{"id":"5607832f-7764-4548-bd14-62595822c02e","arxiv_id":"1908.01533","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A tensor-train policy iteration solver computes feedback controls for a 121-dimensional Allen-Cahn model and reduced Fokker-Planck models with polynomial scaling in dimension.","lead":"Researchers combined tensor-train compression with Newton-style policy iteration to solve Hamilton-Jacobi-Bellman equations from optimal feedback control in more than 100 state-space dimensions. The method produces stabilizing controllers for nonlinear Allen-Cahn and Fokker-Planck systems where conventional grids are impossible.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Polynomial scaling and 100+ dimensional accuracy rest on empirically observed, not proven, TT-rank growth for nonlinear value functions.","rationale":"The reader's CONDITIONAL verdict matches my read. The weakest assumption is indeed low TT representability of the value function for nonlinear problems: the polynomial scaling and >100-dimensional feasibility claims depend on TT ranks growing mildly with d, yet the only proven rank bound is for the quadratic LQ case, and the nonlinear evidence is limited to two dissipative parabolic examples. The authors themselves flag the identification of low-rank problem classes as open, and Remark 3.3 restricts Theorem 3.1 to linear systems. The accuracy of the 121-dimensional Allen-Cahn result is also validated only against LQR and uncontrolled trajectories, not an independent HJB reference, so the 'accurate synthesis' claim is not independently confirmed. Because these are addressable empirical and validation gaps rather than demonstrated contradictions, I recommend no change to the reader's CONDITIONAL verdict.","tokens_in":22065,"tokens_out":11003,"duration_ms":122826,"concrete_test":"Use the released code to run the 1D Allen-Cahn example at d=10,20,40,80 with the nonlinearity coefficient λ multiplying x−x^3 set to λ=1,2,5,10, recording maximum TT rank and CPU time; if the rank-vs-d slope grows with λ beyond linear, the polynomial-scaling claim fails. In the same runs, for d=9, compare the TT-HJB feedback law against a dense full-tensor Galerkin solve (5^9 unknowns) to obtain an independent accuracy benchmark instead of LQR.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline scaling and >100-dimensional feasibility claims require that, for the nonlinear problems to which the method is applied, the value function and the nonlinear system functions have TT ranks that grow only mildly, at most polynomially, with the state-space dimension d. Theorem 3.1 establishes a polylogarithmic rank bound only for the quadratic (LQ) value function; for the Allen-Cahn and Fokker-Planck examples the rank behavior is observed numerically (Figs. 4.2, 4.7, 4.11), and the Conclusion explicitly identifies the class of low-rank representable problems as open: 'An important matter of investigation is the identification of a class of optimal control problems where the value function can be accurately represented with a low-rank tensor train structure.' Remark 3.3 likewise limits Theorem 3.1 to linear systems and cautions that the TT ranks can grow with the number of spatial degrees of freedom in 2D. The two nonlinear examples are strongly dissipative parabolic PDEs; the paper itself notes that convection-dominated PDEs may require a much more complex state-space representation. Thus the 'wide class' and 'polynomial in dimension' claims are not established for nonlinear dynamics beyond the tested dissipative examples. In addition, the 'accurate synthesis' claim for the 121-dimensional Allen-Cahn result is not benchmarked against an independent HJB solution: Figs. 4.3, 4.7, and 4.8 compare costs and controls with LQR and uncontrolled trajectories, which can show that the controller stabilizes and beats LQR but cannot show it approximates the true optimal value function. If TT ranks grow super-polynomially for other nonlinear dynamics, the polynomial-scaling claim fails; if the ranks stay low but the computed value is not verified against the HJB equation, the accuracy claim remains unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops a deterministic tensor-train (TT) method for stationary Hamilton-Jacobi-Bellman (HJB) equations arising in infinite-horizon optimal feedback control. The method combines a Legendre Galerkin discretization, TT representations of the value function and system functions via the TT-Cross algorithm, continuous policy iteration, and a shifted AMEn linear solver. Theorem 3.1 provides a TT-rank bound for quadratic (linear-quadratic) value functions by connecting the Riccati equation to low-rank approximations of a Lyapunov inverse. Numerical experiments cover feedback stabilization of the Allen-Cahn equation in one and two spatial dimensions (up to d=121) and of a reduced Fokker-Planck