{"id":"1cd14a31-9aa7-461e-a0fd-e8bb516db01e","arxiv_id":"2502.08559","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"For value functions with spatially decaying sensitivity, separable sum-of-local-neighborhood approximations achieve dimension-independent error bounds and neural network representations with polynomially many parameters.","lead":"The paper proves that when the influence between parts of a control system decays with their distance, the system's optimal value function can be approximated by a sum of small local pieces, with errors independent of the system dimension. This connects to neural networks, showing how such structures need only polynomially many parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.4's growth bound (8) counts successors of j, while Lemma 3.2's error sum needs predecessors of j; the directed-graph claim is unproved as stated.","rationale":"The reader's weakest_assumption points to the unverified status of Definition 2.1 for nonlinear optimal control. That is a fair conditionality, but it is explicit in the paper: the theorems are conditional, and the LQR cases are proved. The more specific internal weakness is the direction mismatch in the proof of Theorem 3.4. I checked that Lemma 3.2's summation over i with dist_G(i,j)>l matches the in-neighborhood of j, while condition (8) bounds the out-neighborhood of j. Since Theorem 3.4's proof depends directly on (8), the theorem as stated for directed graphs is not established. Because the advertised applications (banded LQR, heat equation, Allen-Cahn discretization) all use undirected or sequential graphs, the fix does not alter the main practical claims, and the reader's CONDITIONAL verdict remains appropriate. I therefore recommend no change to the verdict, while flagging that a revision should correct (8) or restrict Theorem 3.4 to undirected graphs.","tokens_in":24145,"tokens_out":22394,"duration_ms":253510,"concrete_test":"Re-derive Lemma 3.2 keeping the summation sets explicit: for each j, determine whether the telescoping term Psi_j^{l+1}-Psi_j^l sums over {i : dist_G(i,j)=l+1} or over {i : dist_G(j,i)=l+1}. If it is the former, replace condition (8) with r'(l) = max_j |{i : dist_G(i,j)=l}| (or require both in- and out-neighborhood bounds) and check whether the proof of Theorem 3.4 goes through unchanged. To settle whether the current statement is false rather than merely under-proved, exhibit a directed graph with bounded out-neighborhoods but unbounded in-neighborhoods on which Definition 2.1 is satisfiable; if the claimed error bound then fails, Theorem 3.4 needs an added assumption or a restriction to undirected graphs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of the central error estimate, Lemma 3.2 telescopes Psi_j over the newly added subsystems i for which dist_G(i,j)=k+1, i.e. the in-neighborhood of j: it then applies Definition 2.1, whose Lipschitz constant for g_j with respect to z_i is gamma(dist_G(i,j))*||z_j||. But the growth bound (8) is |{i : dist_G(j,i)=l}| <= r(l), i.e. a bound on the out-neighborhood of j. These two cardinalities agree only for undirected graphs. Theorem 3.4 uses (8) to bound the row and column sums of D_l, so for a genuinely directed graph the s,n-independent bound does not follow from the stated assumptions. This is not a cosmetic issue: Section 2 explicitly frames Setting 2.1 with a directed graph and asserts that the main results of Section 3 hold for both directed and undirected graphs, and Theorem 3.4 is the basis of the neural network result Theorem 5.1. The fix is straightforward: restrict Theorem 3.4 to undirected graphs, or add a growth bound on |{i : dist_G(i,j)=l}| (or require both in- and out-neighborhood bounds). All LQR and numerical examples in Section 6 use undirected or sequential graphs, so the practical conclusions are not endangered. The nonlinear verification issue identified by the reader is an honest hypothesis of the theorem; the directed-graph mismatch is an internal proof gap.