{"id":"acc7a393-4b53-4f27-9964-dfa25cb955af","arxiv_id":"1908.04010","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A tensor-train offline/online scheme solves the filtering equation in three dimensions in near real time, with convergence analysis and 100x speedup over finite differences.","lead":"Tracking a moving target from noisy measurements is expensive in high dimensions. This paper compresses the probability cloud with tensor trains, cutting runtime from thousands of seconds to seconds in three-dimensional tests, and proves the method converges.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prop. 4.3's online complexity is contradicted by Table 5: the exponential-factor QTT construction is claimed O(N d r^2), but tEXP scales near O(N^d) and exceeds the FD exponential step at N=128, so real-time high-d support fails.","rationale":"I read the paper as making two connected claims: (i) the FKE solution has low QTT ranks, enabling fast online updates; and (ii) Theorem 5.7 gives convergence. The numerical 3D experiments are credible evidence for (i) in those examples, and the FD/QTT error analysis is a reasonable framework. My concern is with the formal complexity argument that converts low ranks into a real-time algorithm. Proposition 4.3's first complexity term is not derived from a stated structural property of h; it cites a theorem about tensor approximation, but converting a full N^d array to QTT is information-theoretically at least Omega(N^d) unless the array is never formed. The paper's own Table 5 shows tEXP growing superlinearly (roughly N^3 in the last column), inconsistent with O(N d r^2). This is a concrete internal inconsistency, not merely a missing numerical test. The reader's rank-growth concern is related and also valid: Table 5's QTT rank is small and slowly growing in the tested 3D cases, but no evidence is given for d>3. I would keep the CONDITIONAL verdict: 3D results can stand, but the high-dimensional real-time claim needs either a corrected complexity analysis (e.g., using cross approximation for low-order h) or a demonstration on d=4/5. The proposed instrumentation test would settle whether the current implementation actually achieves the claimed complexity.","tokens_in":26211,"tokens_out":12284,"duration_ms":130585,"concrete_test":"Instrument the online routine to isolate the construction of the QTT representation of exp[h^T(x,t_j)(y_{t_j}-y_{t_{j-1}})] from the Hadamard product and TT-rounding, and count the number of function evaluations and SVD operations it performs. Run d=3 with N=2^5, 2^6, 2^7, 2^8 and fit tEXP versus N and versus N^d. If tEXP scales as N^d (or the routine materializes all N^d entries), Proposition 4.3 is false and the real-time high-dimensional claim is unsupported. An analytical check is to inspect the cited Theorem 2.1 of [25] for whether it applies to arbitrary functions h: if it only covers structured matrices, the proof of Prop 4.3 is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central real-time claim rests on Proposition 4.3, which states that the online cost per observation interval is O(N d r^2 + d log_2(N) r^6). The proof attributes the first term to 'constructing the QTT-format of exp[h^T(x,t_j)(y_{t_j}-y_{t_{j-1}})] from a full multidimensional array' via Theorem 2.1 of [25]. A full array has N^d entries, so converting it to QTT cannot cost O(N d r^2) unless the function has exploitable structure and the routine never materializes the full array; no such structural assumption on h is stated. Table 5 reports the CPU time tEXP of this exponential transformation for Example 1: 0.85 s (N=2^4), 1.36 s (N=2^5), 5.20 s (N=2^6), 46.28 s (N=2^7). Doubling N increases tEXP by factors 1.6, 3.8, 8.9, with the last step close to N^3 scaling, not the linear-in-N scaling O(N d r^2) predicts (the QTT rank only grows from 6.83 to 8.89). At N=2^7 the QTT exponential step is even slower than the FD pointwise exponential step (5.33 s). Since this step is repeated at every observation time, the online procedure inherits the curse of dimensionality for d>3 unless a genuinely sublinear QTT construction is supplied. The bounded-rank assumption identified by the reader is real, but this complexity gap is independent of rank growth: even with r uniformly bounded, the stated online complexity and the real-time claim are not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a tensor-train (QTT) method for solving the forward Kolmogorov equation (FKE) that arises in nonlinear