{"id":"ab047443-be34-4384-95b2-995ffdbad8bc","arxiv_id":"2502.01889","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A biased sparsity penalty in the neural OT minimax objective yields displacement-sparse maps, with an adaptive penalty schedule that outperforms fixed strengths on synthetic and real perturbation data.","lead":"The paper adds a sparsity penalty to the neural optimal transport objective so the learned map moves points along only a few coordinates, and it introduces a simulated-annealing schedule to set the penalty strength automatically. If the method holds up, drug perturbation studies could get interpretable low-dimensional transport maps without manual tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported displacement 'dimension' uses a 10^-2 cutoff, but the smoothed-l0 penalty is smooth at zero and yields small nonzero displacements, so sparsity may be a threshold artifact.","rationale":"The reader's weakest assumption focuses on the gap between theoretical guarantees (proven only for convex penalties) and the smoothed-l0 penalty used in experiments; that gap is real and is acknowledged by the authors in Appendix D.4. I agree that this limits the theoretical backing. However, the central claim in the paper is empirical: the method produces displacement-sparse maps with low dimensionality. The most load-bearing threat to that claim is that the measured quantity 'dimensionality' is a threshold-dependent summary of a smoothly regularized displacement field. Because the smoothed-l0 penalty is differentiable at zero and has zero gradient at zero, it does not produce exact zeros; the dimension count depends on an arbitrary 10^-2 cutoff. The 107-vs-2480 comparison could therefore reflect pushing many small displacements just below the cutoff rather than genuine support sparsity. This is testable by recomputing with different thresholds and by checking for exact zeros. The adaptive-lambda result is also partially by construction, since the high-dimensional scheduler is given the ground-truth dimension l=100 as input (Algorithm 4), but that is a lesser issue: the constant-lambda baseline already reaches 120. The theory gaps do not invalidate the empirical method, and the threshold concern is not yet demonstrated to be fatal; it just needs a concrete robustness check. Hence the CONDITIONAL verdict stands, with the added condition that the sparsity metric be shown threshold-robust (or that exact sparsity be verified).","tokens_in":23159,"tokens_out":8580,"duration_ms":83258,"concrete_test":"Vary the displacement threshold across {10^-3, 3e-3, 10^-2, 3e-2, 10^-1} and recompute dim(Delta(x)) for unregularized ICNN OT, constant lambda = 0.005, and adaptive-lambda on the synthetic k=100 dataset. Also report the fraction of coordinates on inactive genes with displacement strictly equal to 0 versus merely below the threshold. If the ordering or reported values (107 vs 2480) change materially, or if no coordinates are exactly zero, the sparsity claim is a threshold artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Table 1, Fig. 5) is that adaptive-lambda ICNN OT achieves displacement dimension 107 vs 2480 for unregularized ICNN OT. However, 'dimension' is defined in Appendix D.3 as counting coordinates with |Delta_i| > 10^-2 via np.isclose(.,0). The smoothed-l0 penalty ||z||_{0,xi} = sum_i (1 - exp(-z_i^2/(2 xi^2))) (section 3.3) is smooth at zero and has zero gradient at zero; with xi = 1.0 (Appendix E), it behaves like a weak L2 penalty for small displacements and never forces exact zeros. Thus a map can have all 3000 coordinates active at small amplitude (say 5e-3) yet be counted as 'dimension 0' under the 10^-2 threshold. The unregularized ICNN baseline may place many coordinates just above threshold and the regularizer just below, producing a large apparent sparsity gain without true support sparsity. This threatens the headline interpretability claim, since the Cuturi et al. baseline produces genuinely zero displacements via proximal operators, while the proposed method does not. The adaptive-lambda advantage (107 vs 120 for constant lambda = 0.005) may also reflect threshold sensitivity rather than a true low-dimensional structure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a regularized minimax formulation for neural optimal transport in which the standard ICNN dual objective is augmented by a penalty λ∫τ(∇g(y)−y)dQ on the displacement vector. This is intended to generalize the elastic-cost displacement-sparse OT of Cuturi et al. to neural solvers and to allow non-convex penalties such as a smoothed ℓ0 norm. The paper proves consistency as λ→0, convergence of regularized solutions to the unregularized solution set, monotonicity of sparsity for convex τ, and a map-error bound depending on optimality gaps and λ. It