{"id":"eb3dfb99-6657-45fb-8394-4cfe0fe3661b","arxiv_id":"2608.00217","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Truncated backpropagation through the final r iterations of a preconditioned PDHG solver recovers inverse-cost maps in mean-field games as accurately as full unrolling while using a fraction of the memory.","lead":"The paper introduces a way to train inverse mean-field-game models by backpropagating through only the last few iterations of a primal-dual solver, rather than the entire solver trajectory. This cuts memory and runtime while keeping recovery accuracy close to full unrolling in several numerical tests.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.5's assumptions (exact prox, fixed active set, exact detach) are all violated in the experiments, and Table 2 shows order-one gradient bias at the tested depths, so the theory does not quantitatively underwrite the headline moderate-depth speedup.","rationale":"I read the paper in good faith: the abstract and Section 4 state the consistency theorem as conditional on Assumption 4.1, exact-proximal PDHG, and exact equilibrium detachment, and the experiments are reported with seed spreads and documented instabilities (Appendix G.3), which is honest. The proofs I checked (metric resolvent identity in Theorem 3.2, kernel argument in Theorem 3.6, fixed-active-set sensitivity in Lemma 4.2 and Proposition 4.4) are structurally sound. The central concern is not an internal inconsistency but a missing quantitative bridge: the implemented algorithm eTθ differs from the analyzed map Tθ in exactly the ways that matter for the finite-depth guarantee, and the paper's own alignment diagnostics show the truncated gradient is only roughly aligned (bias of order one at r=10–20), not close to the implicit gradient. Corollary E.2 is the only finite-depth statement and its key hypothesis (a verified grouped contraction) is not checked. Whether this gap is benign or fundamental is testable by the exact-prox ablation described above. Because the reader already marked the paper CONDITIONAL and our concern sharpens the same conditions without invalidating the empirical results, the verdict should remain CONDITIONAL (i.e., unchanged from the reader's CONDITIONAL).","tokens_in":1295,"tokens_out":1365,"duration_ms":126232,"concrete_test":"On a small (e.g., 16×16×8) instance, run JFB-r in three configurations: (A) exact cell-prox solves to ~1e-13 (much more Newton budget) with detach at the exact equilibrium; (B) exact prox with the paper's warm detach (residual ~1e-7); (C) the implemented 12-Newton prox with warm detach. At the equilibrium and at several outer checkpoints, compute d_JFB_r for r∈{1,5,10,25,50,100} and a branch-matched implicit reference d_ID, and report the relative bias ‖d_JFB_r−d_ID‖/‖d_ID‖ and its cosine. If (A) shows bias → 0 with r while (C) plateaus at order-one bias, the implementation gap is the operative limitation; if (A) also plateaus, Theorem 4.5's conditions are insufficient even in the idealized regime; if (C) also shows bias → 0 at r≈25, the gap is benign and the practical claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is Theorem 4.5: under Assumption 4.1 (locally fixed flux active set with strict complementarity), the exact-proximal map Tθ of (2.15), and exact equilibrium detachment z̄=z0, d_JFB_r → ∇θF as r→∞. Every one of those hypotheses is violated in the experiments: the code uses eTθ with 12 damped Newton prox steps (Section 2.4), the detach residual has median 2e-7 rather than zero (Appendix G.7), and at the true-cost equilibrium 16 flux slots are degenerate while active sets change in bursts of up to 48 slots during training (Appendix G.7). The paper concedes (end of Section 4, Section 5.1) that Theorem 4.5 'does not apply directly' and that the practical claim rests on branch-consistent diagnostics. Those diagnostics, however, show median relative gradient bias of 0.64 (r=10) and 0.60 (r=20) against the branch-matched implicit reference (Table 2): the truncated direction is not a quantitatively accurate surrogate gradient at the depths used. The only finite-depth bound, Corollary E.2, requires a numerically verified grouped contraction q_M<1 that is not supplied. The load-bearing question is whether the exact-proximal resolvent theory has any quantitative connection to the implemented solver at moderate depth; if it does not, the empirical