{"id":"c9f02d05-2a1c-4926-a589-c5f7e5164d40","arxiv_id":"2507.21531","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A hierarchical Brownian bridge SDE model with marked point process inducing points is proposed for latent manifold learning, with linear-time inference via sequential Monte Carlo.","lead":"This paper proposes a new type of state-space model for learning low-dimensional dynamics from high-dimensional time series, using Brownian bridge processes pinned to a sparse set of randomly timed inducing points. It may interest neuroscientists and machine learning researchers who want efficient latent dynamics models with an interpretable structure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The discrete-time transition in Eq. (18) does not discretize the proposed SDE: the noise term scales as Δt instead of √Δt, so inference is run on a different, effectively deterministic process.","rationale":"The paper's central claim is that the hierarchical SDE model admits a tractable, correctly derived inference procedure with linear cost, and that the experiments demonstrate this. That claim fails at the discretization step: Eq. (18) is not a valid discretization of Eq. (3), so the SMC algorithm is propagating a different stochastic process. This is more concrete than the independence assumption in Eq. (2), which the authors explicitly flag as a simplification in Section 4 and which is a modeling choice rather than a mathematical inconsistency. The discretization error, by contrast, is internal: the discrete process does not converge to the continuous model as Δt→0. Still, the error is local and fixable, and the universal approximation argument is unaffected; the underlying model idea is sufficiently plausible that the appropriate disposition remains the reader's CONDITIONAL verdict (major revision), not outright rejection. The reader's rationale already mentions an 'apparent discretization error', so this concern corroborates rather than overturns their assessment; hence UNCHANGED.","tokens_in":14754,"tokens_out":10011,"duration_ms":123186,"concrete_test":"Re-derive the discrete transition of Eqs. (3)–(4) by Euler–Maruyama and compare with Eq. (18): verify that the conditional variance of x_{k+1} is σ_x^2 σ_t^2 Δt^2 in Eq. (18) but should be σ_x^2 σ_t^2 Δt. Then rerun the chirp experiment of Section 3.1 with the corrected transition (replace Δt w_k by √Δt w_k) using the same seeds and hyperparameters; if the posterior inducing points, X-path variance, or reconstruction MSE change materially, the current empirical results depend on the erroneous discretization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is an internal inconsistency between the continuous model and the discretized model used for inference. In Section A.3, Eq. (18) writes x_{k+1} = x_k + (m_{i+1} - x_k)/(t_{i+1} - kΔt) Δt + sqrt(((t_{i+1} - kΔt)(kΔt - t_i))/(t_{i+1} - t_i)) Δt w_k, with w_k ~ N(0, σ_x^2). But Eq. (3) is an Itô SDE with diffusion coefficient σ_t = sqrt(((t_{i+1} - t)(t - t_i))/(t_{i+1} - t_i)); its Euler–Maruyama discretization requires σ_t sqrt(Δt) ε_k (with ε_k ~ N(0, σ_x^2)), not σ_t Δt w_k. The manuscript's factor of Δt instead of √Δt makes the X-increment variance O(Δt^2) rather than O(Δt), so as Δt→0, X converges to the deterministic ODE dx/dt = (m_{i+1} - x)/(t_{i+1} - t). Because Eq. (18) is the transition density used in the SMC weights in Algorithm 1 and Algorithm 2, all reported posteriors over inducing points and latent trajectories are computed for a different, essentially noise-free process than the hierarchical SDE that is the paper's central object. This is not a modeling simplification acknowledged in Section 4; it is a derivation error that invalidates the reported inference pipeline and the empirical validation. The universal approximation theorem is not directly affected, but the paper's central claim of a working inference procedure for the hierarchical SDE model is.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a hierarchical stochastic differential equation model for latent manifold learning in neural time series. Inducing points, defined as time-mark pairs from a marked point process, parameterize Brownian-bridge SDEs whose expectation defines the drift of a second SDE layer; observations are noisy linear projections of the latent process. The authors prove a universal approximation property for the expected latent process, derive an SMC-based EM inference procedure claimed to scale linearly with sequence length, and present experiments on a chirp signal, a projected Lorenz system, and the NLB MC_Maze neural dataset. The central inference pipeline, however, is built on a discrete-time transition that does not correctly discretize the proposed SDE.","tokens_in":15069,"tokens_out":8312,"duration_ms":94452,"significance":"The modeling idea is attractive: a sparse, adaptive set of inducing points connected by Brownian bridges gives an interpretable, nonparametric latent dynamical model that avoids kernel choices and cubic GP scaling. The claimed O(P·N) inference cost and the universal approximation property are, if correct, useful contributions for neural time-series analysis. The paper also deserves credit for explicitly acknowledging the limitations of the independence assumption on waiting times and marks. However, the current manuscript is not reliable in its reported form because the discretization used in inference does not