also introduces a simulated-annealing heuristic to adapt λ, with separate strategies for low- and high-dimensional tasks. Experiments on a synthetic scRNA-like task (d=3000, k=100) report displacement dimension 107 with adaptive λ versus 2480 for unregularized ICNN OT, and over 95% overlap with ground-truth genes; additional experiments on real 4i perturbation data compare the smoothed ℓ0 penalty to ℓ1 and to Cuturi et al.","tokens_in":23500,"tokens_out":5724,"duration_ms":61443,"significance":"If the results hold, the paper addresses a real limitation of neural OT solvers: learned maps can be high-dimensional and difficult to interpret, while exact OT approaches to displacement sparsity require proximal operators that exclude non-convex penalties. The proposed formulation is intuitive and the convex-penalty theory is a useful first step. The synthetic setup with known ground-truth dimension k and gene-overlap evaluation is a good experimental design, and the adaptive-λ scheme is practically relevant for users who need a target displacement dimension. The paper is generally clearly written and does a service by explicitly flagging in Appendix D.4 that its main theoretical guarantees only cover convex penalties and do not cover the smoothed ℓ0 penalty used in the experiments.","major_comments":[{"comment":"The reported 'dimension' of the displacement is defined via np.isclose(.,0) with a threshold of 10^-2 (Appendix D.3), while the smoothed ℓ0 penalty used in the headline experiments is smooth at zero (σ=1.0, Appendix E) and therefore does not produce exact zeros. Consequently, the central comparison '107 vs 2480' may reflect that unregularized ICNN displacements lie just above the threshold and regularized ones just below it, rather than genuine support sparsity. Please report the full distribution of |Δ_i|, perform a threshold-sensitivity analysis (e.g., 10^-3, 10^-2, 5×10^-2), and report how many coordinates are exactly zero, so the interpretability claim is not an artifact of the cutoff.","section":"Appendix D.3, Table 1"},{"comment":"Theorems 4.4 and 4.5 are proved only for convex regularizers, and Appendix D.4 explicitly states that the smoothed ℓ0-norm 'does not benefit from the theoretical guarantee.' Since the main experiments use smoothed ℓ0, the theory in Section 4 does not actually cover the method that carries the empirical claims. Either extend the theory to pseudoconvex penalties (or provide an approximation argument), or clearly restate the contribution as a heuristic method whose theory is limited to the convex case.","section":"Section 4 and Appendix D.4"},{"comment":"In the proof of Theorem 4.3, after equation (13) the argument writes limsup_{λ→0} J_λ ≤ V_{P,Q}(f_λ,g_0) and then substitutes f=f_λ into inequality (11). But f_λ depends on λ, so passing the limit into V_{P,Q}(f_λ,g_0) requires extracting a convergent subsequence (for example by Arzelà-Ascoli under Assumption 3.1) and using continuity of V_{P,Q}. As written this is a gap, and it affects the consistency claim for non-convex τ as well as the limiting argument generally.","section":"Appendix B, proof of Theorem 4.3"},{"comment":"In the high-dimensional mode, Algorithm 4 directly increases λ until the measured dimension falls below the constraint l, then decreases λ while keeping the dimension below l. Thus the claim that 'adaptive-λ achieves a displacement dimensionality closer to the ground truth' (Section 5.1, Table 1) is partly by construction, because the algorithm is optimizing against that very dimension. This is not by itself a flaw, but the evaluation should be reframed as a feasibility-under-constraint result, and the constant-λ baseline should be compared after the same thresholding procedure is applied.","section":"Section 4.1, Algorithm 4, and Section 5.1"}],"minor_comments":[{"comment":"There is a corrupted block of text containing '/uni00000013/...' tokens in the paragraph after the experimental setup; this should be removed or replaced with the intended figure caption or description.","section":"Section 5.1"},{"comment":"The header 'Cuturi et al. 23Oursλ' is missing a space and should read 'Cuturi et al. [23]' and 'Ours (λ)'.","section":"Table 1"},{"comment":"The paper calls the smoothed ℓ0-norm 'non-convex' in Section 3.3 but 'pseudoconvex' in Appendix D.4; please clarify the terminology and be consistent about which property is being invoked.","section":"Section 3.3 vs Appendix D.4"},{"comment":"The proof of Lemma B.2 is labeled 'Lemma A.2' in the text, and the proof of Theorem 4.3 refers to 'Theorem A.1' when the stated minimax theorem is in fact Theorem B.1; internal cross-references should be corrected.","section":"Appendix B"},{"comment":"In the proof of Theorem 4.6, the notation 's(f*)' and 'λs(g′)' is confusing; the penalty function s is applied to dual