speedup, while real on these three-seed runs, is not explained by the paper's analysis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies inverse potential mean-field games, in which an unknown spatial inverse-cost map is inferred from observed densities, and proposes a truncated differentiation scheme called JFB-r applied to a preconditioned primal-dual hybrid gradient (PDHG) forward solver. The main theoretical contributions are: (i) the exact-proximal dual-extrapolated PDHG map is identified as a metric resolvent of the maximal monotone KKT operator (Theorem 3.2); (ii) directional stability of the derivative resolvent is established without assuming flux strict complementarity (Theorem 3.6); and (iii) under a locally fixed active set with strict complementarity and exact equilibrium detachment, the JFB-r direction is shown to converge to the exact implicit gradient as the tracked depth r tends to infinity (Theorem 4.5). Numerically, the paper reports four inverse-MFG experiments in which moderate depths (r=10 or 25) reduce peak memory and runtime relative to full unrolling while achieving recovery errors close to those of full AD. The paper is explicit that the implemented solver uses an approximate proximal map, warm detached starts, and changing active sets, so the exact-proximal fixed-active-set theorem does not apply directly to the experiments.","tokens_in":32475,"tokens_out":6618,"duration_ms":78257,"significance":"If the results hold, the resolvent identification is a valuable theoretical contribution: it extends Jacobian-free backpropagation-style analysis beyond contractive fixed-point maps and gives a clean interpretation of JFB-r as a truncated adjoint iteration. The mathematical derivations in Sections 3-4 and Appendices C-F appear carefully executed, and the paper is unusually transparent about the gap between its assumptions and its implementation. The numerical experiments are useful and suggest that the method has practical merit. However, the central quantitative claim about moderate depths is not connected to the theory by any available bound: the theorem applies to a different map, at exact detachment, with a fixed active set, and the only finite-depth bound is conditional on an unverified contraction certificate. As it stands, the theoretical significance for the actually implemented method is therefore not yet established, while the empirical contribution stands as a separate, weaker claim.","major_comments":[{"comment":"Theorem 4.5 assumes the exact-proximal map T_theta, a locally fixed flux active set with strict complementarity (Assumption 4.1), and exact equilibrium detachment z-bar = z_0. Each of these hypotheses is violated in Algorithm 2.1 and in the experiments: the implemented map eT_theta replaces exact proximal solves by 12 damped Newton iterations (Section 2.4, Appendix A.4), Appendix G.7 reports active-set changes in bursts of up to 48 slots and 16 degenerate flux slots at the true-cost equilibrium, and the detach residual has median 2e-7 rather than zero. The paper concedes in Sections 4 and 5.1 that Theorem 4.5 does not apply directly. Because this theorem is the only theoretical result connecting tracked depth to the implicit gradient, the moderate-depth speedup claims in Sections 5.2 and 5.3 are not underwritten by the analysis. The manuscript should either provide a quantitative extension covering approximate prox and warm detachment, or clearly separate the asymptotic exact-regime theorem from the empirical approximate-regime results and soften the corresponding contribution claims.","section":"Theorem 4.5 and Section 5.1"},{"comment":"The descent-phase alignment diagnostics show median relative gradient bias of 0.64 at r=10 and 0.60 at r=20 against the branch-matched implicit reference; the bias remains of order one at every tested depth. The only finite-depth error bound, Corollary E.2, is conditional on a numerically verified grouped contraction q_M<1 that is not supplied. Consequently, the paper provides no quantitative statement that the directions used at r=10, r=20, or r=25 are accurate surrogates for the true gradient. The recovery comparisons in Tables 4 and 5 are therefore empirical observations rather than consequences of the theory. The text should state this limitation explicitly and, ideally, add a direct attempt to estimate the relevant finite-depth operator norms or a sensitivity study that