{"id":"913f2dc8-c0b9-4a24-9f77-b3820a63af1b","arxiv_id":"2501.00719","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A system of relative-entropy variational problems and Schrodinger functional equations is introduced and shown to have unique solutions, giving a variational route to a stochastic Knothe-Rosenblatt rearrangement.","lead":"This paper constructs a nested ('system') version of Schrodinger's classical problem of matching two probability distributions with minimal random motion cost, together with the associated functional equations. It proves that each stage has a unique solution, yielding a variational characterization of a stochastic analogue of the Knothe-Rosenblatt rearrangement.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Theorem 2.1's inductive argument is internally consistent; the restrictive convexity assumption (A0)(iii) is explicit and is exactly the load-bearing input the reader flags.","rationale":"The reader's verdict is ACCEPT with moderate confidence, and I agree. The central claim is conditional: given (A0)-(A1), Theorem 2.1 constructs measurable h_i with h_i(y_{n_{i-1}},·) ∈ C(R^{d_i}) and identifies the corresponding π_{μ_i,ν_i} as the unique minimizer of the variational problem when finite. I traced the dependencies. Lemma 3.1(ii) is the pivotal technical step: (A0)(iii) is used to prove that y ↦ log I_i(φ_i)(y_{n_{i-1}}, y) + ψ_i(y) is convex, so I_i is continuous on the interior of its domain. This is exactly what Lemma 3.4 needs to show h_i = f_{ν_i}/I_i is continuous in y_i, and Lemma 3.6 then establishes the measurability needed for the selection step. The proof of Proposition 2.1 correctly reduces the conditional Schrödinger equation (2.9) to the auxiliary equation (4.2), which falls under Jamison's theorem; integrating (4.2) in z recovers (2.9), and the uniqueness statement follows because any solution of (2.9) also solves (4.2) with \\tilde h_i = h_i. The selection argument in Theorem 2.1 is valid: the set f_{ν_i}^{-1}((0,∞)) is open hence σ-compact, a Borel selector ξ_i exists, and the measurability of \\tilde h_i is proved through the limit in (4.8), with no circular use of the conclusion of Theorem 2.1. The only genuine presentation weakness is the omitted proof of Lemma 3.2, but the base case can be treated by the same strategy as Lemma 3.4 with i=1 and Theorem 1.1, so it is not a correctness gap. I also agree with the reader's identification of (A0)(iii) as the weakest assumption: it is strong and central, but it is stated explicitly, and the theorem does not claim to hold without it. Hence the verdict should remain ACCEPT.","tokens_in":22114,"tokens_out":41457,"duration_ms":369851,"concrete_test":"Check whether the claimed diffusion example really satisfies (A0)(iii): for a transition density p with Sheu's bound ∇^2_y log p ≥ -2C I, use the marginal identity ∇^2_{y_i} log p_i = E[∇^2_{y_i} log p | y_{n_i}] + Var(∇_{y_i} log p | y_{n_i}) to verify that ∇^2_{y_i} log p_i + 2C I ≥ 0 uniformly in (x_{n_i}, y_{n_{i-1}}). If this fails for some non-Gaussian kernel, Remark 2.1 overclaims, though Theorem 2.1 remains valid under (A0).","verdict_should_be":"UNCHANGED","load_bearing_attack":"I do not find a load-bearing flaw. The theorem is explicitly conditional on (A0)-(A1), and the proof chain Lemma 3.1(ii) to Lemma 3.4 to Lemma 3.6 to Theorem 2.1 is coherent: (A0)(iii) gives convexity of y_i ↦ log p_i(x_{n_i},(y_{n_{i-1}}, y_i)) + ψ_i(y_i), Lemma 3.1 converts this into continuity of I_i(φ_i)(y_{n_{i-1}},·) on the interior of its domain, and Lemma 3.4 obtains h_i(y_{n_{i-1}},·)=f_{ν_i}/I_i ∈ C(R^{d_i}). The measurable-selection step in Theorem 2.1 is handled by defining \\tilde h_i via a Borel selector ξ_i and proving measurability through the limit in (4.8), so no hidden circularity appears. The omitted proof of Lemma 3.2 is a presentation gap: the argument of Lemma 3.4 with i=1 together with Theorem 1.1 supplies it. Assumption (A0)(iii) is restrictive, and the remark that the diffusion example satisfies it for the marginal kernels p_i is only sketched, but this affects the range of applications, not the validity of the conditional theorem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces an inductive system of variational problems of relative entropy (1.26)-(1.27) with two endpoint marginals, together