potentials g, so the expressions should be written consistently as s(g*) and λs(g′).","section":"Appendix B, proof of Theorem 4.6"},{"comment":"The caption says 'The average displacement dimensionality is shown' but does not define the dimension metric; please specify in the caption that dimensionality is computed with the 10^-2 threshold described in Appendix D.3.","section":"Figure 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for cs.LG and the core idea is promising, but the headline empirical claim is currently vulnerable to the threshold artifact: with a smooth-at-zero penalty and a 10^-2 cutoff, 'sparsity' may be a measurement artifact. This is the main risk to the paper's central claim. The theoretical gap in Theorem 4.3 and the explicit exclusion of the non-convex penalty from Theorems 4.4-4.5 are fixable but should be addressed head-on. If the authors can supply threshold-robust evidence of true support sparsity (or acknowledge that the method produces thresholded sparsity only), the paper is a reasonable major-revision candidate; without such evidence, the contribution is substantially weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chen, Xie, Zhang propose a biased minimax formulation for ICNN-based OT that lets you add sparsity penalties directly to the dual objective, plus a simulated-annealing controller for λ. That is a real step beyond Cuturi et al. 2023, which requires proximal operators and exact Sinkhorn, and prior general-cost neural OT work that needed convexity. The smoothed ℓ0 regularizer is a sensible choice, and the gene-overlap accuracy in Table 3 is independent evidence that the selected displacements point at the right coordinates. The adaptive λ framework is practically motivated and the multiple-run trajectories in D.1 give it some credibility.\n\nThe soft spots, in proportion: the headline dimension claim is fragile. Dimension is defined, in D.3, by counting coordinates with |Δ_i| > 10^-2 via np.isclose. The smoothed ℓ0 norm has zero gradient at zero and never forces exact zeros, so a coordinate with a 5e-3 displacement is scored as 'inactive.' That means the 107-vs-2480 comparison may substantially overstate the gain in true support sparsity. The gene-overlap results soften this—if small displacements sit on the true genes, the method is still doing something right—but they don't rescue the absolute dimensionality numbers, and the comparison to Cuturi et al. (which produces genuine zeros) is not apples-to-apples.\n\nThe theory is honest but narrow. Theorems 4.4 and 4.5 are proven only for convex τ, and D.4 explicitly says the smoothed ℓ0 is pseudoconvex and excluded. That is an unusual and welcome disclosure, but it means the convergence and monotonicity results don't cover the actual regularizer used in the experiments. The proof of Theorem 4.3 has a genuine gap: the limsup is passed into V_P,Q(f_λ, g_0) without extracting a subsequence along which f_λ converges. That can likely be fixed, but as written it is not a complete proof. Theorem 4.6's bound is standard, conditional on optimality gaps.\n\nNo code is provided, which is a real limitation for a method paper whose practical appeal is a heuristic controller. The 4i results lack error bars in Figure 6, though the text says ten runs were averaged.\n\nWho this is for: people doing neural OT for single-cell perturbation who care about interpretable maps. It is not a reshaping of OT theory, but it is a plausible, useful methodological contribution. I'd send it to review—the threshold artifact and the proof gap are fixable in revision, and the core idea deserves referee time. I would not cite the dimension numbers without checking what they mean; I might cite the formulation.","headline":"A genuinely new biased-minimax formulation for displacement-sparse neural OT with a clever adaptive λ schedule, but the headline sparsity numbers may be partly a 10^-2 threshold artifact and the theory does not cover the actual penalty used.","tokens_in":23985,"tokens_out":2133,"would_cite":true,"duration_ms":19371,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","68T07","90C08"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a smoothed ℓ0 penalty directly to the dual-potential objective of a neural OT solver produces transport maps whose displacement vector has far fewer active coordinates, while keeping the transported measure close to the target.","keywords":["displacement-sparse optimal transport","neural optimal transport","input convex neural network","smoothed ℓ0 norm","adaptive regularization","simulated annealing","single-cell perturbation","optimal transport interpretability"],"falsifier":"Run the synthetic \\(k=100\\) benchmark with the truly perturbed gene set shuffled after each restart; if the method still reports low displacement dimension but