{"id":"12e99553-f14a-4be4-a8bb-123db28411c2","arxiv_id":"2506.03999","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Large deviation principles for entropic optimal transport are generalized to Schrödinger bridges with uniformly convergent cost functions, covering reflected Brownian reference processes on bounded convex domains.","lead":"This mathematics paper proves a large deviation principle for sequences of Schrödinger bridge problems with changing reference measures, extending prior work to reflected Brownian motion on bounded convex domains. The result gives a rate of convergence to optimal transport plans, which matters for generative models that must respect boundaries.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Proposition 5.7 selects ε′ per pair (x,y), so the claimed uniform lower bound for the reflected heat kernel is not established; Theorem 5.2 depends on this uniformity.","rationale":"The paper's central claim is the large deviation principle for scaled Schrödinger bridges with reflected Brownian reference processes. The decisive input for the reflected case is Theorem 5.2, whose proof relies on the uniform lower bound in Proposition 5.7. I checked the proof of Proposition 5.7 carefully. The earlier parts of the proof are internally consistent: the choice of the bridge point z satisfies |z−y|≤ε/3, which makes the replacement of |z−y|+ε/3 by ε valid; the projection step uses the nonexpansiveness of projections onto the convex set D_{−ε′}, so |x_{ε′}−y_{ε′}|≤|x−y|; and the stochastic estimates are uniform in the relevant parameters. The single genuine soft spot is the final step's parameter ε′: the text chooses ε′ for each pair (x,y) via the condition dist(x,D_{−ε′})≤ε, and then absorbs vol(B(0,ε′/2)) into a constant. Without a statement that one ε′ works simultaneously for all x∈D, the constants in (5.25) can depend on (x,y), invalidating the uniformity claimed in (5.11). This geometric premise is true for bounded convex domains, but it is not proved or even stated. The reader's weakest assumption identifies exactly this issue, and I agree with that assessment. Because the gap is real but readily fixable, the conditional verdict remains appropriate, so no change to the reader's recommendation is needed.","tokens_in":25045,"tokens_out":21075,"duration_ms":200477,"concrete_test":"Prove the missing geometric lemma: for any bounded convex open D⊂R^d and any ε>0, there exists ε′>0 such that D ⊆ D_{−ε′} + B(0,ε), equivalently sup_{x∈D} dist(x,D_{−ε′})≤ε, where D_{−ε′}={x∈D: dist(x,∂D)>ε′}. One route is a compactness contradiction: if no such ε′ exists, choose x_n with dist(x_n,D_{−1/n})>ε; pass to a subsequence x_n→x∈cl(D); if x∈D the inequality fails for large n because x∈D_{−1/n}, and if x∈∂D the distance from x_n to D_{−1/n} tends to 0 for convex D, again a contradiction. After establishing the lemma, revisit the last paragraph of Proposition 5.7 with a fixed ε′; the constants in display (5.25) become independent of (x,y), and (5.11) follows. If the lemma fails for some convex D, exhibit that D and ε; then Theorem 5.2 would need a different proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 5.7 asserts a lower bound for p^r_η(1,x,y) that is uniform in x,y∈D, with constants α_D, β_D, η0 independent of (x,y) (display 5.11). The proof's final Chapman–Kolmogorov step (around display 5.24) introduces ε′≤ε 'chosen so that dist(x,D_{−ε′})≤ε' and then passes through the ball B(z,ε′/2). The resulting constants α̃, β̃ absorb the volume factor vol(B(0,ε′/2)) and the parameter ε′. If ε′ is chosen individually for each pair (x,y), then those constants inherit the (x,y)-dependence and the asserted uniformity does not follow. What the proof requires is a single ε′>0, depending only on ε and D, such that sup_{x∈D} dist(x,D_{−ε′})≤ε. Such an ε′ exists for bounded convex D because the inner parallel sets D_{−ε′} exhaust D as ε′↓0, but this geometric fact is neither stated nor proved in the manuscript. Since Theorem 5.2 uses precisely this uniform lower bound to obtain uniform convergence