model (d=10), including one constrained-control example. The paper's central claims are that the method scales at most polynomially in the state dimension (largest observed order 4) and that accurate optimal feedback maps can be synthesized for nonlinear dynamics with over 100 state variables.","tokens_in":22284,"tokens_out":6367,"duration_ms":68593,"significance":"If the numerical claims hold, the paper is a useful step toward deterministic high-dimensional HJB solution; the combination of policy iteration with TT algebra is natural and the implementation is nontrivial. The manuscript's strengths are the explicit LQ rank theorem in Theorem 3.1, the reproducibility statement with a public code repository, and the systematic numerical convergence studies in the polynomial degree, the TT threshold, the spatial dimension, and the number of finite-difference points. Its main deficit is that the nonlinear low-rank and accuracy claims are not established beyond the tested dissipative examples and are in part explicitly acknowledged as open in the Conclusion. The significance is therefore moderate: the method is promising, but the advertised scope should be qualified to match the evidence.","major_comments":[{"comment":"The polynomial-in-dimension claim in Sec. 1.3 is not established by Theorem 3.1 for the PDE examples. The bound r_k ≲ (M+r_b)(log(1/ε)+C)^(7/2) assumes that the off-diagonal blocks of A D^{-1} have rank at most M uniformly in k and d. For the Chebyshev pseudospectral Allen-Cahn discretization (4.3), A is not banded, and Fig. 4.2 reports the maximal TT rank growing linearly in d; hence the hypothesis rank AD(k+1:d,1:k) ≤ M is not satisfied in the regime advertised. The linear-in-d rank observation is an empirical finding for the tested parabolic problems, not a consequence of Theorem 3.1, and the Conclusion (\"identification of a class ... open\") says exactly this. I request that the claims of a \"wide class\" and \"polynomial scaling\" be reworded to the class of problems for which the off-diagonal rank assumption or an analogous nonlinear rank bound holds, or that the theorem be extended to the semidiscrete elliptic operators used in Section 4.","section":"Sec. 3.1, Theorem 3.1 and Sec. 1.3"},{"comment":"The statement \"accurate synthesis of optimal feedback maps ... over 100 dimensions\" is not benchmarked against an independent HJB solution for the 121-dimensional case. The comparisons in Figs. 4.3 and 4.8 are against LQR and uncontrolled trajectories; they demonstrate that the computed law stabilizes the sampled initial state and gives lower cost than LQR, but they do not quantify the error in the value function or the feedback law. For the one-dimensional problem a dimension-refinement cost error is shown (J_d − J_64 ∼ d^{−2.3}), but no analogous indicator is provided for the two-dimensional d=121 run. I suggest adding a residual-based error indicator, a refinement study in (n,d), or a test with a manufactured or known solution to support the accuracy claim at d=121.","section":"Sec. 4.1, Figs. 4.3 and 4.8"},{"comment":"Algorithm 3.2 relies on heuristic choices of the shift μ and its reduction factor q, and the manuscript provides no convergence analysis for the shifted AMEn iteration on the nonsymmetric, degenerate matrix (3.3). The argument that the transition matrix μ(A+μI)^{-1} has spectral radius less than 1 assumes Re λ(A) ≥ 0, while the text then acknowledges that the eigenvalues are only in the right half-plane \"for a suitable choice\" of domain and polynomial order and that larger shifts or domain sizes can make the stiffness matrix indefinite (Sec. 4.2). Since the policy iteration and the linear solver are not separated in the reported CPU times, the reliability of the method in the advertised high-dimensional regime depends on this heuristic. Please state precise conditions under which Algorithm 3.2 converges, or reformulate the algorithm's guarantees as numerical observation.","section":"Sec. 3.3, Algorithm 3.2"}],"minor_comments":[{"comment":"The notation P^{-1}(µ) is ambiguous; from the surrounding text it is the inverse function of P, not the reciprocal, and this should be stated explicitly.","section":"Eq. (2.7)"},{"comment":"The phrase \"optionally µ := µq\" is unclear about when the shift is reduced relative to the policy iteration loop; a precise schedule (e.g., every outer iteration after the first) would help reproducibility.","section":"Algorithm 3.2, line 3"},{"comment":"The statement that TT ranks \"grow very mildly\" for one-dimensional PDEs is hard to reconcile with Fig. 4.2, which shows a roughly linear growth in d; \"mildly\" should be quantified (e.g., linear with a small slope) to avoid confusion.","section":"Remark 3.3 and Fig. 4.2"},{"comment":"The control set U is introduced as a compact subset of R and then the unconstrained case U ≡ R is used; the transition between