gene-overlap accuracy drops to chance, the claim that it recovers sparse structure rather than merely shrinking displacements is falsified. A second check: at a fixed high \\(\\$\\lambda$\\), measure whether the displacement dimension can be pushed below the true \\(k\\) without the Wasserstein divergence exploding—the theory predicts error growing linearly in \\(\\$\\lambda$\\), so a dramatic superlinear blow-up would contradict it.","tokens_in":22970,"feed_emoji":"🧬","tokens_out":8377,"duration_ms":76483,"temperature":0.7,"pith_summary":"Optimal transport moves one distribution onto another; neural solvers scale to high-dimensional biology but learn dense maps that move every gene or feature, which is hard to interpret. This paper claims that sparsity can be imposed on these maps by adding a penalty \\(\\$\\lambda$ \\int \\tau(\\nabla g(y)-y)\\,dQ\\) directly to the ICNN minimax objective, so the displacement vector \\(\\nabla g(y)-y\\) is encouraged to have few nonzero entries. The paper introduces a smoothed \\(\\ell_0\\) regularizer and an adaptive simulated-annealing schedule for the penalty strength \\(\\$\\lambda$\\). On a synthetic drug-perturbation benchmark with 100 truly affected genes out of 3000, the adaptive method recovers average displacement dimension 107 instead of 2480 for unregularized ICNN OT, while keeping gene-overlap accuracy above 95%. This matters because interpretable, low-dimensional transport maps could make single-cell perturbation analysis more tractable without abandoning neural solvers.","feed_headline":"Sparsity penalty slims neural OT maps from 2480 to 107 dimensions","feed_subtitle":"A smoothed ℓ0 regularizer plus adaptive λ keeps gene overlap above 95 percent on drug-perturbation tasks.","key_machinery":"The load-bearing object is the dual potential \\(g\\) from the ICNN minimax formulation, whose gradient \\(\\nabla g(y)\\) recovers the transport map; the displacement vector is \\(\\nabla g(y)-y\\). The paper's mechanism is to penalize that displacement vector inside the objective (Eq. 7) rather than through an elastic-cost proximal operator, which is what allows arbitrary penalties—including non-convex ones—to be used. The concrete regularizer is the smoothed \\(\\ell_0\\) norm \\(\\|z\\|_{0,\\xi}=\\sum_i (1-\\exp(-$z_i^{2}$/(2\\$xi^{2}$)))\\), a smooth surrogate for counting nonzero coordinates. The adaptive component is a simulated-annealing controller that adjusts \\(\\$\\lambda$\\) according to an evaluation function balancing sparsity and feasibility in low dimension, or according to a direct dimensionality constraint \\(l\\) in high dimension.","core_discovery":"The central claim is that displacement sparsity can be imposed on neural OT maps by regularizing the dual objective rather than the cost: adding \\(\\$\\lambda$ \\int_{\\mathbb{R}^d} \\tau(\\nabla g(y)-y)\\,dQ\\) to the minimax objective (7) and annealing \\(\\$\\lambda$\\) during training yields maps whose displacement vector activates only a few coordinates while still transporting source to target. The paper further claims that a smoothed \\(\\ell_0\\) penalty beats the \\(\\ell_1\\)-based alternative in this neural setting, and that adaptive control of \\(\\$\\lambda$\\) is necessary to avoid early suboptimal convergence in large-scale training. Theoretically, it proves that the biased objective returns to the unbiased one as \\(\\$\\lambda$ \\to 0\\), that convex penalties produce sparser maps as \\(\\$\\lambda$\\) grows, and that the map error is bounded by the optimality gaps plus \\(2\\$\\lambda$ M_\\tau\\), so the main cost of sparsity is controlled by \\(\\$\\lambda$\\). Empirically, on the synthesized scRNA task with ground-truth dimension \\(k=100\\), the adaptive method reaches dimension 107 versus 2480 for unregularized ICNN OT and 120 for the best constant \\(\\$\\lambda$\\), with gene-overlap accuracy above 95%; on four real 4i drug datasets it reduces displacement dimensionality while keeping Wasserstein divergence comparable.","pith_inferences":["The same penalty-in-the-dual trick should extend to other structural priors on the transport map, such as group sparsity over gene modules or monotonicity constraints, since it only requires evaluating a term involving \\(\\nabla g\\).","The adaptive-\\(\\lambda\\) schedule suggests a general curriculum for regularized minimax training beyond optimal transport: anneal regularization intensity to avoid premature convergence to trivial equilibria.","Because the theory covers only convex penalties, a natural testable extension is to prove convergence for pseudoconvex penalties like smoothed \\(\\ell_0\\), or to replace it with a convex surrogate that still approximates the \\(\\ell_0\\) count."],"forward_implications":["Neural OT solvers can now produce displacement-sparse maps with arbitrary