of c_η(x,y)=−η log p^r_η(1,x,y), the gap is load-bearing for the reflected-Brownian application of the LDP. The gap is fixable by choosing ε′ globally before the final step and recording the exhaustion lemma; as written, however, the uniformity claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves a large deviation principle (LDP) for sequences of static Schrödinger bridges / entropic optimal transport plans when the cost functions c_η converge uniformly to a limit c as the noise scale η↓0 (Theorem 4.2), generalizing the fixed-cost LDP of Bernton–Ghosal–Nutz [2]. It then shows that for reflected Brownian motion on a bounded convex domain, the associated cost c_η(x,y) = −η log p^r_η(1,x,y) converges uniformly to |x−y|^2/2 (Theorem 5.2), so the LDP applies to reflected Schrödinger bridges. The upper-bound half of Theorem 4.2 is proved in detail via cyclical-invariance arguments; the lower-bound half is deferred to [2]. Theorem 5.2 relies on heat-kernel upper and lower bounds, the latter proved via Skorokhod-map and Brownian-bridge estimates.","tokens_in":25312,"tokens_out":8202,"duration_ms":73748,"significance":"If the proof gaps are fixed, this would be a meaningful contribution: it addresses an open problem from [2], extends LDPs for entropic optimal transport to non-fixed cost functions, and would provide the first LDP for reflected Schrödinger bridges. The general theorem (Theorem 4.2) is clean, the rate function is explicitly derived from the limit cost with no fitted parameters, and the analysis of reflected heat kernels in Section 5 is nontrivial. The manuscript is transparent about which results are deferred to [2] and which analytic bounds are imported from [8] and [23]. The main claims are clearly stated and the overall strategy is sound, but the two gaps described below are load-bearing.","major_comments":[{"comment":"The lower-bound half of Theorem 4.2 is not proved. The text states that Proposition 4.10 and Corollary 4.11 are \"identical to [2]\" and omits their proofs. This cannot be literally correct: Corollary 4.11 concerns the sequence π_η with varying costs c_η, whereas the corresponding result in [2] is proved for a fixed cost c. Adapting the proof requires controlling the difference c_η − c in the cyclical-invariance identity (2.10)/(3.5) and in the density estimates, exactly as done for the upper bound in Lemmas 4.6 and 4.8. Since the lower bound is essential to the LDP of Theorem 4.2, the authors should provide the proof or a detailed adaptation rather than a citation.","section":"Section 4, Corollary 4.11"},{"comment":"The claimed uniform lower bound (5.11) is not established. In the final Chapman–Kolmogorov step (display 5.24), the parameter ε′ is \"chosen so that dist(x, D_{−ε′}) ≤ ε\" for each pair (x,y), and the constants ᾱ, β̃ absorb the volume factor vol(B(0, ε′/2)) and terms containing ε′. If ε′ varies with (x,y), the constants inherit this dependence, so the asserted uniformity in x,y does not follow. The proof also uses |y − y_{ε′}| ≤ ε (display 5.25) without stating that ε′ is chosen for both coordinates. The fix is to select a single ε′ > 0, depending only on ε and D, such that sup_{x∈D} dist(x, D_{−ε′}) ≤ ε (e.g., ε′ = ε/2 for bounded convex D), and to prove the required exhaustion property of the inner parallel sets D_{−ε′}. Without this, Theorem 5.2 is unsupported.","section":"Section 5, Proposition 5.7"}],"minor_comments":[{"comment":"In the display (4.4), the notation \"(xi,yi)^k_{2=1}⊆Γ\" is a typo; it should be \"(x_i,y_i)^k_{i=2}⊆Γ\".","section":"Section 4, Theorem 4.2"},{"comment":"There is a typo \"TThis establishes\" in the proof, and the final sentence should read \"lim inf_{η↓0} η log π^k_η(A) ≥ −δ′\" (the minus sign is missing in the sentence after (4.13)).","section":"Section 4, Lemma 4.6"},{"comment":"The text contains minor typos: \"Kullback-Liebler\" should be \"Kullback-Leibler\", and \"the the strict convexity\" has a duplicated article.","section":"Section 2.1"},{"comment":"The statement \"it is straightforward to show that c_η converges uniformly\" would benefit from explicitly naming the mode of convergence (uniform on D×D) before referencing Figure 1, since Figure 1 only demonstrates pointwise slices.