network) system's estimates. The neural network's output consists of mean estimate and the rank-1 square root factor of the covariance estimate. The proposed algorithm essentially bounds the two norms of these variables. Then the key result is the theorem that claims that under some assumptions about the data model and the boundedness of residual model outputs, the state tracking is stable. In the numerical experiments section, the authors provide experiments to demonstrate the improvement obtained by their proposed bounded residual approach, especially for the model mismatch (between the generative dynamical model and the model based estimator) cases.

Critical strengths
  • The article addresses an important open problem of stable hybrid (model based+data driven) state tracking problem. The authors propose to bound the mean and covariance part of the residual (data-driven) neural network component. They provide a theorem which claims the stability of this proposed approach.

  • The article has been written in a clear manner in general.

  • The experimental results provided in Section 4 illustrate the relative success of the proposed approach, especially considering the model mismatch cases.

Critical weaknesses
  • The main reason the hybrid models are introduced is to compensate for the unmodelled dynamics (or model misfit). I was expecting that the authors would make some assumption on the misfit and provide a stability result under that condition. However, it seems like the proposed theorem assumes that the model based estimator and the data generation model matches. Therefore, the residual (data-driven) component seems to be only potentially correcting the approximation of the model based part such as extended Kalman filter. So the model mismatch case, which is the main motivation of the hybrid system is not covered in this article. As the article is essentially a theory article, the experimental evidencing in Section 4 is not sufficient for this purpose. This weakens the contribution scope of the article.

  • The article's appendix contains a very long proof of the stability theorem, and this part is critical as the main contribution of the article is this theorem. I believe in its current form, the article seems to be more suitable for a control theory journal. Regarding this proof, the authors make strong assumptions on the boundedness of the estimation error (for example in equations (20) in Hypothesis 1 ). They attempt to justify this assumption, as far as I understand, by an induction based proof: they assume small error at the start. Then they show the monotonic expected decrease condition in (169), which assumes previous errors are constrained to a bounded domain. However, the noise can potentionally take large values (despite the fact that its bounded variance) which can drive the error out of the assumed small domain. This seems to invalidate the inductive structure of the proof.

Following are some notes about the presentation of the article:

Section 2

  • it is better mean and covariance are indicated next to μMB,t∣i\mu_{MB,t|i} and PMB,t∣iP_{MB,t|i} respectively.

  • I was expecting that the authors would provide meaning and motivation of the BRC concept here before providing an example.

  • As a general comment, a diagram describing the state estimation system with the details of the neural network component would be really helpful.

  • I believe Algorithm 1 could be better clarified if the text before Algorithm 1 explains z_\mu, z_P,a,b. I though z_P was vectorized version of covariance matrix (which would be hard to impose positive definitiness), instead it is a rank-1 factor.

Section 3 The authors could provide a accessible and tutorial discussion on how the boundedness of the data-driven component of the state estimator guarantees stability. This could be based on explaining the reasoning chain used in the proof of the theorem.

Section 4 There is a mismatch between the theory in the earlier sections and the examples in this section. As the theory does not allow mismatch between the generative model and the model based part of the estimator, whereas the examples in Section 4 are on the model mismatch.

Questions for authors
  • Regarding Eq. (1), is there an uncorrelatedness assumption about the model based and BRC based state estimates (as theirs covariances are added.)?
  • Is not the drift assumption a strong assumption in practice? I know that it is inherited from Reif et. al (1999) about model based state estimation, but still seemed strong.
  • is not having rank-1 constraint on P_{BRC} too restrictive? what if the unmodelled dynamics is richer than that?
  • is the probability-one boundedness of theorem valid, given that the noise can take arbitrarily large values?
9 comments
AuthorsOpenReview ↗

Response to Reviewer pSuV (1/3)

We thank the reviewer for their thoughtful comments. Our responses to the same are below; all references to texts within our manuscripts, cited using §/Eq. numbers, are to the original manuscript unless mentioned:

Length and Validity of Theorem and Proof (Weakness 2, Question 4)

We kept the proof detailed to make the paper more self-sufficient (it still builds on Reif et al., 1999), easier to check step by step, and potentially reusable for the other filter families we leave to future work (§6). The reviewer's concern, as far as the validity our proof is concerned, appears to be that the 'unbounded' nature of noise terms detracts from the non-positivity of the conditional expectation of the forward difference of Lyapunov function ('CELF')---the reviewer references Eq. (169)---which, as they rightfully point out, is essential to our proof by induction. We address this concern as follows:

  • As a result of the bounded state assumption (Assumption 3), whose strength we discuss further below in our response, the process noise vtv_t is deterministically bounded, as we show in lemma L.17.
  • In contrast, the measurement noise term wtw_t, as the reviewer perhaps points out, is not assumed to be deterministically bounded.
  • However, the Lyapunov candidate VtV_t is a quadratic in the error ete_t (Vt:=et⊤Pt∣t−1etV_t:=e_t^\top P_{t|t}^{-1}e_t).
  • Independently, we show that the sole noise-driven component of ete_t is the term rtr_t of the filter's exact error dynamics, Eq. (18).
  • In lemmas L.22-L.24, by applying Jensen's inequality (in L.22), we show that the contribution of rtr_t towards the CELF is upper bounded by the expectation of the squared norm E[∥rt∥22]\mathbb{E}[\|r_t\|_2^2], consistent with the quadratic relationship of ete_t with VtV_t.
  • In lemma L.17 itself, we use the sub-multiplicity of spectral norms, to show that the contribution of wtw_t towards the upper bound of E[∥rt∥22]\mathbb{E}[\|r_t\|_2^2] arises via the expectation of its square, E[wtwt⊤]=Cov(wt)\mathbb{E}[w_t w_t^\top] = \mathrm{Cov}(w_t).
  • Thus, insofar as the measurement noise wtw_t is concerned, the contraction of the CELF only requires that the covariance of wtw_t be bounded and 'small', which is in fact a condition of the theorem. Specifically, lemma L.25 identifies the corresponding threshold σˉ\bar{\sigma}.

We add that Lemma L.25 secures the non-positivity of the CELF (via the satisfaction of the supermartingale descent condition of Lemma L.24), provided its conditions are met. One of these conditions is that the inductive error bound ϵsup\epsilon_{\text{sup}} is below the threshold ϵˉ\bar{\epsilon}. Thus the inductive hypothesis is indeed assumed to hold until L.25. However, we show subsequently (§I.3) that the limited time-horizon existence---granted by the Theorem's conditions and the inductive hypothesis---of the Jacobian bounds Fˉ,Hˉ\bar{F}, \bar{H} (lemma L.4), covariance bounds pˉ,p‾\bar{p}, \underline{p} (lemma L.11), the linearisation residue bounding constants cφ,cχc_{\varphi}, c_{\chi} (lemmas L.18-L.19) and the abovementioned supermartingale descent condition are, together with regularity conditions, sufficient to obtain a uniform, infinite-horizon error bound (with probability one); this uniform bound is in turn obtained via the invocation of the stopping-time result, Gard (1988, Theorem 5.2), following Reif et al., 1999. The infinite-horizon error bound, in particular, yields a bound on the filter error at the time step immediately outside the inductive hypothesis's temporal window, thus closing the induction (validating the inductive hypothesis for all time t≥0t \geq 0) and granting time-independent validity of all the lemmas.

AuthorsOpenReview ↗

Response to Reviewer pSuV (2/3)

Strength of Drift Assumption (Question 2) As the reviewer notes, the bounded trajectory drift assumption (part 3 of Assumption 3) is inherited from Reif et al., 1999. Given the bounded deterministic solution xt~\tilde{x_t} and state xtx_t (parts 1 and 2 of Assumption 3), the trajectory drift xt~−xt\tilde{x_t} - x_t is bounded; part 3 merely names the constant ϵv\epsilon_v and is not a strong assumption given parts 1 and 2, which are also inherited from Reif et al., 1999.