filtering, with an offline stage that precomputes low-rank QTT representations of the discretized PDE operators and propagator, and an online stage that assimilates observations through an exponential (Hadamard) update and propagates the density in QTT format. A convergence theorem (Theorem 5.7) is stated that combines the observation-time splitting error, the finite-difference discretization error, and the QTT rounding errors. Numerical experiments on two 3D filtering problems compare the method against finite differences and particle filters, reporting large speedups at moderate resolution and small mean-square errors relative to the FD reference.","tokens_in":26600,"tokens_out":9199,"duration_ms":92810,"significance":"If the assumed low-rank structure persists, the method is a promising route toward real-time nonlinear filtering: the paper gives explicit convergence analysis, identifies the three sources of error separately, and demonstrates on two 3D examples that the FKE density has low effective QTT ranks, yielding speedups of two orders of magnitude over finite differences. The numerical comparison is honest in reporting averaged MSEs over 100 paths and in showing convergence with respect to the TT-rounding precision. However, the high-dimensional and real-time claims go beyond what is demonstrated: only d=3 is tested, and the complexity analysis contains a load-bearing gap that the paper's own timing data contradicts.","major_comments":[{"comment":"The claimed O(N d r^2) cost for constructing exp[h^T(x,t_j)(y_{t_j}-y_{t_{j-1}})] in QTT format is contradicted by the paper's own timing data. The proof says this is obtained 'from a full multidimensional array' via Theorem 2.1 of [25], but a full array has N^d entries; the cited theorem does not justify a construction that avoids materializing the full array for a general nonlinear h(x)\\cdot y_t. In Table 5, doubling N from 64 to 128 increases tEXP from 5.20 s to 46.28 s (a factor of 8.9), while the averaged QTT rank grows only from 8.78 to 8.89; this scaling is close to O(N^d), not O(N d r^2). Since this step is repeated at every observation time, the online complexity stated in Proposition 4.3 and the associated real-time claim are not supported.","section":"Section 4.3, Proposition 4.3 and Table 5"},{"comment":"The convergence theorem requires |2h^T y_t| \\le c2 on the observation times, but for the SDE model (2) the observation process y_t is not bounded. Even with h bounded on the numerical domain, h^T y_t is unbounded along a Brownian path; for a fixed path the maximum is finite, but the resulting constant c2 is path-dependent and typically grows with the number of observation times, and the paper does not quantify it. The constant c2 controls the factor e^{c2} in Eq. (42) and enters the final bounds (36) and (53). The numerical experiments do not indicate that the simulated observation paths are truncated or that Asm.4 is verified, so the convergence statement as formulated does not cover the stochastic model actually simulated in Section 6.","section":"Section 5.1, Assumption (Asm.4) and Lemma 5.4"},{"comment":"The QTT contribution to the error is controlled by powers c7^{Nt}, with c7 = (1+ǫ1)(1+ǫ2)c6 e^{c2} defined after Eq. (49). The paper does not show that c7 < 1, and for a stable explicit scheme with c6 potentially larger than 1 this will generally not be the case. Consequently, the term ǫ1 c7^{Nt} in Eq. (46) and the analogous terms in (53) grow exponentially in the number of observations, so the displayed bound does not establish an error of order O(Nt^{-α} + τ + h^2 + ǫ1 + ǫ2) uniformly in Nt. As stated, the convergence theorem is therefore not proven by the given estimates unless the authors add a mechanism (such as explicit conditions guaranteeing c7 < 1 or a choice of ǫ1, ǫ2 that compensates the exponential growth).","section":"Section 5.3, Lemma 5.5 and Theorem 5.7"},{"comment":"The paper claims in the abstract and conclusions that it solves 'high-dimensional' NLF problems in a real-time manner, but all numerical experiments are for d=3 (Examples 1 and 2 in Section 6.2). The central assumption of bounded QTT ranks (Proposition 4.3) is verified only for these 3D problems, as shown in Tables 1-4; no evidence is provided for d>3, and Section 7 itself defers d>3 to future work. The high-dimensional claim in the title and abstract is therefore an overstatement of the demonstrated results.","section":"Abstract, Section 7, and Section 6"}],"minor_comments":[{"comment":"The