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a genuine open problem and the overall strategy is sound, but the two gaps above are substantial. The fix for Proposition 5.7 is straightforward: add a global choice of ε′ and a short exhaustion lemma for inner parallel sets of bounded convex domains. The lower-bound proof in Section 4 should be included or clearly adapted; a bare citation to [2] for a result in a non-identical setting is not sufficient for the central LDP claim. If the authors supply these, the paper would be a solid contribution to the literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is Section 5: for reflected Brownian motion on a bounded convex domain, the cost c_eta(x,y) = -eta log p^r_eta(1,x,y) converges uniformly to |x-y|^2/2. That gives a large-deviation principle for reflected Schrödinger bridges, which is a real extension of Bernton–Ghosal–Nutz and speaks directly to their open question. The Section 4 extension to uniformly convergent costs is, as the authors themselves say, a straightforward modification of [2]; I appreciate that they do not oversell it.\n\nThe heat-kernel analysis is the meat of the paper. The upper bound comes from Davies; the lower bound goes through the Skorokhod map and Brownian-bridge estimates, and most of it is convincing. But Proposition 5.7 has a genuine gap. In the final Chapman–Kolmogorov step, epsilon' is chosen for each pair (x,y) so that dist(x,D_{-epsilon'}) <= epsilon. The constants alpha~ and beta~ absorb vol(B(0,epsilon'/2)), so if epsilon' varies with (x,y), the uniformity claimed in (5.11) does not follow. You need a single epsilon' depending only on epsilon and D, and the exhaustion property of the inner parallel sets D_{-epsilon'}. That property is true for bounded convex domains, but it is not stated or proved. This is load-bearing for Theorem 5.2, and a referee should insist on the missing lemma.\n\nTwo smaller things. The lower bound of the LDP is not proved in the text: Proposition 4.10 and Corollary 4.11 are quoted from [2]. That is acceptable if the journal tolerates reliance on a companion paper, but it does mean the paper is not self-contained at the main theorem level. Also, the last display of Lemma 4.6 has a sign typo: it should be lim inf >= -delta', not delta'.\n\nI think the main claims are correct. The gap is fixable and does not indicate a wrong approach. This paper deserves a serious referee: it advances the subfield, the exposition is honest, and the reflected-Brownian application is new. I would send it to review and ask for a revision that patches the uniformity lemma in 5.7 and clarifies exactly which parts of [2] are being imported.","headline":"A real but modest extension of the Bernton–Ghosal–Nutz LDP to reflected Schrödinger bridges, held up by one fixable gap in the uniform heat-kernel lower bound.","tokens_in":25888,"tokens_out":12632,"would_cite":true,"duration_ms":126179,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F10","49Q22","93C10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Schrödinger bridges with reflected Brownian reference satisfy a large deviation principle, quantifying their exponential convergence to the quadratic optimal transport plan.","keywords":["large deviations","Schrödinger bridges","entropic optimal transport","reflected diffusions","cyclical monotonicity","Neumann heat kernel","Skorokhod map","generative modeling"],"falsifier":"Using the explicit density (5.6) for $D=[0,1]$, compute $\\sup_{x,y\\in[0,1]} | -\\eta\\log p^r_\\eta(1,x,y) - |x-y|^2/2 |$ for decreasing $\\eta$; the theorem asserts this error goes to zero uniformly, so a persistent positive error near the boundary at any fixed $\\eta$ would falsify the uniform convergence that feeds the large deviation principle.","tokens_in":24797,"feed_emoji":"📈","tokens_out":10638,"duration_ms":103680,"temperature":0.7,"pith_summary":"The