with an inductively defined system of functional equations (1.29), (1.31), generalizing Schrodinger's problem and functional equation. Under assumptions (A0)-(A1), Theorem 2.1 asserts, for each i=2,...,k0, the existence of a measurable solution h_i of the conditional equation (2.9) with h_i(y_{n_{i-1}},.) continuous, uniqueness up to a multiplicative measurable function of y_{n_{i-1}}, and the unique minimality of the measure pi_{mu_i,nu_i} in (2.13) for V_i whenever V_i is finite. The proof is inductive, using Jamison's theorem for an auxiliary Schrodinger equation, convexity of log p_i + psi_i (Lemma 3.1), weak-continuity and measurability results (Lemmas 3.3-3.6), and a measurable selection argument to regularize h_i.","tokens_in":22343,"tokens_out":26018,"duration_ms":237197,"significance":"The framework is a plausible new route to a stochastic optimal transport analog of the Knothe-Rosenblatt rearrangement. The assumptions are explicit, including the restrictive log-concavity type condition (A0)(iii), and the paper honestly records the limitation that h_i need not be jointly continuous (Remark 2.3). The proof couples standard tools (Jamison's representation, convexity, Lusin's theorem, and a selection lemma) rather than introducing a radically new technique, but the resulting existence and uniqueness theorem for a system of Schrodinger equations appears to be new. I found no circularity or internal inconsistency: the inductive use of previously constructed minimizers is recursion, not a logical loop.","major_comments":[{"comment":"Lemma 3.2 is stated without proof ('We omit the proof'), but it is used directly in the proof of Theorem 2.1 and in Lemmas 3.3 and 3.4. The indicated argument by analogy with Lemma 3.4 is credible, since the i=1 case can be derived from Lemma 3.1(i) and Theorem 1.1 without Proposition 2.1; however, the manuscript as submitted does not contain that derivation. Because this is a load-bearing step of the induction, please supply a self-contained proof of Lemma 3.2 or an explicit reference to a published theorem that yields continuity of the Schrodinger solution h_1 under (A0)(i)-(ii) and (A1).","section":"Section 3, Lemma 3.2"}],"minor_comments":[{"comment":"The assertion that a solution to (2.9) is also a solution to (4.2) is used for the uniqueness claim but is not demonstrated; adding one line with tilde h_i(y_{n_{i-1}},z,y)=h_i(y_{n_{i-1}},y) and integrating z against q_{i-1}dz would make the equivalence explicit.","section":"Section 4, proof of Proposition 2.1"},{"comment":"After (4.8), the measurability of the displayed integral is attributed to Lemma 3.6, but the integrand depends on y_{n_{i-1}} also through xi_i(y_{n_{i-1}}); please state that this dependence can be absorbed into the test function in (3.14).","section":"Section 4, proof of Theorem 2.1"},{"comment":"There are several minor typographical errors, e.g., 'Foy any phi' should be 'For any phi'; a final proofreading pass is needed.","section":"Section 3, proof of Lemma 3.6"},{"comment":"The verification that (A0)(iii) holds for the marginal kernels p_i is only sketched; a sentence noting that log-concavity is preserved under taking marginals (e.g., by Prekopa-Leindler) would make the example more transparent.","section":"Remark 2.1(i)"}],"recommendation":"major_revision","confidential_remarks":"The paper is in scope for the journal and the mathematics appears sound. My recommendation is driven by the single missing proof of Lemma 3.2; if the authors provide that proof or a suitable citation, I expect the paper can be accepted without further substantive review. The self-citation rate is high, but the citations are relevant to the technical tools used."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Genuinely new and mostly clean. The paper extends Schrödinger's classical problem to an inductive system whose minimizers give a variational, stochastic analogue of the Knothe–Rosenblatt rearrangement. Theorem 2.1 is the real content: existence and uniqueness (up to a multiplicative function of the conditioning variable) for the system of functional equations (1.31), and identification of the unique minimizer of the relative-entropy variational problem (1.27). For k0=1 it reduces to the standard Schrödinger problem.