As for the strengths of parts 1 and 2, to our knowledge, all EKF stability results for discrete-time dynamical systems are conditional upon the existence of the Jacobian bounds Fˉ,Hˉ\bar{F}, \bar{H}, covariance bounds pˉ,p‾\bar{p}, \underline{p}, the linearisation residue bounding constants cφ,cχc_{\varphi}, c_{\chi} and a bound on the process noise (Reif et al., 1999, La Scala et al., 1995, Boutayeb et al., 1997, Song and Grizzle, 1992). Of these, Reif et al. are the only ones to cover general systems with process noise and nonlinear measurement functions; later works such as Kluge et al., 2010 and Barrau and Bonnabel, 2017 cover special cases such as missing observations and Lie-group state spaces. In Reif et al., 1999 and our work, the bounded state and deterministic solution (parts 1 and 2) serve as a basis for deriving for all of these. La Scala et al., 1995 also assume bounded states (part 2) for some of the above bounds. The works differ in which of these bounds they derive from their own assumptions and which they assume to exist directly, and we submit that these assumptions could all be considered strong:

(a) Jacobian bounds: La Scala et al., 1995 derived the transition Jacobian bound from a bounded state and assumed linear observation function; both bounds are directly assumed to exist, along the filter's trajectory in Boutayeb et al., 1997 and globally in Song and Grizzle, 1992.

(b) Linearisation residue bounds: La Scala et al., 1995 derived the transition residue from a bounded state; Song and Grizzle, 1992 derived both bounds via assumed global Hessian bounds; Boutayeb et al., 1997 directly assumed existence of both bounds.

(c) Process noise bound: La Scala et al., 1995 assumed it directly; Boutayeb et al., 1997 and Song and Grizzle, 1992 only cover deterministic systems.

(d) Covariance bounds: We derive the covariance bounds from the point-wise observability rank condition (Assumption 2) and the bounded deterministic solution; Boutayeb et al., 1997 and La Scala et al., 1995 assumes uniform observability (stronger on its own than a point-wise rank assumption) and derive them from it, and Song and Grizzle, 1992 assumed them directly.

Complementary Roles and Scopes of the Theorem and Experiments (Weakness 1, Presentation Note on §4) We clarify that we did not intend for our experiments to verify our theorem. Our experiments evaluate BRC under misspecification, which is the setting hybrid filters target and relevant for applications. As we had hoped to communicate in §4.1, we explain this choice below:

  • The theorem describes one of BRC's properties as a method; we covered its role and scope in the common comment (sections B, D and E); contributions 1 and 4, which our experiments (§4) evidence, are claims in their own right.
  • Contribution 1: our implementations of ResEKF and ResRORF, together with our algorithmic expositions in §A.1, cover BRC's compatibility with multiple nominal filters and neural networks.
  • Contribution 4: the key issue studied is the tension, experienced by hybrid filters in prior work, between expressivity and stability---in principle and as is common in learned systems, greater expressivity via a large norm-bound or unbounded residuals holds the potential to better compensate for misspecification, whereas our theorem guarantees stability only within a threshold.
  • The purpose of our empirical evaluations (§4.1) was therefore to test whether norm-bounding trades off performance for stability.
  • Further, since exact state estimation is intractable for nonlinear systems, formal guarantees for these systems yield bounded errors but not optimality or performance guarantees (Särkkä, 2013); performance metrics like accuracy should, in these cases, be assessed empirically.
  • Empirically, no BRC-augmented filter diverged in any seed on the linear benchmark or the Lorenz attractor, both under misspecification, whereas multiple baselines did (DANSE and APBM; KNet, RKNet and KFormer) (§4, §F).
AuthorsOpenReview ↗

Response to Reviewer pSuV (3/3)

Uncorrelatedness Assumption (Question 1)} We do not assume that the model-based (MB) estimates and BRC's learned residuals are uncorrelated. These norm-bounded residuals may depend on the MB estimates and/or on each other, as in our implementation in §4. Our theorem's proof treats the learned residuals' effect on the error dynamics as a bounded disturbance, whose cross terms with the filter's error are handled pathwise (Young's inequality) using only the norm bound cmax⁡c_{\max}, rather than via independence or uncorrelatedness. Further, the proof assumes the learned covariance residuals to be positive semi-definite, so they inflate the MB covariance estimates, much as the specified QtQ_t and Σt\Sigma_t do in the standard EKF.

Rank-1 Paremeterisation (Question 3) The rank-1 parameterisation of learned covariance residuals is not required. Algorithm 1 only presents it as a possible approach, as we mentioned in §A.2. Our theorem, as such, only requires the learned covariance residuals to be positive semi-definite and norm-bounded, so any rank, up to full rank, is covered by it. In §4.2 and §4.3, we used rank-1 parameterisation because it is more parsimonious in terms of the neural parameter budget and sufficed to produce good results.