the two settings should be explicit.","section":"Sec. 2"},{"comment":"The text reports \"time ∼ d^4\", whereas the complexity estimate O(d n^2 r^4) with r ∼ d would give O(d^5); the discrepancy is attributed to non-uniform ranks and is a useful observation, but it should be stated as an empirical fit rather than a complexity bound.","section":"Fig. 4.2 and Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely acceptable after the central claims are narrowed and the 121-dimensional accuracy claim is supported by an error indicator. The novelty relative to existing TT-based dynamic programming is incremental but real, and the LQ rank theorem is a genuine contribution. I would invite a revision rather than reject, because the component algorithms are sound and the numerical evidence is internally consistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is the first deterministic, PDE-based tensor-train solver I know of for stationary fully nonlinear HJB equations that genuinely goes past 100 dimensions. The Allen-Cahn 121-dimensional example is the real headline, and the paper is honest about what it can and cannot prove.\n\nThe core engineering is sound. They combine continuous policy iteration with a Galerkin Legendre discretization, assemble the linearized HJB operator in the tensor-train format, and solve the nonsymmetric degenerate systems with a shifted AMEn iteration. That last piece matters: the shift makes the linear solves work despite the gradient operator's nullspace, and the geometric reduction of the shift as the policy iteration converges is a practical touch that clearly took experimentation to get right. The paper ships Matlab code, gives convergence studies in polynomial degree, TT threshold, and dimension, and reports CPU times on one core. That is reproducible work and should be credited.\n\nThe main theoretical contribution is Theorem 3.1: a parameter-free polylogarithmic TT-rank bound for linear-quadratic value functions, derived through Riccati/Lyapunov analysis and low-rank approximants to the matrix exponential. I checked the structure of the proof, not every constant, and it is plausible and genuinely useful. It is also clearly labeled as LQ-only, and Remark 3.3 correctly warns that TT ranks can grow with spatial degrees of freedom for 2D discretizations.\n\nThe soft spots are real but not fatal. The load-bearing assumption for the nonlinear results is that value functions of these controlled parabolic PDEs keep low TT rank as the dimension grows. That is observed numerically, with ranks growing roughly linearly in d, but not proven for nonlinear dynamics. The paper itself flags this as open in the conclusion, so the authors are not hiding it. Still, the polynomial-scaling claim is conditional on that empirical regularity. Second, the 121-dimensional Allen-Cahn result is validated by comparing running costs and controls against LQR and uncontrolled trajectories, not against an independent HJB solution. That shows the controller stabilizes and beats LQR, which is meaningful, but it does not quantify how close the computed value is to the true optimal value. Third, the Fokker-Planck pipeline inherits an unquantified bilinear balanced-truncation error; they estimate cost differences with respect to reduced dimension, but there is no a priori bound. These are addressable weaknesses, and none of the paper's equations show an internal contradiction.\n\nWho should read this: anyone working in deterministic optimal feedback synthesis, tensor methods for PDEs, or high-dimensional dynamic programming. It deserves a serious referee. My recommendation is to send it to peer review and push for a modest revision: add a discussion of when TT ranks can be expected to stay low for nonlinear systems, and benchmark at least one high-dimensional case against a policy computed by another method or against the HJB residual itself.","headline":"A real deterministic TT solver for stationary nonlinear HJB equations that reaches 100+ dimensions, with an honest LQ rank theorem and shipped code; the high-dimensional accuracy claims rest on empirically observed low-rank structure and LQR/uncontrolled comparisons, not yet on independent HJB verification.","tokens_in":22928,"tokens_out":1407,"would_cite":true,"duration_ms":17157,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","15A23","65F10","65N22","49J20","49LXX","49MXX"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that tensor-train methods can solve high-dimensional HJB equations for nonlinear feedback control with cost polynomial in the state dimension, demonstrated up to 121 dimensions.","keywords":["tensor train","Hamilton-Jacobi-Bellman equation","optimal feedback control","curse of dimensionality","policy iteration","Allen-Cahn equation","Fokker-Planck equation","high-dimensional PDE"],"falsifier":"Fix a target accuracy and compute the maximal