word 'roblem' in 'NLF roblem' is a typo and should be corrected to 'problem'.","section":"Abstract"},{"comment":"The table should specify whether tFKE and tEXP are total costs over the whole simulation or per observation interval; the text says the FD method at N=2^7 would cost about 13 hours, but without a per-step breakdown the real-time claim cannot be checked against the observation interval ΔT=0.05.","section":"Section 6.3, Table 5"},{"comment":"The definition of the effective QTT-rank r_eff is nonstandard and not derived; the authors should either explain its motivation or report the maximum QTT-rank, which is the quantity that enters the complexity estimates.","section":"Section 6.1, Eq. (60)"},{"comment":"Reference [14] has a typo in the author name: 'Vlanimir A. Kazeev' should be 'Vladimir A. Kazeev'.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The numerical results for the two 3D examples are encouraging, but the central complexity claim of Proposition 4.3 is contradicted by Table 5, where the exponential-factor construction scales roughly like N^d rather than linearly in N. This is a load-bearing point for the real-time and high-dimensional claims. I recommend a major revision that either supplies a genuinely sublinear construction of the exponential factor or explicitly restricts the claims to the tested 3D setting, and that addresses the path-dependence of Assumption Asm.4 and the exponential error growth in Theorem 5.7."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about tensor methods for nonlinear filtering. The paper combines the pathwise robust DMZ formulation with a quantized tensor train offline/online scheme and gives a convergence theorem. That combination is new relative to the 1D Hermite and 2D POD predecessors, and the 3D numerical results are genuinely encouraging: the QTT solution tracks the finite-difference reference with small MSE and runs two orders of magnitude faster. The slow rank growth in Tables 1–4 is real evidence that the FKE solution has low-rank structure in these examples.\n\nThe soft spot is the online complexity. Proposition 4.3 claims the exponential observation factor can be built in O(N d r^2) using Theorem 2.1 of [25], but a full array has N^d entries; reading it already costs O(N^d). No structural assumption on h is stated that would let you construct the QTT representation without materializing the full tensor. The paper's own Table 5 shows tEXP growing roughly like N^3 as N increases from 2^4 to 2^7, not linearly, and at N=2^7 it is slower than the FD exponential step. Since this step runs at every observation time, the real-time and high-dimensional claims are not supported as written. This is independent of rank growth; even with r bounded, the stated complexity fails.\n\nThe convergence analysis is honest but loose: the global error bound carries c7^{N_t} factors and relies on Asm.4, which bounds the observation increments |2 h^T y_t|. Real SDE observations are not bounded, so the bound does not apply to the stochastic model in the numerical experiments. The authors acknowledge this assumption but do not reconcile it with the Brownian observation model. The paper also tests only d=3, so 'high-dimensional' is an extrapolation.\n\nThe citation pattern is fine, including the self-citation to [30] used for the well-posedness background; nothing circular there. No code is released, which limits reproducibility but is not unusual for this kind of paper.\n\nBottom line: a solid 3D numerical study with an interesting convergence framework, but the central complexity theorem is contradicted by the authors' own timing data. This deserves peer review, not desk rejection, and the referee should insist on either fixing the exponential-factor construction or scaling the claim back to d=3.","headline":"A promising QTT-based filter for 3D problems whose central complexity claim does not survive its own Table 5.","tokens_in":27115,"tokens_out":2385,"would_cite":false,"duration_ms":23951,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","35R60","65M12","60G35","65M99"],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantized tensor trains solve 3D nonlinear filtering in seconds","keywords":["nonlinear filtering","forward Kolmogorov equations","Duncan-Mortensen-Zakai equation","tensor train decomposition","quantized tensor train","low-rank approximation","convergence analysis","real-time algorithm"],"falsifier":"Run a filtering problem whose conditional density becomes more