Presentation Notes on §2 and §3 We thank the reviewer for their notes. We address them as follows:

  • Mean and covariance are now indicated next to their symbols.
  • We have now added to the start of §2 a paragraph covering the meaning and motivation of BRC.
  • A diagram illustrating the steps of Algorithms 1 and 2 has now been added as Figure 4 in §A, but referenced in §2; the figure's caption and illustration together detail the steps of the Algorithms.
  • We have now added an intuitive description of the chain of reasoning in the proof, following the sketch of proof in §I.1.

References: Attached as a supplemental file (`Review_References.pdf').

Comments on the paper

6 comments
AuthorsOpenReview ↗

Common Comment: Theorem's Reliance on Model Fidelity (1/4)

We thank the reviewers for their thoughtful comments.

We observe that the assumption of model fidelity in the statement of our theorem comes across as counter-intuitive, considering our motivation for hybrid filters as a way to address model misspecification. To address this issue, we present a corollary that extends the result to the bounded model misspecification setting. The theorem's results transfer to the corollary easily under bounded model misspecification. The stability results of Reif et al., 1999, on the other hand, do not transfer trivially under BRC's bounded residual conditions. The theorem is therefore the more fundamental result which is why we chose to focus on it in the original submission. To support this claim, we present the following; all references to texts within our manuscripts, cited using §/Eq. numbers, are to the original manuscript unless mentioned:

[A.] A Summary of the Corollary: what it guarantees, what assumptions it makes, how it follows from the theorem under these assumptions.

[B.] Input-to-State Stability (ISS) Approaches from Theorem to Corollary: as the reviews suggest, existing literature shows potentially alternative approaches from the theorem to the corollary.

[C.] ISS does not transfer Reif et al's results to our Theorem: building on [B.] we show why ISS approaches do not trivialise our theorem.

We explain why we chose to formulate the theorem under model fidelity by making the below arguments:

[D.] The Theorem itself validates Norm-Bounding: we explain how the theorem validates norm-bounding as a useful approach towards stabilising hybrid filters---our main motivation for any theoretical analysis.

[E.] The Theorem itself isolates Failure Modes: we explain why the theorem, together with our cmax⁡c_{\max} sweep (§4.2), isolates the failure modes of a BRC-augmented EKF and points to a practical remedy: reducing cmax⁡c_{\max} to stabilise it.

[F.] Precedents in the Literature: not presenting a formal misspecification result in the original submission is consistent with the literature, not a gap particular to our work.

A. Corollary

  • The corollary, its assumptions and proof are presented in §3 and §K of the revised manuscript.
  • True Model. We consider the general case of a system where the latent state, its dimensionality, transition and measurement functions and noise terms may all be different from the specified model; the measurements, by definition, are the only quantities shared by the nominal and true models.
  • Assumptions. - In exchange for the full relaxation of the model fidelity assumption (Assumption 6), we assume that model misspecification is bounded. - We restate (exchange) parts of two assumptions, bounded state (Assumption 3) and bounded measurement noise (Assumption 4) in terms of the true rather than nominal model.
  • Strength of New Assumptions. Bounded misspecification is a necessary condition for guaranteeing the stability of a BRC-augmented EKF (and an EKF, in general) (we prove this using a scalar linear counterexample in §M.4, in the revised manuscript). Moreover, this assumption is consistent with the premise of the hybrid filtering paradigm: nominal model and data must be sufficiently informative (§1). Finally, the modified assumptions 3 and 4 are no stronger than in the model fidelity case.
  • Nature of Guarantees Offered. Under the conditions the corollary specifies, the estimation errors of the BRC-augmented EKF are (a) bounded with probability one over the infinite horizon and (b) exponentially bounded in the mean-square. These are the same types of guarantees offered by the theorem and Reif et al., 1999.
  • Proof Mechanism. The corollary's proof uses the lemmas and induction closure argument of the theorem either (a) exactly as they are or (b) with substitutions resulting from the changed assumptions 3 and 4, plus an added disturbance due to the bounded misspecification.
AuthorsOpenReview ↗

Common Comment: Theorem's Reliance on Model Fidelity (2/4)