bridges the exact-proximal regime.","section":"Section 5.1, Table 2"},{"comment":"The adjoint-iteration interpretation in (4.8) and Appendix F is elegant, but the alignment is established in the M-geometry of the adjoint space, while the parameter-space gradient is obtained by a final compression through B B^T. Appendix F acknowledges that descent can be lost in this compression, and Table 2 confirms that the parameter-space bias is large in practice. Since the paper's practical claim is about parameter-space optimization, the M-alignment result does not by itself explain the observed recovery behavior. The authors should either prove a parameter-space bound under additional assumptions on B, or avoid phrasing that suggests the adjoint alignment transfers to the gradient used in training.","section":"Section 4.3 and Appendix F"},{"comment":"Corollary E.2 is the only quantitative finite-depth statement in the paper, but it is conditional and non-constructive: the existence of (r*, q_M) follows from Gelfand's formula, with no explicit estimate of r* or q_M. For a paper whose headline is that moderate tracked depths suffice, a non-constructive existence result does not provide the needed quantitative certificate at the tested depths. The authors should report an attempt to verify the grouped contraction numerically at representative checkpoints, or state plainly that no finite-depth theoretical guarantee is available for the depths used in the experiments.","section":"Corollary E.2 and Section 5.1"}],"minor_comments":[{"comment":"The sentence 'JFB-10 gives up about 0.2 percentage points of accuracy for a 2.4x speedup' is ambiguous; it should read that JFB-10 is about 0.2 percentage points less accurate than AD, since the reported errors are 1.13% for JFB-10 versus 0.95% for AD.","section":"Section 5.1, Table 1 discussion"},{"comment":"The alignment table evaluates r=5,10,20 retrospectively at states visited by JFB-10 trajectories, so the depth comparison is not made along each depth's own optimization trajectory. This caveat is present in the appendix but should be stated more prominently next to Table 2.","section":"Section 5.1, Table 2 and Appendix G.2.2"},{"comment":"Over all 63 checkpoints the depth ordering by median cosine is non-monotone, with r=5 having median cosine -0.53 and r=1 having +0.21. The main text should not overinterpret the descent-phase ordering without acknowledging this full-set variability.","section":"Appendix G.2.2, Table 7"},{"comment":"The repository link in Appendix G says the code 'will be released'; for a submission claiming reproducibility, a permanent and accessible code artifact should be provided at review time.","section":"Appendix G"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and the mathematical core is sound, but the theoretical and experimental parts are not quantitatively connected. I recommend major revision rather than rejection because the resolvent theory is a genuine contribution and the empirical results are promising. The main requested changes are to provide a quantitative bridge between Theorem 4.5 and the approximate/warm/changing-active-set regime, or to substantially re-scope the claims so that the theoretical consistency result is not presented as explaining the moderate-depth performance in the experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The contribution is real but narrower than the title suggests. The genuinely new content is Theorem 3.6, directional stability of the PDHG derivative resolvent at degenerate flux kinks without strict complementarity, and the large-depth consistency theorem under a locally fixed active set. The resolvent identity itself is classical, as the paper itself concedes; what is useful is the fixed-active-set truncation analysis and the branch-consistent gradient diagnostics.\n\nWhat the paper does well: the mathematics in Sections 3–4 is structurally sound. The kernel argument in Theorem 3.6 uses strong convexity, density interiority, and injectivity of D_t^T, and it works. The experiments are reported honestly: divergent 64^2 runs are included, active-set changes are documented (bursts of up to 48 slots), detach residuals are measured, and recovery is measured against ground truth rather than fitted to the data. No fitted value is disguised as a prediction. That is good practice and it earns credit.