\n\nThe proof is solid. The authors use Jamison's theorem for the conditional problems, convexity (Lemma 3.1) to turn (A0)(iii) into continuity of the integral operator, and a selection lemma to handle measurability. I checked the induction and the measurable selection step; it holds together. The assumption (A0)(iii) is strong, but it is explicitly stated and exactly what the proof needs. The finiteness example (Example 2.1) is a nice touch.\n\nSoft spots are minor. Lemma 3.2 is stated without proof, with a note that it follows like Lemma 3.4; I believe that's correct, but in a refereed version they should either prove it or spell out the reduction. The verification that a diffusion transition density satisfies (A0)(iii) is only sketched; for a theorem conditional on (A0), this weakens the sense of applicability, but it doesn't undermine the result. The significance is subfield-level: first existence and uniqueness for a nested Schrödinger system, but not the full stochastic Knothe–Rosenblatt program, which the authors explicitly leave open.\n\nI'd send this to a serious referee. It deserves refereeing rather than desk rejection, and I'd expect acceptance after a modest revision that fills in Lemma 3.2. I would cite it if I worked on multi-marginal entropic transport or triangular transport.","headline":"A genuinely new existence and uniqueness theorem for an iterative Schrödinger system, with a solid proof and an explicit but restrictive convexity assumption; worth refereeing.","tokens_in":22910,"tokens_out":2431,"would_cite":true,"duration_ms":24390,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","93E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"An inductively defined chain of Schrödinger-style relative-entropy problems has measurable solutions and unique minimizers at each stage.","keywords":["Schrödinger's problem","Schrödinger's functional equation","Knothe–Rosenblatt rearrangement","stochastic optimal transport","relative entropy","Bernstein process","h-path process"],"falsifier":"The cleanest falsifier is to exhibit a triple (μ,ν,p) satisfying (A0)–(A1) with V_2 finite for which (1.31) has two solutions not differing by a multiplicative function of y_{n_1}; the theorem's uniqueness assertion would then be false. Numerically, this can be probed by solving the stage-two functional equation by fixed-point iteration on slices and checking whether the assembled coupling reproduces the direct minimizer of V_2.","tokens_in":21874,"feed_emoji":"🧩","tokens_out":10216,"duration_ms":93581,"temperature":0.7,"pith_summary":"This paper studies a system of Schrödinger problems: rather than one relative-entropy minimization between two endpoint measures on R^d, the authors minimize a sequence of entropies, one per coordinate block, where each stage's reference measure is the previous stage's optimal coupling tensored with the conditional distribution of the next block under the starting measure μ. The central claim is that this induction is well-posed: at each stage the associated functional equation has a solution, unique up to a multiplicative function of the past coordinates, and, when the entropy is finite, the displayed coupling is the unique minimizer. The result supplies a variational route to a stochastic optimal-transport analog of the Knothe–Rosenblatt rearrangement, an object for which no general existence theorem had been available. A sympathetic reader should care because the construction converts a multi-marginal transport problem into a chain of one-block Schrödinger bridges, each solvable from the previous one.","feed_headline":"Iterated Schrödinger bridges exist and are unique step by step","feed_subtitle":"Each step solves one functional equation; together they produce the stochastic analog of Knothe–Rosenblatt rearrangement.","key_machinery":"The load-bearing object is the inductively defined reference measure π_{0,i} = π_{μ_{i-1},ν_{i-1}} ⊗ μ_{i|i-1} p_i(· | y_{n_{i-1}}) dy-block, together with the functional equation (1.31) whose unknown h_i is the Radon–Nikodym factor making the i-th marginal correct. The proof mechanism is the convexity assumption (A0)(iii): for each new block, y ↦ log p_i(x_{n_i},(y_{n_{i-1}},y)) + ψ_i(y) is convex, which forces the integral operator I_i(φ)(y_{n_{i-1}},·) to be continuous on the interior of its convex domain (Lemma 3.1). Continuity in the new block is