match the continuous SDE, so the empirical validations do not yet support the central claim that the proposed hierarchical SDE can be trained and inferred successfully.","major_comments":[{"comment":"The discrete-time transition in Eq. (18) is not a valid Euler-Maruyama discretization of the SDE in Eqs. (3)-(4). For an Ito SDE dX_t = mu_t dt + sigma_t dW_t, the increment over a step of length dt is mu_t dt + sigma_t dW_t with dW_t ~ N(0, dt); hence the diffusion term must carry a factor sqrt(dt), not dt. As written, the variance of the X-increment is O(dt^2) rather than O(dt), so in the limit dt -> 0 the process converges to the deterministic ODE dX_t/dt = (m_{i+1} - X_t)/(t_{i+1} - t). Because Eq. (18) supplies the transition density p(X_k | tau, m) used in the importance weights of Algorithms 1 and 2, all reported posteriors over inducing points and latent trajectories are computed for a different, effectively noise-free process than the hierarchical SDE defined in Section 2.1. This is a load-bearing error in the inference pipeline and must be corrected (the sqrt(dt) factor) and the experiments rerun.","section":"A.3, Eq. (18)"},{"comment":"The renewal assumption p(tau_i) p(m_i) in Eq. (2) is load-bearing: it is what allows the non-Markovian X process to be treated as Markovian in Algorithm 1 by conditioning on the next event time and mark. The authors acknowledge in Section 4 that this 'may reduce its ability to adjust inducing points to capture changes in data happening at different temporal scales.' This limitation is not merely cosmetic: because tau_i and m_i are generated independently of the data, the generative model has no mechanism for placing more inducing points where the dynamics are fast; any such adaptation in Figures 1B and 2 must come entirely from posterior reweighting. The paper should temper the claims of adaptive inducing-point placement or provide a quantitative comparison of posterior waiting times against data-driven event times.","section":"Section 2.1, Eq. (2); Section 4"},{"comment":"The diffusion parameters sigma_x and sigma_y and observation noise variance R are set by hand (e.g., sigma_x = 10^{-1}, sigma_y = 10^{-4} in Section 3.1; see also Section 4, where the authors state 'we adjust their values by checking different values for both'), yet the model's behavior is strongly controlled by these choices: sigma_x sets the flexibility of the interpolation paths and sigma_y controls how much observation information propagates back to X. The paper reports no sensitivity analysis or posterior uncertainty for these parameters. Since the central empirical claims are qualitative, the authors should at least show that the reported latent trajectories and inducing-point placements are stable across reasonable ranges of these parameters.","section":"Sections 3.1, 3.2, and 4"}],"minor_comments":[{"comment":"The reference to Shannon's sampling theorem is misleading: the proof uses piecewise linear interpolation and density in C([0,T]), not bandlimited sampling; the connection to Shannon should be removed or clarified.","section":"Section 2.2.1"},{"comment":"Chebyshev's inequality bounds the probability that a single time point deviates from its mean; the text writes 'Y_t concentrates around E[Y_t]' but the uniform-in-t statement used later requires a union bound or a modulus-of-continuity argument. This does not affect the main approximation claim, but the proof is slightly imprecise.","section":"Section 2.2.1, proof"},{"comment":"Sampling new tau and m from the prior with no data-dependent proposal is a weak proposal and may cause weight degeneracy; the paper does not report effective sample size (ESS), which would help assess the reliability of the SMC results.","section":"Algorithm 1, line 10"},{"comment":"The comparison with the GP reports MSE values 0.22 and 0.30 without error bars or multiple restarts; the sentence 'We expect our model to have a slightly larger MSE' is not a substitute for a statistical comparison.","section":"Section 3.1"},{"comment":"The claim that 'more inducing points are placed around instances where there is a fast transition' is not quantified; a scatter of inferred tau_i against local |dZ/dt| or a similar diagnostic would strengthen the claim.","section":"Section 3.2"},{"comment":"There are several typographical errors (e.g., 'ne ural' in the abstract, 'Y ousefi' in the author list) and inconsistent notation (e.g., p used for both densities and probabilities in Eq. (2)); a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The discretization error in Eq. (18) is the main obstacle; if the authors fix the sqrt(dt) factor and rerun the experiments, the paper could become publishable after a rigorous revision. I would also ask the editor to require either code/data release or detailed reproducibility information, since the empirical results are central to the paper's claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere are two things you should know about arXiv:2507.21531.\n\nFirst, the architecture is genuinely new: a hierarchical SDE model where a marked point process generates inducing points, the latent X process is a Brownian bridge pinned to those points, and a second SDE (Y) is driven by X. The authors derive a sequential Monte Carlo inference scheme and claim linear-time scaling. The universal approximation theorem is correct, though it is a standard piecewise-linear density result. The model is interpretable and the paper is clearly written.