\n\nThe soft spots are real but proportionate. Theorem 4.5 assumes the exact-proximal map, a locally fixed flux active set with strict complementarity, and exact equilibrium detachment. The experiments violate all three. Table 2 shows median relative gradient bias of 0.64 at r=10, so the truncated direction is aligned with the implicit reference but not quantitatively close. The paper states this repeatedly and does not pretend the theory underwrites the moderate-depth speedup; it leans on diagnostics. That is honest, but the load-bearing question remains: does the exact-proximal resolvent theory have a quantitative connection to the implemented 12-damped-Newton map at moderate depth? The only finite-depth bound offered, Corollary E.2, requires a numerically verified grouped contraction q_M that is not supplied. Code is not yet released. These are addressable and are the natural referee asks.\n\nMinor: the depth ordering in the full-set alignment statistics is non-monotone (r=5 median cosine is negative over all 63 checkpoints), so the empirical story leans on the descent-phase restriction. The 128^2 comparison is a single resource-bound run, which the paper labels correctly.\n\nWho this is for: researchers doing inverse MFG or training through fixed-point solvers. The practical claim, that moderate depths give AD-comparable recovery at reduced memory and runtime, is plausibly supported by the experiments. The theory-experiment gap should be fixed by a quantitative deviation bound, but its absence does not sink the paper. It deserves a serious referee. I would send it out, with requests for code release and a direct comparison of exact-proximal versus finite-Newton directions at the same states.","headline":"Solid analysis of truncated differentiation through PDHG for inverse MFGs, with an honest but unresolved gap between the exact-proximal theory and the finite-Newton experiments.","tokens_in":32976,"tokens_out":1631,"would_cite":true,"duration_ms":20823,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N80","65K10","49M41","47H05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Truncated gradients through 25 PDHG steps match full backpropagation in inverse mean-field games at a quarter of the memory.","keywords":["inverse mean-field games","Jacobian-free backpropagation","primal-dual hybrid gradient","monotone operator resolvent","implicit differentiation","truncated backpropagation","inverse-cost map recovery"],"falsifier":"Measure the relative bias $\\lVert d^{\\mathrm{JFB}}_r - d^{\\mathrm{ID}} \\rVert / \\lVert d^{\\mathrm{ID}} \\rVert$ against a cold-solved implicit reference on a problem with degenerate flux slots and persistent active-set changes: if increasing $r$ from 10 to hundreds does not reduce the bias toward zero, or if cosine alignment with the implicit gradient stays below about 0.5 at large $r$, Theorem 4.5's consistency is contradicted where it was meant to apply. A cheaper check: on a task where the warm start is far from equilibrium (large detach residual), increasing $r$ should still improve recovery if the practical claim holds; failure would show the near-equilibrium detachment is indispensable.","tokens_in":31772,"feed_emoji":"🎯","tokens_out":11726,"duration_ms":114848,"temperature":0.7,"pith_summary":"The paper asks whether an inverse potential mean-field game — recovering the unknown mobility cost that moves a population, from observed density data — can be trained by differentiating through only the tail of the forward solver instead of its entire trajectory, which would avoid both the large tape storage of full unrolling and the extra linear solves of adjoint-based implicit differentiation. It claims yes: run the full preconditioned primal-dual hybrid gradient (PDHG) solve, but record and backpropagate only the final $r$ iterations from a detached warm start, and the resulting JFB-$r$ direction converges to the exact implicit gradient $\\nabla_\\theta F(\\theta_0)$ as the tracked depth grows. The load-bearing identity is that the dual-extrapolated PDHG map with exact proximal steps is the metric resolvent $(I + M^{-1}A_\\theta)^{-1}$ of the game's maximal monotone KKT operator, which replaces the contraction condition that earlier Jacobian-free backpropagation proofs required and yields, under a locally fixed active set at an exact equilibrium detach point, geometric bias decay with depth. In experiments, moderate depths suffice: JFB-25 is statistically