tensor-train rank of the value function for a family of nonlinear control problems (for instance, transport-dominated or nonlocal dynamics) as the state dimension $d$ increases; if the rank required to keep the error fixed grows faster than polynomial in $d$, the polynomial-scaling claim collapses even though each algorithmic step remains correct.","tokens_in":21799,"feed_emoji":"🧮","tokens_out":6471,"duration_ms":62186,"temperature":0.7,"pith_summary":"The paper tries to show that Bellman's 'curse of dimensionality' can be partially circumvented for a broad class of nonlinear optimal feedback control problems by writing the value function of a Hamilton-Jacobi-Bellman (HJB) equation in tensor-train form. On top of this compressed representation it runs a Newton-like continuous policy iteration, solving each linearized equation with a shifted alternating least-squares method. The reported outcome is that the method's cost scales at most polynomially (observed order about 4) with the state dimension, allowing accurate feedback synthesis for nonlinear PDE-controlled dynamics with more than 100 variables. If this holds, deterministic PDE methods become a practical route to optimal feedback control in regimes where grid-based dynamic programming and linearized approximations such as LQR are either intractable or too crude.","feed_headline":"Tensor method tames 100-dimensional nonlinear control equations","feed_subtitle":"Tensor-train value functions make feedback synthesis polynomial; tested on Allen-Cahn and Fokker-Planck control.","key_machinery":"The central object is the tensor-train (TT) decomposition, which stores a d-dimensional array using about $dnr^2$ parameters by chaining small three-dimensional blocks with $r$ ranks; here the array holds the coefficients of a Legendre spectral expansion of the value function. Around this object the method wraps three mechanisms: continuous policy iteration (a Newton-type scheme that linearizes the HJB equation at each step), TT-cross interpolation to assemble the nonlinear drift and control terms directly in low-rank form, and a shifted AMEn (alternating minimal energy) iteration that solves the resulting nonsymmetric, degenerate linear systems while preserving the TT structure. The linear-quadratic rank bound comes from writing the Riccati solution as the inverse of a Kronecker-sum operator and approximating that inverse by a sum of matrix exponentials of bounded quasi-separable rank.","core_discovery":"The central claim is that the value function of the stationary HJB equation associated with deterministic infinite-horizon optimal control can be represented and computed in the tensor-train format with low ranks, making the overall algorithm scale polynomially in the dimension instead of exponentially. The paper proves this rank structure rigorously for linear-quadratic problems: Theorem 3.1 bounds the TT ranks of the quadratic value function by the off-diagonal ranks of the linearized system matrix times a polylogarithmic factor in the desired accuracy. For nonlinear dynamics, the same low-rank behavior is established numerically: the Allen-Cahn stabilization problem is solved with state dimensions up to 121 and the Fokker-Planck problem with reduced dimensions up to about 20, with maximal TT ranks growing linearly (or stabilizing) in the dimension. The paper also shows that the resulting feedback laws stabilize unstable equilibria at substantially lower cost than the LQR feedback computed from the linearized system, and that control bounds can be enforced through penalty functions.","pith_inferences":["The observed near-linear rank growth for Allen-Cahn suggests a testable conjecture: value functions for feedback stabilization of semilinear parabolic PDEs with localized actuators form a low-rank class, and a systematic study of rank growth across reaction-diffusion, convection-diffusion, and nonlocal systems would map the method's real boundary.","The Fokker-Planck treatment already uses balanced truncation before HJB synthesis; a natural extension is to feed the TT-HJB solver with data-driven or reduced-order coordinates for transport-dominated problems, where direct ranks are likely higher.","Because the rank bound ties feasibility to the off-diagonal ranks of the Jacobian, one could build a cheap pre-screening procedure: compute those ranks for a candidate dynamics before committing to the full HJB solve.","The shifted AMEn linear solver may transfer to time-dependent and stochastic (parabolic) HJB equations, where the same nonsymmetric-degenerate linearized systems appear; that connection is implicit in the paper but not developed."],"forward_implications":["For dynamics whose value function admits small TT ranks, optimal feedback synthesis cost becomes polynomial in state dimension, so state-space dimension is no longer the decisive obstacle for deterministic HJB