\n\nSecond, the discrete-time inference is built on a wrong noise scaling. In Eq. (18), the diffusion term is written as σ_t Δt w_k with w_k ~ N(0, σ_x^2). The SDE (3) is an Itô equation; the Euler-Maruyama discretization requires σ_t √Δt ε_k. With the paper's scaling, the noise variance is O(Δt^2), so in the continuous limit X converges to a deterministic ODE. This means the SMC weights in Algorithms 1 and 2 are computed for a different process than the one the paper defines. It is not an acknowledged approximation; it is an internal inconsistency.\n\nThis does not damage the universal approximation theorem, but it does invalidate the empirical results as evidence for the proposed model. The only quantitative comparison (chirp) shows the model underperforming a GP (MSE 0.30 vs 0.22). The Lorenz and neural results are qualitative, with no error bars or baselines. The independence assumption in Eq. (2) is a real limitation that the authors acknowledge, but it is load-bearing for tractability.\n\nFor a reader: the model family is worth pondering, and the paper could be a useful starting point for work on adaptive inducing points. But do not cite the experimental claims until the discretization is fixed and code is available. I would send it to peer review because the idea is novel and the flaws look addressable, but I would require a major revision with a corrected derivation, reproducible code, and quantitative comparisons on multiple benchmarks.\n\nBest,","headline":"New hierarchical SDE architecture worth knowing, but the discrete-time inference in Eq. (18) uses the wrong noise scaling, so the reported experiments are for a different, nearly deterministic process.","tokens_in":15645,"tokens_out":4539,"would_cite":false,"duration_ms":50179,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sparse inducing points, joined by Brownian-bridge SDEs, let a hierarchical model uniformly approximate any continuous trajectory while keeping inference linear in time.","keywords":["hierarchical stochastic differential equations","latent manifold learning","neural time series","Brownian bridge","marked point process","sequential Monte Carlo","universal approximation","state-space models"],"falsifier":"Simulate a continuous $f$ with a sharp localized feature, generate observations from the true model with known inducing points, run the SMC inference at increasing particle counts, and measure $\\sup_t |\\mathbb{E}[Y_t] - \\int_0^t f(s)\\,ds|$; if the error does not shrink toward zero as $N$ grows—or if the inferred waiting times stay long near the sharp feature—the practical claim of adaptive, linear-cost universal approximation is contradicted.","tokens_in":14518,"feed_emoji":"🧠","tokens_out":5865,"duration_ms":57570,"temperature":0.7,"pith_summary":"This paper proposes a hierarchical stochastic differential equation model for discovering low-dimensional latent manifolds in high-dimensional time series, especially neural recordings. The central idea is that a sparse set of event-value pairs—called inducing points—can be turned into smooth latent trajectories by Brownian bridges, and those trajectories form the drift of a second SDE whose noisy linear projection produces the observations. The authors prove that, in the limit of many inducing points, the expected latent process can uniformly approximate the integral of any continuous function, and they derive a sequential Monte Carlo training and inference scheme whose cost grows linearly with the length of the data. On simulated chirps, a Lorenz attractor, and a macaque reaching dataset, the model recovers the latent manifold and places more inducing points where dynamics change quickly. The value of the claim is that it offers a nonparametric, interpretable alternative to Gaussian-process and deep state-space models without requiring a hand-chosen kernel.","feed_headline":"Sparse bridge points reconstruct any continuous signal","feed_subtitle":"Hierarchical Brownian-bridge SDEs infer latent manifolds from neural time series at linear cost.","key_machinery":"The load-bearing object is the hierarchical pair of SDEs. Inducing points $(\\tau_i, m_i)$, drawn from a marked point process with Gamma waiting times and Gaussian marks, define a Brownian bridge process $X_t$—an SDE pinned to reach each mark at its time—whose path then serves as the drift of a second SDE $Y_t$; observations are a noisy linear projection of $Y_t$. The renewal-process simplification (waiting times and marks independent of history) breaks the non-Markovian dependence and makes sequential Monte Carlo inference possible. The universal approximation result rests on the density of piecewise-linear interpolants of uniformly spaced samples, with the bridge noise concentrated by Chebyshev's inequality.","core_discovery":"The paper's central claim is that a cascade of two SDEs—an inner Brownian bridge constrained to pass through sampled marks at sampled times, and an outer process whose drift is the inner path—is expressive enough to model any continuous low-dimensional signal while remaining tractable. The formal statement is a universal approximation theorem: for any continuous $f$ and any $\\varepsilon > 0$, there exists a choice of inducing points such that $\\mathbb{E}[Y_t]$ uniformly approximates $g(t) = \\int_0^t f(s)\\,ds$ within $\\varepsilon$. The authors also claim that the discretized model can