B. Input-to-State Stability (ISS) Approaches from Theorem to Corollary We consider ISS results (e.g., Sontag, 2008; Jiang and Wang, 2001; Teel et al., 2014) that the reviews mention. These results consider a system of the following kind, where ete_t represents the `state' and ξt\xi_t represents an input or perturbation, both of which can be multivariate:

et=Φt(et−1,ξt), for all ξte_t = \Phi_t(e_{t-1}, \xi_t), \text{ for all } \xi_t

Here Φt\Phi_t is a function of et−1e_{t-1} and ξt\xi_t only. In particular, the earlier perturbations ξ1:t−1\xi_{1:t-1} affect ete_t only through et−1e_{t-1}; they do not change the function Φt\Phi_t itself. Φt\Phi_t may depend on tt, and through tt on anything that is fixed in advance (here, the data and the nominal model), but not on the earlier perturbations ξ1:t−1\xi_{1:t-1}.

Each ISS result assumes its own set of regularity conditions. But let us assume that they hold, so that we may apply such a result (as the reviews suggest). Then, if the unperturbed system et=Φt(et−1,0)e_t = \Phi_t(e_{t-1}, 0) is stable in some sense (determined by the particular ISS result applied), the perturbed system et=Φt(et−1,ξt)e_t = \Phi_t(e_{t-1}, \xi_t) is also stable in some sense, for any ξt\xi_t that is bounded.

Now, consider the BRC-augmented EKF (ResEKF)'s error dynamics in the general case where there may or may not be misspecification. Let the measurements y1:Ty_{1:T} (data) be produced by the dynamical system, and let its true data generating model be unknown. Further, let the nominal model be specified as in Eq. (2), with nominal functions and parameters M={f,h,Qt,Σt,μ0∣0,P0∣0}\mathcal{M} = \{f, h, Q_t, \Sigma_t, \mu_{0|0}, P_{0|0}\}. In this general case, the derivation of ResEKF's error dynamics follows exactly the same steps as in the model fidelity case presented originally in §H. We present this in Eq. (225) of §L in the revised manuscript:

et=Etet−1+st+rt+ut+dt, with :e_t = E_t e_{t-1} + s_t + r_t + u_t + d_t, \text{ with }: Et=(I−KtHt)FtE_t = (I - K_tH_t)F_t st=(I−KtHt)φ(zt−1,μt−1∣t−1)−Ktχ(zt,μt∣t−1)s_t = (I - K_tH_t)\varphi(z_{t-1}, \mu_{t-1|t-1}) - K_t\chi(z_t, \mu_{t|t-1}) rt=(I−KtHt)vt~−Ktwt∗r_t = (I - K_tH_t)\tilde{v_t} - K_tw_t^{\ast} ut=−(I−KtHt)μBRC,t∣t−1−μBRC,t∣tu_t = -(I - K_tH_t)\mu_{\text{BRC}, t|t-1} - \mu_{\text{BRC}, t|t} dt=−Ktmthd_t = -K_tm^h_t

where mthm^h_t is the measurement misspecification term arising from the new assumption of bounded misspecification, wt∗w_t^* is the true measurement noise (arises from modified Assumption 4) and vt~:=zt−f(zt−1)\tilde{v_t} := z_t - f(z_{t-1}) is a residual that reduces to process noise under model fidelity, with ztz_t the true state (projected to the nominal state space). Other variables are as defined by the specified nominal model, Eq. (2): wtw_t is the nominal measurement noise and vtv_t is the nominal process noise.

We can reformulate sts_t so that it depends on the state et−1e_{t-1} rather than on the true states zt−1z_{t-1} and ztz_t. Since zt−1=μt−1∣t−1+et−1z_{t-1} = \mu_{t-1|t-1} + e_{t-1}, and the a priori error is et∣t−1=f(μt−1∣t−1+et−1)−f(μt−1∣t−1)+vt~−μBRC,t∣t−1e_{t|t-1} = f(\mu_{t-1|t-1} + e_{t-1}) - f(\mu_{t-1|t-1}) + \tilde{v_t} - \mu_{\text{BRC},t|t-1} (§L of the revised manuscript), so that zt=μt∣t−1+et∣t−1z_t = \mu_{t|t-1} + e_{t|t-1}, we have: st=(I−KtHt)φ(μt−1∣t−1+et−1,μt−1∣t−1)−Ktχ(μt∣t−1+et∣t−1,μt∣t−1),s_t = (I - K_tH_t)\varphi(\mu_{t-1|t-1} + e_{t-1}, \mu_{t-1|t-1}) - K_t\chi(\mu_{t|t-1} + e_{t|t-1}, \mu_{t|t-1}),  with et∣t−1=f(μt−1∣t−1+et−1)−f(μt−1∣t−1)+vt~−μBRC,t∣t−1\text{ with } e_{t|t-1} = f(\mu_{t-1|t-1} + e_{t-1}) - f(\mu_{t-1|t-1}) + \tilde{v_t }- \mu_{\text{BRC}, t|t-1}