then upgraded to measurability in the past coordinates by a selection lemma that chooses a Borel point (y_{n_{i-1}},ξ_i(y_{n_{i-1}})) where the conditional density is positive, allowing the multiplicative ambiguity to be normalized. Finally, uniqueness of the minimizer comes from disintegrating the relative entropy along the past coordinate and invoking the classical I-projection/Schrödinger-bridge uniqueness at almost every slice.","core_discovery":"The paper's main theorem (Theorem 2.1) states that under assumptions (A0)–(A1), for each i=2,…,k0, the functional equation (1.31) admits a measurable solution h_i defined on $R^{{n_i}}$, with h_i(y_{n_{i-1}},·) continuous on $R^{{d_i}}$, satisfying (2.9) for fν_{i-1}(y_{n_{i-1}})dy_{n_{i-1}}-almost every past coordinate. The solution is unique up to a multiplicative measurable function of y_{n_{i-1}}, and the measure π_{μ_i,ν_i} defined in (2.13) belongs to the admissible class A(μ_i,ν_i;π_{μ_{i-1},ν_{i-1}}) and is the unique minimizer of V_i whenever V_i is finite. At the base, h1 is a continuous solution of the classical Schrödinger functional equation (1.29), and the first coupling π_{μ_1,ν_1} is the classical Schrödinger bridge. The authors thereby extend Jamison's existence theory for a single bridge to a chain of bridges whose reference measures are built recursively, and they frame the whole system as a stochastic counterpart of the Knothe–Rosenblatt rearrangement.","pith_inferences":["A natural test is the zero-noise limit: replacing the transition density by a Brownian kernel with vanishing variance should make the chain of bridges concentrate on the deterministic Knothe–Rosenblatt map, a limit the paper leaves as future work.","Under stronger smoothness or log-concavity of the kernels, the measurable h_i are likely to be continuous in the past coordinates as well, which would resolve the paper's open question about continuous solutions to (1.31).","The recursive construction is not tied to Euclidean space in an essential way; the same block-by-block scheme could be run for any family of positive continuous Markov kernels satisfying the analogous convexity assumption, e.g., on manifolds or graphs.","Computationally, the paper implies a conditional-slice Sinkhorn algorithm: at block i, solve a classical Schrödinger problem for each past coordinate slice and then glue the solutions by the measurable selection; convergence analysis for such an algorithm is not provided here."],"forward_implications":["At every stage with finite entropy, the inductively built coupling π_{μ_i,ν_i} genuinely belongs to A(μ_i,ν_i;π_{μ_{i-1},ν_{i-1}}), so the chain defines a stochastic analogue of the Knothe–Rosenblatt rearrangement without constructing an explicit triangular map.","When k0=1 the system collapses to the classical Schrödinger problem and its functional equation, recovering the known existence and uniqueness theory as the base case.","For the Gaussian product kernel of Example 2.2, the theorem produces an explicit Bernstein-type probability law on C([0,1];R^2), giving a process-level object from the variational construction.","Because (1.31) is the Euler equation of V_i, the minimizer can be sought by solving one functional equation per coordinate block, so the system suggests a block-by-block Sinkhorn algorithm.","The uniqueness statement fixes the continuation of the chain: once π_{μ_{i-1},ν_{i-1}} is known, the next coupling has no free parameter when V_i is finite."],"supporting_citations":[{"why":"Jamison's existence theorem for the Schrödinger system with a positive continuous kernel supplies the solution of the conditional functional equation (4.2) from which h_i is built.","marker":"[16]"},{"why":"Jamison's construction of the Schrödinger Markov process and its positive continuous transition density establishes the classical base case and the h-path measure behind the first bridge.","marker":"[17]"},{"why":"Rüschendorf–Thomsen identifies the Schrödinger bridge as the unique I-projection, which the paper invokes to identify π_{μ_i,ν_i} as the unique minimizer of the conditional variational problem.","marker":"[33]"},{"why":"Fleming–Rishel's measurable selection lemma provides the Borel selector ξ_i used to normalize h_i and prove measurability in the past coordinates.","marker":"[11]"},{"why":"Sheu's estimates