indistinguishable from full automatic differentiation under noisy, partial-in-time observation while using roughly a quarter of the peak memory and half the runtime, and JFB-10 reduces memory by up to a factor of nine in a multi-instance task.","feed_headline":"25 tracked solver steps match full backprop at a quarter of the memory","feed_subtitle":"A resolvent identity shows truncated gradients converge to the true gradient at half the runtime.","key_machinery":"The carrying object is the metric resolvent identity $T_\\theta = (I + M^{-1}A_\\theta)^{-1}$ (Theorem 3.2): the dual-extrapolated, Schur-preconditioned PDHG update coincides with the unit resolvent of the maximal monotone KKT operator $A_\\theta(x,\\phi) = (\\partial f_\\theta(x) + K^\\top\\phi,\\; b - Kx)$ in the weighted metric $M$ built from the primal and dual step operators. This converts the analysis from the contraction assumption that standard JFB guarantees require — which PDHG lacks, being merely averaged and nonexpansive — into firm nonexpansiveness in the $M$-inner product, and makes the map semi-differentiable with derivative resolvent $R_0 = (I + M^{-1}DA)^{-1}$. The derivative resolvent has trivial kernel (strong convexity of the forward objective plus the injective temporal block $D_t$), which forces grouped directional decay even at flux kinks without strict complementarity; under the locally fixed active set of Assumption 4.1 the resolvent becomes the linear map $L = (M+H)^{-1}M$ with spectral radius below one, so the tracked sensitivities obey $u_{k+1} = Lu_k + P$ and JFB-$r$ is the truncated adjoint iteration of (4.8).","core_discovery":"On the paper's own terms, the central discovery is a sharp trade-off: in inverse potential mean-field games the differentiation depth $r$ can be truncated severely without losing gradient quality. The paper proves (Theorem 4.5) that for the exact-proximal map, with exact equilibrium detachment and a locally fixed active set, $d^{\\mathrm{JFB}}_r(\\theta_0) \\to \\nabla_\\theta F(\\theta_0)$ as $r \\to \\infty$, with error $-P^\\top (L^\\top)^r (I - L^\\top)^{-1} g$; the truncated direction is exactly $r$ metric-preconditioned iterations toward the implicit adjoint $H^{-\\top}g$. The practical demonstration is that moderate depths suffice: JFB-25 matches full unrolling within three-seed variability in the hardest setting, using roughly a quarter of AD's peak memory and half its runtime, and JFB-10 runs 3.5x faster than AD in a multi-instance task with one-ninth the memory. The authors state the scope precisely: the consistency theorem is local and holds for the exact-proximal map, not directly for the implemented finite-Newton map with changing active sets.","pith_inferences":["The metric-resolvent argument is not specific to mean-field games, since it only uses maximal monotonicity of the KKT operator and metric positivity: the same truncated-differentiation guarantee should carry over to other PDHG-solved saddle problems such as dynamic optimal transport, imaging inverse problems, and convex games, a transfer the paper mentions as future work only in passing.","If the grouped-decay bound on $L$ is taken literally, a principled adaptive schedule suggests itself: grow $r$ early in training when the warm-start detach residual is large, and shrink it once the model is near equilibrium — a testable extension that would automate the manual $r = 25$ choice.","The analysis predicts that JFB-$r$ becomes more reliable as training converges, because the detached warm start approaches a fixed point; one could verify this by tracking alignment of JFB-$r$ with the implicit reference across outer iterations, which the reported diagnostics already show improving in the descent phase."],"forward_implications":["Inverse-MFG training can recover the mobility map with $r = 10$–$25$ tracked steps, reaching errors statistically equal to full unrolling while using roughly a quarter to a ninth of the peak memory and half or less of the runtime.","At a $128^2$ spatial grid plain full unrolling exceeds a 48 GB GPU, while JFB-10 finishes at 17.5 GB and about 5.2x faster than checkpointed AD, at a modest recovery-error cost.","For any differentiable tracked map JFB-$r$ is the exact gradient of an $r$-step surrogate, so it is always a valid descent direction for that surrogate — but $r=1$ can badly misalign (negative cosine in 13 of 15 checkpoints), which explains