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops a framework for approximating high-dimensional value functions by separable sums of local contributions. Setting 2.1 formalizes a gamma-decaying sensitivity property of a function V with respect to a directed graph of interacting subsystems. The central construction (4)-(5) approximates V(x) by V(0)+Psi_l(x)=V(0)+sum_j Psi^j_l(z_{j,l}), where each local term depends on a bounded graph neighborhood of subsystem j. Lemma 3.2 bounds the approximation error by x^T D_l x, and Theorems 3.4 and 3.5 derive explicit error bounds under exponential and polynomial sensitivity decay together with graph growth bounds, with constants independent of the number of agents s and the dimension n. Section 4 verifies the sensitivity property for discrete-time LQR (citing prior work) and for a continuous-time LQR class with symmetric A, B=I, Q=cI, R=gamma I (Proposition 4.3). Section 5 proves a neural-network representation result: under the exponential-decay assumptions, a feedforward network with polynomially many neurons and parameters achieves accuracy epsilon(1+||x||^2) on [-R,R]^n. Section 6 reports numerical experiments for LQR problems, non-quadratic functions with prescribed decay, and an Allen-Cahn SDRE surrogate, with code available online.","tokens_in":24477,"tokens_out":12038,"duration_ms":118221,"significance":"If the main theorems hold as stated, the paper gives explicit, dimension-independent approximation guarantees for a well-defined class of value functions and a constructive neural-network complexity bound, thereby offering a principled route around the curse of dimensionality for separable structures. The strongest assets are the explicit constants in Theorems 3.4 and 3.5, the closed-form continuous-time LQR decay result in Proposition 4.3, the explicit connection to anchored decompositions, and a reproducible numerical study that compares the separable architecture against fully connected networks and tests non-quadratic and polynomial-decay examples. The paper is also honest in labeling the nonlinear verification in Section 6.3 as numerical evidence rather than proof. The main caveat is the directed-graph proof gap identified below, which is localized and does not affect the undirected and sequential examples actually treated in Section 6.","major_comments":[{"comment":"The dimension-independent bound for genuinely directed graphs is not proved as stated. Lemma 3.2's telescoping sum runs over nodes i with dist_G(i,j)=k+1, i.e. over the in-neighborhood of j, and the column sums of D_l that give ||D_l||_1 in Theorem 3.4 require a bound on |{i : dist_G(i,j)=l}|. The growth bound (8), however, is stated as |{i : dist_G(j,i)=l}| <= r(l), i.e. an out-neighborhood bound. These two cardinalities agree only when G is undirected. Therefore the assertion in Setting 2.1 that the main results of Section 3 hold for directed graphs is unsupported, and Theorem 3.4 -- and hence the neural-network bound in Theorem 5.1 -- needs either to be restricted to undirected graphs or supplemented by an analogous in-neighborhood growth assumption. The fix is local, and the practical conclusions are unaffected because all LQR and numerical examples use undirected or sequential graphs.","section":"Section 3.2, Eq. (8), Lemma 3.2, Theorem 3.4"},{"comment":"The applicability of the theory to nonlinear control is presented as empirically supported rather than proven, and this should be stated more explicitly as a limitation. The Lipschitz estimates in Figure 13 are first-order difference quotients computed at two reference points with one perturbation per coordinate; they do not certify the uniform, state-dependent inequality in Definition 2.1 on the whole domain. I do not regard this as an error, because the wording 'strong numerical evidence' is honest, but the abstract and Section 7 should state that for nonlinear problems the gamma-decaying sensitivity assumption is a conjecture supported by numerical evidence rather than a theorem, so that readers do not infer that Theorem 5.1 has been verified for the Allen-Cahn problem.","section":"Section 6.3 and Section 7"}],"minor_comments":[{"comment":"The set of newly added nodes is written once as {i : dist_G(j,i)=l~} and later as {i : dist_G(i,j)=k+1}. Please use a single orientation, preferably dist_G(i,j), throughout the proof, since the definition of z_{j,l} in (4)-(5) uses the distance from i to j.","section":"Proof of Lemma 3.2"},{"comment":"As typeset, the displayed chain of inequalities contains an extra factor delta^{l+1}: it reads '<= C b_C delta^{l+1}/(1-delta) delta^{l+1}', which would make the constant \tilde C depend on l. The intended bound should be C b_C delta^{l+1}/(1-delta), with \tilde C = C b_C/(1-delta). Please correct this display.","section":"Proof of Theorem 3.4"},{"comment":"There is an inconsistency in the reported maximum number of epochs for the 50-dimensional case with nu_g=nu_z=0: Table 1 says training stopped after max epochs 1000, while the text says the MSE was 1.3e-2 after 10^4 epochs. Please align the numbers.","section":"Table 1 and surrounding text"},{"comment":"The sentence beginning 'or a detailed discussion of the theoretical underpinnings' should read 'For a detailed discussion...'