complex over time — for example, a bimodal drift that splits the density into well-separated peaks, or a cubic sensor in four dimensions — and record the effective QTT rank at each observation time. If the rank exceeds the assumed constant bound and keeps growing, the stated online complexity and convergence guarantee will not hold.","tokens_in":26019,"feed_emoji":"🎯","tokens_out":12602,"duration_ms":114760,"temperature":0.7,"pith_summary":"The paper sets out to establish that nonlinear filtering — estimating a hidden state from noisy observations — can be solved in real time even in dimension three, by storing the conditional density in the quantized tensor train (QTT) format. The key observation is that the solution of the forward Kolmogorov equation, which evolves the density between observations, has low-dimensional structure that can be extracted once and reused. After an offline phase builds a low-rank QTT approximation of the evolution operator, each new observation is assimilated by fast QTT operations. The paper proves convergence to the pathwise robust DMZ solution with an $L^1$ error of $O(N_t^{-\\alpha}+\\tau+h^2+\\text{QTT rounding terms})$ and demonstrates on two 3D examples that filtering updates take seconds instead of thousands of seconds.","feed_headline":"Tensor trains shrink 3D filtering from hours to seconds","feed_subtitle":"A precomputed low-rank propagator updates state estimates in ~15 s, versus ~40 min for finite differences.","key_machinery":"The central object is the quantized tensor train (QTT) decomposition: a $d$-dimensional tensor is written as a product of small cores $G_1(i_1)\\cdots G_d(i_d)$ with TT-ranks $r_k$, and each spatial dimension is reshaped into virtual binary dimensions so storage drops to $O(d\\log_2(N) r^2)$. The method's engine is the exponential transformation $\\sigma(x,t)=\\exp(h^T(x,t)S^{-1}(t)y_t)u(x,t)$, which converts the stochastic DMZ equation into a deterministic FKE, plus the offline assembly of the operator $A=\\frac12\\Delta_d-C_d-\\frac12 Q_d$ in QTT format and its power $(\\tau A+I)^{\\Delta T/\\tau}$. The QTT rank of the discrete Laplacian is bounded by 4, the convection operator's rank by $5dr$ when the drift functions have rank $r$, and TT-SVD rounding controls the ranks of everything else; these bounds are what make the online operations — Hadamard products and matrix-vector products — fast.","core_discovery":"On the paper's own terms, the central discovery is that the forward Kolmogorov equation (FKE) — a parabolic PDE for the unnormalized conditional density — admits low-rank quantized tensor train representations over the filtering horizon, so the whole filtering loop can be run in compressed form. Starting from the pathwise robust DMZ equation (the stochastic PDE for the unnormalized conditional density), the paper freezes each observation interval, solves the FKE with a finite-difference operator assembled in QTT format, and assimilates new observations through the exponential transformation. Theorem 5.7 bounds the $L^1$ error between the QTT-recovered density and the DMZ solution by three contributions: the time-frozen approximation of the observation path, the finite-difference discretization error $O(\\tau+h^2)$, and the QTT rounding errors. Numerically, the method matches finite-difference estimates on two 3D examples (mean-square errors $0.007$ and $0.023$) while cutting the per-run CPU time from $2052$ s and $4079$ s to about $15$ s and $17$ s.","pith_inferences":["A natural stress test, not performed in the paper, is a time-dependent drift or observation function: the paper itself notes the low-rank structure may shift over time, so an adaptive or dynamic low-rank update would be needed.","The offline phase must build low-rank QTT tensors from full discretized operators; the paper reports this cost is modest for the 3D tests, but for much larger $d$ or $N$ the offline construction could become the bottleneck unless cross-approximation is used.","Assumption 4 requires the observation term $|2h^T y_t|$ to be bounded by a constant, but real observations driven by Brownian motion are not bounded; a fully rigorous guarantee would need a truncation of observations or a probabilistic bound on the exceptional paths.","If the small-rank conjecture survives in $d>3$, the $O(N d r^2)$ cost of assimilating the exponential observation factor becomes the main bottleneck, so the scaling in very high dimension deserves direct measurement."],"forward_implications":["If the