methods.","HJB-based feedback from the method can stabilize unstable nonlinear equilibria at much lower control cost than LQR linearization, as demonstrated for Allen-Cahn and Fokker-Planck models.","Control constraints can be incorporated through smooth penalty functions, at the price of larger but still manageable TT ranks.","The method extends from semi-discretized one-dimensional PDEs to two-dimensional spatial domains, producing 121-dimensional state systems that are solved in reasonable CPU times on a single core.","The linear-quadratic rank bound provides an a priori criterion: the off-diagonal ranks of the linearized system matrix (and actuator rank) indicate whether the TT approach is promising for a given control system."],"supporting_citations":[{"why":"Defines the tensor-train decomposition and the $O(dnr^2)$ complexity estimate that the polynomial-scaling claim rests on.","marker":"[50]"},{"why":"Supplies the continuous policy iteration/Galerkin framework used to linearize the HJB equation into a sequence of linear PDEs.","marker":"[8]"},{"why":"Provides the AMEn linear solver that the paper modifies with a shift to handle the nonsymmetric degenerate systems.","marker":"[21]"},{"why":"Gives the TT-cross algorithm used to build TT approximations of nonlinear drift and control terms from selected samples.","marker":"[52]"},{"why":"Establishes the polynomial approximation setting for high-dimensional HJB equations that this work extends to tensor-train form.","marker":"[39]"},{"why":"Yields the approximate Kronecker-product inverse and quasi-separable rank bounds used in the proof of Theorem 3.1.","marker":"[28]"},{"why":"Provides the Chebyshev pseudospectral discretization that converts the PDE dynamics into high-dimensional ODE systems.","marker":"[63]"},{"why":"Supplies the regularity analysis for the Fokker-Planck bilinear control value function used in the numerical study.","marker":"[15]"}],"fun_headline_variants":["Tensor trains tame 121-dimensional HJB control","Low-rank HJB scales to 121 variables","Polynomial-time HJB for 121-D feedback","Tensor train HJB: 121D, no exponential blowup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the value functions of the nonlinear problems stay low-rank in tensor-train form as the dimension grows; for nonlinear examples this is observed numerically, not proved, and the conclusion explicitly leaves the identification of low-rank problem classes open.","fun_headline_variants_meta":{"raw":{"variants":["Tensor trains tame 121-dimensional HJB control","Low-rank HJB scales to 121 variables","Polynomial-time HJB for 121-D feedback","Tensor train HJB: 121D, no exponential blowup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001075,"raw_usage":{"total_tokens":4451,"prompt_tokens":850,"completion_tokens":3601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":3537}},"tokens_in":466,"tokens_out":3601,"duration_ms":28779,"temperature":1.0,"reasoning_tokens":3537,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:10:41.281146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a target accuracy and compute the maximal tensor-train rank of the value function for a family of nonlinear control problems (for instance, transport-dominated or nonlocal dynamics) as the state dimension $d$ increases; if the rank required to keep the error fixed grows faster than polynomial in $d$, the polynomial-scaling claim collapses even though each algorithmic step remains correct.","supporting_citations":[{"cited_title":"Beard, George N","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous policy iteration/Galerkin framework used to linearize the HJB equation into a sequence of linear PDEs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the AMEn linear solver that the paper modifies with a shift to handle the nonsymmetric degenerate systems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the TT-cross algorithm used to build TT approximations of nonlinear drift and control terms from selected samples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the polynomial approximation setting for high-dimensional HJB equations that this work extends to tensor-train form."},{"cited_title":"Grasedyck , Existence and computation of low Kronecker-rank approximations for large systems in tensor product structure , Computing, 72 (2004), pp","cited_arxiv_id":null,"evidence_quote":"Yields the approximate Kronecker-product inverse and quasi-separable rank bounds used in the proof of Theorem 3.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Chebyshev pseudospectral discretization that converts the PDE dynamics into high-dimensional ODE systems."},{"cited_title":"Control Op- tim., 56 (2018), pp","cited_arxiv_id":null,"evidence_quote":"Supplies the regularity analysis for the Fokker-Planck bilinear control value function used in the numerical study."}],"review_version":1}