show that for a uniformly nondegenerate diffusion, log transition density plus a quadratic is convex, providing the concrete class where (A0)(iii) holds.","marker":"[36]"},{"why":"Mikami's continuity result for Schrödinger's functional equation in the weak topology supplies Lemma 3.5, the continuous-dependence step in the measurability induction.","marker":"[26]"},{"why":"This paper's Knothe–Rosenblatt process is the object the variational system aims to provide an existence route for, and the paper notes no general existence theorem was previously available.","marker":"[25]"}],"fun_headline_variants":["Iterated Schrödinger bridges: unique solutions at each step","Generalized Schrödinger equations have unique solutions","Stochastic analog of Knothe-Rosenblatt via Schrödinger bridges","Schrödinger bridge chain: existence and uniqueness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole proof leans on the assumption that, at every stage, the logarithm of the conditional transition density in the new coordinate block plus a fixed continuous function is convex in that block; if that convexity fails, the continuity and measurability arguments used to construct h_i and the minimizer no longer go through.","fun_headline_variants_meta":{"raw":{"variants":["Iterated Schrödinger bridges: unique solutions at each step","Generalized Schrödinger equations have unique solutions","Stochastic analog of Knothe-Rosenblatt via Schrödinger bridges","Schrödinger bridge chain: existence and uniqueness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001669,"raw_usage":{"total_tokens":6626,"prompt_tokens":953,"completion_tokens":5673,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":5609}},"tokens_in":569,"tokens_out":5673,"duration_ms":38153,"temperature":1.0,"reasoning_tokens":5609,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:44:04.888318+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The cleanest falsifier is to exhibit a triple (μ,ν,p) satisfying (A0)–(A1) with V_2 finite for which (1.31) has two solutions not differing by a multiplicative function of y_{n_1}; the theorem's uniqueness assertion would then be false. Numerically, this can be probed by solving the stage-two functional equation by fixed-point iteration on slices and checking whether the assembled coupling reproduces the direct minimizer of V_2.","supporting_citations":[{"cited_title":"Jamison, Reciprocal processes , Z","cited_arxiv_id":null,"evidence_quote":"Jamison's existence theorem for the Schrödinger system with a positive continuous kernel supplies the solution of the conditional functional equation (4.2) from which h_i is built."},{"cited_title":"Jamison, The Markov process of Schr¨ odinger, Z","cited_arxiv_id":null,"evidence_quote":"Jamison's construction of the Schrödinger Markov process and its positive continuous transition density establishes the classical base case and the h-path measure behind the first bridge."},{"cited_title":"R¨ uschendorf and W","cited_arxiv_id":null,"evidence_quote":"Rüschendorf–Thomsen identifies the Schrödinger bridge as the unique I-projection, which the paper invokes to identify π_{μ_i,ν_i} as the unique minimizer of the conditional variational problem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Fleming–Rishel's measurable selection lemma provides the Borel selector ξ_i used to normalize h_i and prove measurability in the past coordinates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sheu's estimates show that for a uniformly nondegenerate diffusion, log transition density plus a quadratic is convex, providing the concrete class where (A0)(iii) holds."},{"cited_title":"Mikami, Regularity of Schr¨ odinger’s functional equation in the weak topology and moment measures , J","cited_arxiv_id":null,"evidence_quote":"Mikami's continuity result for Schrödinger's functional equation in the weak topology supplies Lemma 3.5, the continuous-dependence step in the measurability induction."},{"cited_title":"Mikami, A characterization of the Knothe–Rosenblatt processes by a convergence result , SIAM J","cited_arxiv_id":null,"evidence_quote":"This paper's Knothe–Rosenblatt process is the object the variational system aims to provide an existence route for, and the paper notes no general existence theorem was previously available."}],"review_version":1}