sparsity penalties, not only the convex penalties that admit proximal operators.","The smoothed \\(\\ell_0\\) regularizer gives a direct handle on the number of active displacement coordinates, which is valuable when perturbed and noise genes are not separable by magnitude.","Adaptive \\(\\lambda\\) via simulated annealing removes manual tuning in high-dimensional training and reaches displacement dimensions closer to known ground truth than any fixed \\(\\lambda\\).","The map-error bound implies that once the ICNN training gap is small, users can trade a controlled amount of map accuracy for interpretability by choosing \\(\\lambda\\)."],"supporting_citations":[{"why":"Supplies the minimax dual-potential formulation and the map-recovery identity \\(\\nabla g\\); Eq. (7) is this objective plus the new penalty.","marker":"[33]"},{"why":"Introduces displacement-sparse OT maps via elastic costs in exact solvers; the paper generalizes this to neural solvers and treats it as the exact baseline.","marker":"[18]"},{"why":"Defines the smoothed \\(\\ell_0\\) norm used as the paper's main sparsity regularizer.","marker":"[34]"},{"why":"Provides the drug-perturbation application, the real 4i dataset, and the ICNN-OT baseline the method is designed to improve.","marker":"[12]"},{"why":"Defines the Input Convex Neural Network architecture used to parameterize the convex dual potentials.","marker":"[2]"},{"why":"Supplies the sliced Wasserstein approximation used to accelerate the feasibility metric in the adaptive evaluation function.","marker":"[8]"}],"fun_headline_variants":["Sparse neural OT: smoothed ℓ0 regularizer cuts map size 23-fold","Adaptive λ in neural OT reduces displacement space to 107 dims","Smoothed ℓ0 beats ℓ1 for displacement-sparse neural OT","Neural OT maps shrinkage: adaptive sparsity with smoothed ℓ0","Displacement-sparse neural OT via adaptive smoothed ℓ0"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proofs assume the neural-network training comes very close to solving the optimal-transport problem before the sparsity penalty matters, and the main experiments use a smoothed-\\(\\ell_0\\) penalty for which the convergence and sparsity theorems are not proved because it is not convex.","fun_headline_variants_meta":{"raw":{"variants":["Sparse neural OT: smoothed ℓ0 regularizer cuts map size 23-fold","Adaptive λ in neural OT reduces displacement space to 107 dims","Smoothed ℓ0 beats ℓ1 for displacement-sparse neural OT","Neural OT maps shrinkage: adaptive sparsity with smoothed ℓ0","Displacement-sparse neural OT via adaptive smoothed ℓ0"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000682,"raw_usage":{"total_tokens":3148,"prompt_tokens":1045,"completion_tokens":2103,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":2004}},"tokens_in":661,"tokens_out":2103,"duration_ms":14693,"temperature":1.0,"reasoning_tokens":2004,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:06:13.423372+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the synthetic \\(k=100\\) benchmark with the truly perturbed gene set shuffled after each restart; if the method still reports low displacement dimension but gene-overlap accuracy drops to chance, the claim that it recovers sparse structure rather than merely shrinking displacements is falsified. A second check: at a fixed high \\(\\$\\lambda$\\), measure whether the displacement dimension can be pushed below the true \\(k\\) without the Wasserstein divergence exploding—the theory predicts error growing linearly in \\(\\$\\lambda$\\), so a dramatic superlinear blow-up would contradict it.","supporting_citations":[{"cited_title":"Optimal transport mapping via input convex neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the minimax dual-potential formulation and the map-recovery identity \\(\\nabla g\\); Eq. (7) is this objective plus the new penalty."},{"cited_title":"Monge, bregman and occam: Interpretable optimal transport in high-dimensions with feature-sparse maps","cited_arxiv_id":null,"evidence_quote":"Introduces displacement-sparse OT maps via elastic costs in exact solvers; the paper generalizes this to neural solvers and treats it as the exact baseline."},{"cited_title":"Hosein Mohimani, Massoud Babaie-Zadeh, and Christian Jutten","cited_arxiv_id":null,"evidence_quote":"Defines the smoothed \\(\\ell_0\\) norm used as the paper's main sparsity regularizer."},{"cited_title":"Zico Kolter","cited_arxiv_id":null,"evidence_quote":"Defines the Input Convex Neural Network architecture used to parameterize the convex dual potentials."},{"cited_title":"Sliced and radon wasser- stein barycenters of measures","cited_arxiv_id":null,"evidence_quote":"Supplies the sliced Wasserstein approximation used to accelerate the feasibility metric in the adaptive evaluation function."}],"review_version":1}