.","section":"Section 6.3"},{"comment":"The definition of l uses nonstandard bracket notation and should specify the floor/ceiling operation and require epsilon small enough that the chosen l is nonnegative; currently the formula is hard to parse and appears as 'l := l ... m'.","section":"Proof of Theorem 5.1"}],"recommendation":"major_revision","confidential_remarks":"The directed-graph gap is the only substantive correctness issue I see; it is localized and fixable by restricting Theorem 3.4 to undirected graphs or adding an in-neighborhood growth assumption. The nonlinear verification concern is a claim-scope issue rather than a proof error. Neither ground justifies rejection, so I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper deserves a serious referee. The core contribution is a clean, conditional framework: if a value function satisfies a decaying-sensitivity property with respect to a graph, and the graph's neighborhoods grow slowly enough, then the separable approximation from the authors' CDC 2023 paper has error independent of the number of agents and dimension. The new material beyond [33] is real: the s-independent exponential and polynomial error bounds (Theorems 3.4 and 3.5), the closed-form continuous-time Riccati decay proof via Benzi–Golub (Proposition 4.3), the anchored-decomposition link, and the polynomial-size neural network theorem (Theorem 5.1). The proofs for undirected graphs are coherent, and the paper is honest about the conditional nature of the main assumption.\n\nThe soft spot is exactly where the stress-test lands. Theorem 3.4 is stated for directed graphs, but its proof uses a growth bound on out-neighborhoods (condition (8)) to control both row and column sums of the matrix D_l. For a genuinely directed graph, bounding the column sum requires a bound on in-neighborhoods, and those two counts are not interchangeable. The same confusion appears in the proof of Lemma 3.2, where the set of indices {i : dist_G(j,i)=l} is written when the construction of z_{j,l} actually relies on dist_G(i,j). This is an internal proof gap, not a load-bearing flaw for the paper's main message: every example in Section 6 uses undirected or sequential graphs, and Section 4 explicitly restricts to undirected graphs. The fix is straightforward—restrict Theorem 3.4 to undirected graphs or add a matching in-neighborhood growth bound.\n\nOn the numerical side, the nonlinear Allen–Cahn example uses an SDRE surrogate as ground truth, which the authors disclose. That makes the experiment a consistency check with the surrogate rather than a direct test of the true value function. It is a reasonable first step, but not proof of the assumption for nonlinear problems. The lack of error bars is minor given the large performance gaps between S-NN and fully connected networks.\n\nOverall, the central conditional claim is well supported for the undirected setting, the main theorems are new relative to the conference version, and the paper is not circular—the decaying-sensitivity assumption is defined independently and verified for LQR. I would send this to peer review with a note to the authors about the directed-graph gap.","headline":"Worth a serious referee: genuine new results and a clear conditional framework, but the directed-graph claim of Theorem 3.4 is not proved as stated.","tokens_in":716,"tokens_out":1813,"would_cite":true,"duration_ms":54039,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N10","49L20","65D15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A value function with decaying sensitivity has a separable approximation whose error does not grow with the number of agents or dimension, and a neural network with polynomially many parameters can represent it.","keywords":["separable approximations","decaying sensitivity","value functions","neural networks","optimal control","curse of dimensionality","Riccati equations"],"falsifier":"For a fixed $\\gamma(l)=C\\rho^l$ and a graph whose neighborhood growth $r(k)$ exceeds $b\\mu^k$ with $\\mu\\ge\\rho^{-1}$, compute $\\|D_l\\|_2$ and the actual separable error $|V-\\Psi_l-V(0)|$ as $s$ grows; Theorem 3.4 predicts the bound loses its $s$-independence in this regime, so observing the error remain uniformly bounded would falsify the role of the growth condition.","tokens_in":2161,"feed_emoji":"🧠","tokens_out":4844,"duration_ms":125175,"temperature":0.7,"pith_summary":"Decaying