The terms that may depend on misspecification are collected in the perturbation ξt:=[mth,vt~−vt,wt∗−wt]\xi_t := [m^h_t, \tilde{v_t} - v_t, w_t^* - w_t].

The ResEKF's estimates use only the given nominal M\mathcal{M} and the given data y1:Ty_{1:T}, and are computed in the same way whether or not ξt=0\xi_t = 0; the filter does not use the true model, only the data y1:Ty_{1:T} already generated by it. Hence the quantities KtK_t, HtH_t, FtF_t, μt−1∣t−1\mu_{t-1|t-1}, μt∣t−1\mu_{t|t-1} and BRC's learned residuals do not depend on the perturbations ξ1:t\xi_{1:t}, and the error dynamics equation above holds for all ξt\xi_t, not only under model fidelity (ξt=0\xi_t = 0). Consequently, EtE_t and utu_t do not depend on ξ1:t\xi_{1:t} at all, while rtr_t and dtd_t depend on the perturbations only through ξt\xi_t, as is seen from their definitions above. So does sts_t, which, in the reformulation above, is a function of et−1e_{t-1}, vt~\tilde{v_t} and the quantities listed above only. The error dynamics are therefore of the ISS form, et=Φt(et−1,ξt)e_t = \Phi_t(e_{t-1}, \xi_t), with Φt(e,ξt):=Ete+st(e,ξt)+rt(ξt)+ut+dt(ξt)\Phi_t(e, \xi_t) := E_t e + s_t(e, \xi_t) + r_t(\xi_t) + u_t + d_t(\xi_t).

Observe that if ξt=0\xi_t = 0 for all tt, i.e. mth=0m^h_t = 0, v~t=vt\tilde{v}_t = v_t and wt∗=wtw_t^* = w_t (as under model fidelity, where additionally zt=xtz_t = x_t and the true system coincides with the nominal model), then these error dynamics reduce to Eq. (18). Then using the theorem's guaranteed stability under model fidelity, one could possibly apply an ISS result to obtain stability in the ISS sense under bounded misspecification (bounded ξ\xi).

AuthorsOpenReview ↗

Common Comment: Theorem's Reliance on Model Fidelity (3/4)

C. ISS does not transfer Reif et al.'s results to our Theorem Assume model fidelity (as the theorem needs), and treat the BRC's learned residuals to be the bounded `perturbations', ξt:=[μBRC,t∣t,PBRC,t∣t,μBRC,t∣t−1,PBRC,t∣t−1]\xi_t := [\mu_{\text{BRC},t|t},P_{\text{BRC},t|t},\mu_{\text{BRC},t|t-1},P_{\text{BRC},t|t-1}], as the reviews suggest. ResEKF's error dynamics, Eq. (18), are then, with ξ1:t:=(ξ1,…,ξt)\xi_{1:t} := (\xi_1, \dots, \xi_t):

et=Et(ξ1:t)et−1+st(ξ1:t)+rt(ξ1:t)+ut(ξ1:t), for all ξ1:te_t = E_t(\xi_{1:t}) e_{t-1} + s_t(\xi_{1:t}) + r_t(\xi_{1:t}) + u_t(\xi_{1:t}), \text{ for all } \xi_{1:t}

Note that the terms depend on the whole history ξ1:t\xi_{1:t}, not on ξt\xi_t alone. The reason is that the learned residuals are fed back into the filter's own recursions (Algorithm 2, Eq. (1)), which carry them forward:

μt∣t−1=f(μt−1∣t−1)+μBRC,t∣t−1\mu_{t|t-1} = f(\mu_{t-1|t-1}) + \mu_{\text{BRC},t|t-1} Pt∣t−1=FtPt−1∣t−1Ft⊤+Qt+PBRC,t∣t−1P_{t|t-1} = F_tP_{t-1|t-1}F_t^\top + Q_t + P_{\text{BRC},t|t-1} Pt−1∣t−1=(I−Kt−1Ht−1)Pt−1∣t−2+PBRC,t−1∣t−1P_{t-1|t-1} = (I - K_{t-1}H_{t-1})P_{t-1|t-2} + P_{\text{BRC},t-1|t-1} Kt=Pt∣t−1Ht⊤(HtPt∣t−1Ht⊤+Σt)−1K_t = P_{t|t-1}H_t^\top(H_tP_{t|t-1}H_t^\top + \Sigma_t)^{-1}

Here the gain KtK_t depends on PBRC,t−1∣t−1P_{\text{BRC},t-1|t-1} through Pt−1∣t−1P_{t-1|t-1}, and so on back to t=1t = 1. Likewise, the Jacobians Ft=Jf(μt−1∣t−1)F_t = \mathbf{J_f}(\mu_{t-1|t-1}) and Ht=Jh(μt∣t−1)H_t = \mathbf{J_h}(\mu_{t|t-1}) are evaluated along estimates that contain all earlier μBRC\mu_{\text{BRC}}. Since Et,st,rtE_t, s_t, r_t and utu_t are built from KtK_t, FtF_t and HtH_t, all four depend on ξ1:t\xi_{1:t}. In contrast, in [B.], the perturbation never entered the filter: EtE_t and utu_t did not depend on it, and rtr_t, sts_t and dtd_t depended only on the current ξt\xi_t. In [B.], for given values of et−1e_{t-1} and ξt\xi_t, the value of ete_t did not change with the earlier perturbations ξ1:t−1\xi_{1:t-1}; here it does, because ξ1:t−1\xi_{1:t-1} changes EtE_t, sts_t, rtr_t and utu_t.

On the other hand, the EKF's error dynamics are exactly as shown below, with ut(0)≡0u_t(0) \equiv 0:

et=Et(0)et−1+st(0)+rt(0), for all ξ1:te_t = E_t(0) e_{t-1} + s_t(0) + r_t(0), \text{ for all } \xi_{1:t}

Comparing the two equations, we see that for any value of ξ1:t\xi_{1:t}, the EKF's error dynamics are fixed by Et(0),st(0)E_t(0), s_t(0) and rt(0)r_t(0). This is consistent with the fact that the EKF operates without the influence of BRC's learned residuals. This means that the EKF's error dynamics coincide with the ResEKF's only when ξt=0\xi_{t} = 0 for all tt. ISS, on the other hand, requires the error dynamics to be et=Φt(et−1,ξt)e_t = \Phi_t(e_{t-1}, \xi_t) with one fixed function Φt\Phi_t for all perturbations ξt\xi_t, so that the earlier ξ1:t−1\xi_{1:t-1} act only through et−1e_{t-1}. Here the map from et−1e_{t-1} to ete_t is itself changed by ξ1:t−1\xi_{1:t-1}, so this premise does not hold.

The ISS argument of [B.], therefore, does not apply here.

D. The Theorem itself validates Norm-Bounding The principle that the guaranteed stability of an unperturbed system is necessary, if not sufficient, for guaranteeing the stability of perturbed systems is established in the literature (Reif et al., 1999; Sontag, 2008; Jiang and Wang, 2001; Teel et al., 2014). In Reif et al., 1999, for instance, the guarantee for noise-free dynamical systems (Theorem 3.1) was essential for the guarantee of noise-perturbed systems (Theorem 4.1). The remaining references above are the ISS references of [B.], where the same principle holds. Thus, under this principle, for the BRC-augmented EKF to be guaranteed as stable under bounded misspecification, it is necessary for it to be guaranteed as stable under model fidelity; in saying this we are essentially treating misspecification as a bounded perturbation as in [B.] above.

Our theorem shows that norm-bounding learned residuals and keeping the norm-bound (cmax⁡c_{\max}) below threshold (cˉ\bar{c}) is one of the sufficient conditions to keep the 0-misspecification system of error dynamics stable. Therefore, the theorem shows norm-bounding---which is our distinct approach (§1, contribution 1) to hybrid filtering---to be a valid approach towards stabilising hybrid filters, including under misspecification, because of the necessity principle above. Validating our design principle of norm-bounding was the main reason we developed our theorem.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.