ranks stay bounded, the online per-interval cost is $O(d\\log_2(N) r^6)$, so refining the spatial grid makes the solver only logarithmically more expensive.","The offline/online split means the same low-rank propagator can be reused for every observation path of a given system, making repeated or long-horizon filtering much cheaper after a one-time setup.","Theorem 5.7 gives explicit error control: reduce $N_t$ for the path-frozen error, $\\tau$ and $h$ for the finite-difference error, and the TT-rounding precisions for the QTT error.","In the two 3D examples, the method matches the finite-difference density estimates while reducing CPU time from 2052 s and 4079 s to about 15 s and 17 s."],"supporting_citations":[{"why":"Supplies the pathwise robust DMZ equation and the exponential transformation (5) that turns the filtering SPDE into the deterministic FKE (8).","marker":"[30]"},{"why":"Defines the TT-format, its storage and operation complexities, and the TT-rounding procedure used throughout.","marker":"[26]"},{"why":"Introduces the QTT format and the O(N d r^2) construction of the exponential factor in Proposition 4.3.","marker":"[25]"},{"why":"Supplies the QTT operator-power and rounding techniques used to precompute (τA+I)^{ΔT/τ} offline.","marker":"[7]"},{"why":"Gives the rank-4 QTT bound for the discrete Laplacian used in assembling the operator A.","marker":"[14]"},{"why":"Provides the TT-SVD error bound (Proposition 4.2) that controls rounding in the QTT approximation.","marker":"[27]"}],"fun_headline_variants":["Tensor trains solve high-dimensional filtering in real time","Real-time nonlinear filtering via tensor train low-rank structure","Tensor train method accelerates 3D filtering from hours to seconds","Low-rank tensor trick enables fast, accurate nonlinear filtering","Compressed tensor propagator speeds up filtering by over 100x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method works only if the compressed representation of the density stays small over the entire filtering window; the paper assumes exactly this (Proposition 4.3) and checks it numerically on only two three-dimensional examples. If the compression size grows with time or dimension, the promised speed-up and the error bound both collapse.","fun_headline_variants_meta":{"raw":{"variants":["Tensor trains solve high-dimensional filtering in real time","Real-time nonlinear filtering via tensor train low-rank structure","Tensor train method accelerates 3D filtering from hours to seconds","Low-rank tensor trick enables fast, accurate nonlinear filtering","Compressed tensor propagator speeds up filtering by over 100x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000303,"raw_usage":{"total_tokens":1742,"prompt_tokens":946,"completion_tokens":796,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":716}},"tokens_in":562,"tokens_out":796,"duration_ms":8336,"temperature":1.0,"reasoning_tokens":716,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:54:24.952171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a filtering problem whose conditional density becomes more complex over time — for example, a bimodal drift that splits the density into well-separated peaks, or a cubic sensor in four dimensions — and record the effective QTT rank at each observation time. If the rank exceeds the assumed constant bound and keeps growing, the stated online complexity and convergence guarantee will not hold.","supporting_citations":[{"cited_title":"Real time solution of the n onlinear filtering problem without memory II","cited_arxiv_id":null,"evidence_quote":"Supplies the pathwise robust DMZ equation and the exponential transformation (5) that turns the filtering SPDE into the deterministic FKE (8)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the TT-format, its storage and operation complexities, and the TT-rounding procedure used throughout."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the QTT format and the O(N d r^2) construction of the exponential factor in Proposition 4.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the QTT operator-power and rounding techniques used to precompute (τA+I)^{ΔT/τ} offline."},{"cited_title":"Kazeev, and Boris N","cited_arxiv_id":null,"evidence_quote":"Gives the rank-4 QTT bound for the discrete Laplacian used in assembling the operator A."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the TT-SVD error bound (Proposition 4.2) that controls rounding in the QTT approximation."}],"review_version":1}