paper establishes a large deviation principle (LDP) for sequences of Schrödinger bridges / entropic optimal transport plans in the small-noise limit, when the costs vary with the noise level and converge uniformly to a fixed cost function. This extends the fixed-cost LDP of [2] and answers an open problem posed there. The main application is to reflected Brownian motion on bounded convex domains: the paper proves the associated cost functions converge uniformly to the quadratic cost, so reflected Schrödinger bridges satisfy the LDP. The result matters because it quantifies, at exponential scale, how fast entropic transport plans converge to the limiting optimal transport plan, a question directly relevant to diffusion Schrödinger bridge generative models on constrained domains.","feed_headline":"Large deviations proved for reflected Schrödinger bridges","feed_subtitle":"Entropic transport plans with reflected Brownian motion converge exponentially to quadratic optimal transport.","key_machinery":"The argument is carried by four pieces. (1) Cyclical invariance (2.10): each EOT minimizer $\\pi_\\eta$ is characterized by a density identity involving the cost differences $c_\\eta(x_i,y_i) - c_\\eta(x_i,y_{i+1})$, which is the exact bridge between the plan's mass and the rate function. (2) The rate function $I$ in (4.4), read as the maximal cost improvement obtainable by permuting partners in a cycle. (3) A slack-variable modification of the fixed-cost lemmas of [2] that absorbs the uniform error $\\|c_\\eta - c\\|_\\infty$, converting the fixed-cost LDP into the varying-cost version. (4) For the reflected application: the Neumann heat-kernel upper bound of [8] and a matching lower bound derived from the continuity of the Skorokhod map on the bounded convex domain, giving uniform control on $-\\eta \\log p^r_\\eta(1,x,y)$.","core_discovery":"On the paper's own terms, the central discovery is Theorem 4.2: under Assumptions 2.1 and 4.1, if the cost functions $c_\\eta$ converge uniformly to $c$ as $\\eta \\downarrow 0$, then the entropic optimal transport plans $\\pi_\\eta$ satisfy a weak-type LDP with rate function $I(x,y) = \\sup_{k\\ge 2} \\sup_{(x_i,y_i)\\subseteq \\Gamma} \\sup_{\\sigma\\in\\Sigma(k)} \\sum_{i=1}^k [c(x_i,y_i) - c(x_i,y_{\\sigma(i)})]$, where $\\Gamma$ is the support of the limiting plan. The rate function can also be written as $c - (-\\psi \\oplus \\psi^c)$ on the support for any Kantorovich potential $\\psi$. Theorem 5.2 shows that for reflected Brownian motion on an open bounded convex domain, the costs $c_\\eta(x,y) = -\\eta \\log p^r_\\eta(1,x,y)$ converge uniformly to $c(x,y) = |x-y|^2/2$, so the LDP applies to reflected Schrödinger bridges. Under compact supports the weak-type bound upgrades to a full LDP (Corollary 4.3).","pith_inferences":["The same theorem should apply to other reference dynamics whose transition densities admit matching Gaussian-type upper and lower bounds, e.g., reflected diffusions with bounded drift on bounded convex domains; verifying uniform convergence of $-\\eta \\log q_\\eta(1,x,y)$ would be the only missing step.","The rate function $I$ admits a sample-based interpretation in generative modeling: for a learned transport plan, evaluating the cycle-cost sums against the OT plan gives a quantitative measure of how far the model is from being optimal, which could serve as a diagnostic for diffusion Schrödinger bridge training.","A path-space LDP for the dynamic reflected Schrödinger bridge remains open; because the static plans satisfy the LDP and the bridge interpolation is the reference process's own bridges, a path-space statement would need a control on the full bridge laws, and the Gaussian-bridge argument used for Brownian motion in [19] will not carry over to reflected Brownian motion."],"forward_implications":["Reflected Schrödinger bridge plans on any open bounded convex domain converge to the quadratic optimal transport plan at an exponential rate governed by the rate function $I$.","The LDP is full (not merely weak-type) whenever the supports