sensitivity means the influence of one subsystem's state on another fades as their graph distance grows. The paper proves that any function with this property, and in particular optimal-control value functions from LQR problems, can be approximated by the sum of local terms involving only nearby subsystems. Its main estimate, Theorem 3.4, holds when sensitivity decays exponentially and graph neighborhoods grow at most exponentially with a compatible base; then the approximation error is bounded by a constant times $\\delta^{l+1}\\|x\\|^2$, with constants independent of the number of agents $s$ and dimension $n$. That result is used to show that a feedforward neural network with a number of neurons and parameters polynomial in $s$ can approximate the value function to any desired accuracy on a bounded box.","feed_headline":"Decaying sensitivity yields polynomial-size neural nets","feed_subtitle":"For value functions whose influence fades with distance, approximation costs grow polynomially, not exponentially.","key_machinery":"The central object is the $\\gamma$-decaying sensitivity property (Definition 2.1): for the maps $g_j(x)=V(x)-V(x\\text{ with }z_j=0)$, the Lipschitz constant with respect to $z_i$ is at most $\\gamma(\\operatorname{dist}_G(i,j))\\|z_j\\|$. This condition links the value function to the graph geometry. The second ingredient is the separable ansatz $\\Psi_j^l(z_{j,l})=V(x_{j,l})-V(\\Lambda_j x_{j,l})$, where $x_{j,l}$ keeps only $z_j$ and its graph ball of radius $l$. Lemma 3.3, the norm inequality $\\|A\\|_2\\le\\sqrt{\\|A\\|_1\\|A\\|_\\infty}$, converts row- and column-sum estimates of the truncated decay matrix $D_l$ into the final error. The graph growth bound $r(k)$ controls these sums uniformly as $s$ grows. In Theorem 5.1, the same local terms are the targets of independent shallow subnetworks, which are then concatenated into one feedforward network.","core_discovery":"Under the $\\gamma$-decaying sensitivity property, the separable approximation $V(x)\\approx V(0)+\\sum_{j=1}^s \\Psi_j^l(z_{j,l})$ has error bounded by $\\|x\\|_2^2 \\|D_l\\|_2$, where $D_l$ collects the decay constants for pairs at graph distance greater than $l$. Theorem 3.4 sharpens this to $|V(x)-\\Psi_l(x)-V(0)|\\le \\tilde C \\delta^{l+1}\\|x\\|_2^2$ when $\\gamma(l)=C\\rho^l$ and the graph growth satisfies $r(k)\\le b\\mu^k$ with $\\mu<\\rho^{-1}$, with $\\delta\\in(\\rho\\mu,1)$ and constants independent of $s$ and $n$. For polynomial decay $\\gamma(l)=C(l+1)^{-\\alpha}$ and $r\\in O(l^\\beta)$ with $\\beta<\\alpha-1$, Theorem 3.5 gives an analogous bound with a tail sum. Theorem 5.1 then constructs a feedforward network with $s^{d+1}C^d+n+1$ neurons and $s^{d+1}C^d(d+s+1)+1$ parameters that approximates $V$ to error $\\varepsilon(1+\\|x\\|_2^2)$ on $[-R,R]^n$. The paper also proves an explicit off-diagonal decay rate for the solution of the continuous-time Riccati equation when $A$ is symmetric and banded, and provides numerical evidence for exponential sensitivity decay in a nonlinear control problem.","pith_inferences":["The same separable construction might be used to design control laws for PDE-constrained multi-agent systems if the decaying-sensitivity property is verified for their value functions; the nonlinear example in the paper uses an SDRE surrogate rather than a direct proof.","Replacing the fixed anchor $0$ by data-dependent anchors could improve the constants, since the construction is a graph-localized anchored decomposition.","The numerical finding that two neurons per sublayer suffice across dimensions suggests the worst-case polynomial exponents in Theorem 5.1 may be pessimistic, and a sharper bound might depend on the range of local sensitivity rather than the maximal neighborhood dimension.","An independent test would be to train the $s$ local terms $\\Psi_j^l$ separately and sum them, then compare accuracy and training cost against the end-to-end separable network."],"forward_implications":["For a family of problems with uniformly decaying sensitivity and uniformly bounded graph-neighborhood dimension, value functions can be represented by neural networks with polynomial parameter growth, bypassing the exponential curse of dimensionality.","The local radius $l$ sets an explicit accuracy-versus-sparsity trade-off: larger $l$ reduces the error as $\\delta^{l+1}\\|x\\|^2$ but increases the dimension of the local terms.","Discrete-time LQR problems that satisfy the uniform stabilizability and detectability conditions inherit the