why moderate depths matter.","JFB-$r$ needs no equilibrium-adjoint linear solve: it is entirely $r$ metric-preconditioned backward-Euler iterations toward the adjoint system $H^\\top w = g$, with geometric bias decay in depth."],"supporting_citations":[{"why":"Defines the JFB direction that the paper extends from contractive fixed-point maps to the non-contractive PDHG map.","marker":"[65]"},{"why":"Supplies the monotone-operator facts (sum theorem, Minty theorem, firm nonexpansiveness, Krasnosel'skii–Mann) used to prove the resolvent and directional-stability results.","marker":"[7]"},{"why":"Provides the averaged/nonexpansive characterization of PDHG that motivates replacing contraction assumptions with the metric-resolvent analysis.","marker":"[57]"},{"why":"Gives the PDHG algorithm and the step-size condition that Assumption 2.2 verifies in closed form as $\\tau\\sigma < 1$.","marker":"[13]"},{"why":"Supplies the G-prox preconditioned PDHG with the exact Schur preconditioner $W = KK^\\top$ that the forward solver applies via DCT and a small eigendecomposition.","marker":"[40]"},{"why":"Establishes the primal-dual treatment of time-dependent MFGs that the space-time discretization follows.","marker":"[11]"},{"why":"Formulates the inverse MFG problem of recovering the cost from density observations, the objective (2.9) the paper trains.","marker":"[18]"},{"why":"Provides the local regularity and strong-convexity assumptions that Lemma 3.5 adapts for the forward objective's curvature.","marker":"[68]"}],"fun_headline_variants":["Truncated differentiation matches full backprop at quarter memory","25 tracked steps equal full unrolling, half runtime","JFB-10: 3.5x speedup, one-ninth memory with comparable accuracy","Truncated gradients converge to true ones as depth grows","Inverse MFGs: truncate solver trajectory, keep gradient quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of large-depth consistency assumes three things the implementation only approximates: the proximal solves inside PDHG are exact (the code uses 12 damped Newton iterations), the detachment point is exactly the equilibrium (the measured residual has median $2\\times 10^{-7}$), and the set of binding flux constraints stays fixed during differentiation (in practice active sets shift in bursts of up to 48 slots during training). The authors state plainly that the local consistency theorem therefore does not apply directly to their experiments.","fun_headline_variants_meta":{"raw":{"variants":["Truncated differentiation matches full backprop at quarter memory","25 tracked steps equal full unrolling, half runtime","JFB-10: 3.5x speedup, one-ninth memory with comparable accuracy","Truncated gradients converge to true ones as depth grows","Inverse MFGs: truncate solver trajectory, keep gradient quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1481,"prompt_tokens":967,"completion_tokens":514,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":423}},"tokens_in":583,"tokens_out":514,"duration_ms":6099,"temperature":1.0,"reasoning_tokens":423,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:13:11.976113+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the relative bias $\\lVert d^{\\mathrm{JFB}}_r - d^{\\mathrm{ID}} \\rVert / \\lVert d^{\\mathrm{ID}} \\rVert$ against a cold-solved implicit reference on a problem with degenerate flux slots and persistent active-set changes: if increasing $r$ from 10 to hundreds does not reduce the bias toward zero, or if cosine alignment with the implicit gradient stays below about 0.5 at large $r$, Theorem 4.5's consistency is contradicted where it was meant to apply. A cheaper check: on a task where the warm start is far from equilibrium (large detach residual), increasing $r$ should still improve recovery if the practical claim holds; failure would show the near-equilibrium detachment is indispensable.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the G-prox preconditioned PDHG with the exact Schur preconditioner $W = KK^\\top$ that the forward solver applies via DCT and a small eigendecomposition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the primal-dual treatment of time-dependent MFGs that the space-time discretization follows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the local regularity and strong-convexity assumptions that Lemma 3.5 adapts for the forward objective's curvature."}],"review_version":2}