of the two marginals are compact, giving the usual large-deviation upper and lower bounds on all open and closed sets.","The rate function is computable from the limit cost $c$ and the Kantorovich potential $\\psi$ as $c - (-\\psi \\oplus \\psi^c)$, so the exponential asymptotics are explicit for the quadratic-cost case.","The uniform-convergence hypothesis transfers the entire large-deviation machinery of [2] from one fixed cost to any family of costs that stabilize uniformly, which is the partial answer to the open problem posed there."],"supporting_citations":[{"why":"Provides the fixed-cost LDP and the cyclical-monotonicity framework that Theorem 4.2 extends to uniformly convergent cost sequences; the paper modifies its Lemmas 3.1 and 4.1 with a slack parameter.","marker":"[2]"},{"why":"Supplies well-posedness and 1/2-Hölder continuity of the Skorokhod map for the bounded convex domain, used in Lemma 5.5 and Proposition 5.7 for the heat-kernel lower bound.","marker":"[23]"},{"why":"Gives the Neumann heat kernel upper bound (Theorem 3.2.9) used to prove the matching upper estimate in the cost convergence proof for Theorem 5.2.","marker":"[8]"},{"why":"Provides the Brownian bridge computations and modulus-of-continuity estimates used in Proposition 5.7 to control the probability that an unreflected bridge stays inside the required cylinder.","marker":"[18]"}],"fun_headline_variants":["Reflected Schrödinger bridges satisfy large deviations","LDP for reflected Schrödinger bridges in small-noise limit","Reflected bridges: large deviation principle extended","Entropic transport LDP extends to reflected bridges","Small-noise limits for reflected Schrödinger bridges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single shrink of the domain by one common margin lets every pair of start and end points stay within that margin of the shrunk set, which is what makes the reflected heat-kernel lower bound uniform across all pairs.","fun_headline_variants_meta":{"raw":{"variants":["Reflected Schrödinger bridges satisfy large deviations","LDP for reflected Schrödinger bridges in small-noise limit","Reflected bridges: large deviation principle extended","Entropic transport LDP extends to reflected bridges","Small-noise limits for reflected Schrödinger bridges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001195,"raw_usage":{"total_tokens":4978,"prompt_tokens":1046,"completion_tokens":3932,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":3862}},"tokens_in":662,"tokens_out":3932,"duration_ms":27752,"temperature":1.0,"reasoning_tokens":3862,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:51:19.410912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Using the explicit density (5.6) for $D=[0,1]$, compute $\\sup_{x,y\\in[0,1]} | -\\eta\\log p^r_\\eta(1,x,y) - |x-y|^2/2 |$ for decreasing $\\eta$; the theorem asserts this error goes to zero uniformly, so a persistent positive error near the boundary at any fixed $\\eta$ would falsify the uniform convergence that feeds the large deviation principle.","supporting_citations":[{"cited_title":"Bernton, P","cited_arxiv_id":null,"evidence_quote":"Provides the fixed-cost LDP and the cyclical-monotonicity framework that Theorem 4.2 extends to uniformly convergent cost sequences; the paper modifies its Lemmas 3.1 and 4.1 with a slack parameter."},{"cited_title":"Lions and A.-S","cited_arxiv_id":null,"evidence_quote":"Supplies well-posedness and 1/2-Hölder continuity of the Skorokhod map for the bounded convex domain, used in Lemma 5.5 and Proposition 5.7 for the heat-kernel lower bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the Neumann heat kernel upper bound (Theorem 3.2.9) used to prove the matching upper estimate in the cost convergence proof for Theorem 5.2."},{"cited_title":"Karatzas and S","cited_arxiv_id":null,"evidence_quote":"Provides the Brownian bridge computations and modulus-of-continuity estimates used in Proposition 5.7 to control the probability that an unreflected bridge stays inside the required cylinder."}],"review_version":1}