separable error bounds through the spatially exponential decay of the Riccati solution.","For continuous-time LQR with symmetric banded $A$, $B=I$, $Q=cI$, $R=\\gamma I$, the paper proves an explicit off-diagonal decay rate for the Riccati solution, which feeds directly into the separable approximation.","Standard low-rank approximation cannot exploit this structure, so the separable network architecture offers a complementary route for high-dimensional value functions."],"supporting_citations":[{"why":"Introduces the separable approximation construction and shows exponential decay of the discrete-time LQR Riccati solution.","marker":"[33]"},{"why":"Proves the spatially exponential decay of the optimal feedback matrix for networked LQR, the starting point for the discrete-time analysis.","marker":"[30]"},{"why":"Defines spatially exponentially decaying matrices, used in Corollary 3.1 to connect SED structure to decaying sensitivity.","marker":"[34]"},{"why":"Provides the matrix-function off-diagonal bound used in Proposition 4.3 for the continuous-time Riccati solution.","marker":"[5]"},{"why":"Supplies the norm inequality $\\|A\\|_2\\le\\sqrt{\\|A\\|_1\\|A\\|_\\infty}$ used in the proofs of Theorems 3.4 and 3.5.","marker":"[11]"},{"why":"Gives the one-hidden-layer approximation rate for smooth functions with non-polynomial activation, used in Theorem 5.1.","marker":"[23]"},{"why":"Extends the approximation result to cubes, enabling the statement of Theorem 5.1 on $[-R,R]^n$.","marker":"[17]"}],"fun_headline_variants":["Decaying sensitivity yields polynomial-size neural nets","Dimensionality curse lifted by decaying sensitivity","Separable approximations beat exponential scaling","Neural networks of polynomial size via decaying sensitivity","Decaying sensitivity eases value function approximation"],"cache_read_input_tokens":27008,"weakest_assumption_plain":"The entire argument rests on Definition 2.1, the assumption that the value function's cross-sensitivity has a Lipschitz constant decaying as $\\gamma(\\operatorname{dist}_G(i,j))\\|z_j\\|$; for nonlinear problems this property is only verified numerically from an SDRE surrogate, not proved.","fun_headline_variants_meta":{"raw":{"variants":["Decaying sensitivity yields polynomial-size neural nets","Dimensionality curse lifted by decaying sensitivity","Separable approximations beat exponential scaling","Neural networks of polynomial size via decaying sensitivity","Decaying sensitivity eases value function approximation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000635,"raw_usage":{"total_tokens":2957,"prompt_tokens":1001,"completion_tokens":1956,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":1891}},"tokens_in":617,"tokens_out":1956,"duration_ms":13800,"temperature":1.0,"reasoning_tokens":1891,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T04:38:48.255568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed $\\gamma(l)=C\\rho^l$ and a graph whose neighborhood growth $r(k)$ exceeds $b\\mu^k$ with $\\mu\\ge\\rho^{-1}$, compute $\\|D_l\\|_2$ and the actual separable error $|V-\\Psi_l-V(0)|$ as $s$ grows; Theorem 3.4 predicts the bound loses its $s$-independence in this regime, so observing the error remain uniformly bounded would falsify the role of the growth condition.","supporting_citations":[{"cited_title":"Sperl, L","cited_arxiv_id":null,"evidence_quote":"Introduces the separable approximation construction and shows exponential decay of the discrete-time LQR Riccati solution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves the spatially exponential decay of the optimal feedback matrix for networked LQR, the starting point for the discrete-time analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines spatially exponentially decaying matrices, used in Corollary 3.1 to connect SED structure to decaying sensitivity."},{"cited_title":"Benzi and G","cited_arxiv_id":null,"evidence_quote":"Provides the matrix-function off-diagonal bound used in Proposition 4.3 for the continuous-time Riccati solution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the norm inequality $\\|A\\|_2\\le\\sqrt{\\|A\\|_1\\|A\\|_\\infty}$ used in the proofs of Theorems 3.4 and 3.5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the one-hidden-layer approximation rate for smooth functions with non-polynomial activation, used in Theorem 5.1."},{"cited_title":"Kang and Q","cited_arxiv_id":null,"evidence_quote":"Extends the approximation result to cubes, enabling the statement of Theorem 5.1 on $[-R,R]^n$."}],"review_version":1}