be trained and inferred with a custom particle filter whose cost per iteration is $O(P\\cdot N)$, avoiding the cubic cost of Gaussian process inversion, and that the inducing points adapt to the data's temporal structure. They demonstrate the framework on synthetic chirp and Lorenz data and on neural population recordings, where the inferred latent trajectory changes around target onset, consistent with preparatory motor activity.","pith_inferences":["The universal approximation theorem guarantees existence of good inducing points, not that the SMC inference will find them; a direct test of posterior concentration as $N$ grows would separate expressiveness from recoverability.","The independence assumption in Eq. (2) forces waiting times to be drawn afresh regardless of how fast the signal is changing, so the model may systematically under-sample sharp transitions; a history-dependent or repulsive prior, which the authors mention, would be a natural extension to test.","The linear-cost claim depends on the number of particles $P$ staying fixed; comparing effective sample size and accuracy at fixed compute against Kalman-filter hybrids could show whether the SMC overhead is actually necessary.","Applied to neural decoding, the model's inferred $Y_t$ could be used as a real-time readout of preparatory dynamics; this is a testable extension the paper leaves open."],"forward_implications":["Inference cost scales as $O(P\\cdot N)$ per SMC iteration, so long, high-resolution neural time series become feasible where Gaussian-process models hit $O(N^3)$.","As the number of inducing points grows, the expected latent process can uniformly approximate the integral of any continuous function, so the model's expressiveness is not limited by a fixed parametric form.","Inducing points are inferred from data and concentrate where dynamics change rapidly—for example, around fast Lorenz transitions—giving an interpretable summary of where the manifold is complex.","The same latent process can be paired with point-process observation models, as in the reaching-task analysis, making it applicable to spike-train recordings rather than only continuous observations.","Because the model needs no pre-specified kernel, it offers a nonparametric alternative to GP-based latent dynamic models for exploratory neural data analysis."],"supporting_citations":[{"why":"Provides the Brownian bridge process used to pin the inner SDE to the inducing points.","marker":"[Pitman and Yor, 1999]"},{"why":"Supplies the marked point process and renewal-process formalism behind the inducing-point distribution.","marker":"[Daley and Vere-Jones, 2006]"},{"why":"Supplies the sampling-theorem intuition that a continuous signal can be reconstructed from sufficiently dense samples, which the universal approximation proof builds on.","marker":"[Shannon, 1949]"},{"why":"Provides the sequential Monte Carlo methodology the paper adapts for particle-based inference.","marker":"[Doucet et al., 2001]"},{"why":"Establishes the $O(N^3)$ Gaussian-process cost baseline that the paper's $O(P\\cdot N)$ claim is measured against.","marker":"[Seeger, 2004]"},{"why":"Provides the point-process observation model used to map latent states to neural spike trains.","marker":"[Eden et al., 2004]"},{"why":"Supplies the Neural Latents Benchmark MC_Maze dataset used for the reaching-task validation.","marker":"[Pei et al., 2021]"},{"why":"Provides the experimental finding about preparatory neural population dynamics that the inferred latent trajectory is compared with.","marker":"[Churchland et al., 2012]"}],"fun_headline_variants":["Hierarchical SDEs learn any continuous signal from sparse points","Sparse Brownian bridges scale latent manifold inference linearly","From neural spikes to continuous manifolds via two-stage SDEs","Linear-complexity SDEs approximate arbitrary continuous trajectories"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that waiting times between inducing points and their mark values can be treated as independent of each other and of past events (Eq. 2); if that independence fails to hold for a given dynamic, the model cannot adapt its inducing points to the data's temporal scales, as the authors acknowledge.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical SDEs learn any continuous signal from sparse points","Sparse Brownian bridges scale latent manifold inference linearly","From neural spikes to continuous manifolds via two-stage SDEs","Linear-complexity SDEs approximate arbitrary continuous trajectories"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000744,"raw_usage":{"total_tokens":3323,"prompt_tokens":953,"completion_tokens":2370,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":2301}},"tokens_in":569,"tokens_out":2370,"duration_ms":16866,"temperature":1.0,"reasoning_tokens":2301,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:38:43.834307+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a continuous $f$ with a sharp localized feature, generate observations from the true model with known inducing points, run the SMC inference at increasing particle counts, and measure $\\sup_t |\\mathbb{E}[Y_t] - \\int_0^t f(s)\\,ds|$; if the error does not shrink toward zero as $N$ grows—or if the inferred waiting times stay long near the sharp feature—the practical claim of adaptive, linear-cost universal approximation is contradicted.","supporting_citations":[],"review_version":1}