{"id":"13240ab5-0091-47f9-9fc0-b053d66ce190","arxiv_id":"2501.09097","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The minimizer of a distributional least-squares problem recovers a conditional distribution under phi-divergence and a projected marginal under Wasserstein distance.","lead":"The paper solves a least-squares analogue on probability distributions: which input distribution best explains a target distribution after a forward map. The answer depends on the chosen divergence: phi-divergences reconstruct the conditional distribution on the reachable range, while Wasserstein distance reconstructs the projected (marginal) distribution.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2 is overclaimed: for non-strictly convex φ (e.g. total variation) the φ-divergence minimizer is not unique, and only one minimizer is the conditional; the theorem needs strict convexity or an existence quantifier.","rationale":"The reader's verdict CONDITIONAL is appropriate, and my stress-test supports it rather than overturning it. I identify a different load-bearing weakness than the reader's primary one. The reader focused on the measurable left-inverse/single-valued projection gap. The measurable-selection gap is real as a proof gap but is patchable: for a Borel map G, a universally measurable selection exists, so every measure supported on R can be lifted to P(Θ); the proof's 'left inverse' can be replaced by this standard result. The more serious issue is the quantifier: the paper proves that ρ^c_{y|R} is a minimizer, but its theorem and abstract claim that the minimizer (any minimizer) is recovered. For non-strictly convex φ-divergences such as total variation, the minimizer set is genuinely larger, as the discrete TV computation shows. For W_p, the analogous non-uniqueness arises whenever the metric projection onto R is set-valued. Both failures are independent of the right-inverse issue; they are inherent to the statement 'G#ρ*_x = ...' for an arbitrary minimizer ρ*_x. The central conceptual insight — that φ-divergences point to a conditional and Wasserstein distances point to a projected marginal — is still correct in the sense that there exists a minimizer of that form, and it becomes a full characterization if φ is strictly convex and the projection is single-valued. Therefore the paper needs a corrected theorem statement, but not a rejection of the underlying idea. Hence the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":5276,"tokens_out":17058,"duration_ms":178949,"concrete_test":"Run the discrete two-atom computation: Θ = {a,c}, G = identity, ρ_y = 0.2δ_a + 0.3δ_b + 0.5δ_c, and D = total variation φ(t)=|t−1|. Evaluate D(p_λ||ρ_y) = |λ−0.2| + |0.5−λ| + 0.3 at λ = 2/7 (the conditional ρ^c_{y|R}) and at λ = 0.3. Both values equal 0.6, so λ = 0.3 is a minimizer whose pushforward is not the conditional. This disproves the uniqueness assertion in Theorem 3.2 as stated. If the authors instead intend 'there exists a minimizer recovering the conditional' or add strict convexity of φ, the theorem statement and abstract must be revised accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 3.2 (Section 3) uses Jensen's inequality to show Dφ(ρ'_y||ρ_y) ≥ ν(1)φ(1/ν(1)) + ν(0)φ(0), with equality at ρ'_y = ρ^c_{y|R}. The final line, 'completing the proof', implicitly treats this as the unique minimizer. That is only justified when φ is strictly convex. For non-strictly convex φ, equality in Jensen is not unique. A concrete instance: let Θ = R = {a,c}, G = id, and ρ_y = 0.2δ_a + 0.3δ_b + 0.5δ_c, so ν(1) = 0.7. Every feasible pushforward has the form p_λ = λδ_a + (1−λ)δ_c. With the total-variation divergence φ(t)=|t−1|, D_TV(p_λ||ρ_y) = |λ−0.2| + |0.5−λ| + 0.3, which is minimized for every λ ∈ [0.2, 0.5]. The conditional ρ^c_{y|R} corresponds to λ = 2/7, but λ = 0.3 is another minimizer whose pushforward is not the conditional. Thus the theorem's statement 'G#ρ*_x = ρ^c_{y|R}' for an arbitrary minimizer ρ*_x is false as written. The same quantifier problem appears in Theorem 4.2: if the metric projection PG is set-valued (e.g., R is a circle and ρ_y is a Dirac at the center), many measures on R achieve the same W_p distance, so not every minimizer equals a fixed selection PG#ρ_y. The paper proves existence of a minimizer of the claimed form, not uniqueness. This is independent of the measurable-selection issue: even with a perfect global left-inverse of G, the non-uniqueness remains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the variational problem ρ_x^∗ = argmin_{ρ_x} D(G#ρ_x, ρ_y) over probability measures on the input space Θ, where G is a deterministic forward map to R^n and D is either a φ-divergence or the p-Wasserstein distance. The main claims are Theorem 3.2, asserting that for any φ-divergence with convex φ, every minimizer's pushforward equals the conditional distribution of ρ_y on the range R = G(Θ), and Theorem 4.2, asserting that for any W_p with p ≥ 1, every minimizer's pushforward equals P_G#ρ_y where P_G is the metric projection onto R. The proofs are short: Theorem 3.2 uses measure disintegration and Jensen's inequality; Theorem 4.2 compares transport cost through the projection inequality. The paper also includes a discussion distinguishing the formulation from Bayesian inversion. Existence of a minimizer is explicitly assumed in footnote 1 and in both theorems.","tokens_in":5645,"tokens_out":6889,"duration_ms":72579,"significance":"If the dichotomy were true as stated, it would be a clean and useful conceptual result: the choice of divergence in a distribution-space least-squares problem selects which aspect of the target ρ_y is recovered. The proofs are elegant and self-contained, and the explicit assumption that a minimizer exists is honest. However, the theorems as written are too strong: non-uniqueness of minimizers for non-strictly convex φ and for set-valued metric projections breaks the claimed equality for arbitrary minimizers, and the proof relies on an unstated measurable-selection condition on G. The central idea is defensible under additional hypotheses, but the paper's main statements require revision before they can be accepted.","major_comments":[{"comment":"The theorem states that any minimizer ρ_x^* satisfies G#ρ_x^* = ρ_{y|R}^c. The proof shows that ρ_{y|R}^c attains the Jensen lower bound, but it does not prove that this minimizer is unique. For non-strictly convex φ, equality in Jensen's inequality is not unique. Concretely, take Θ = R = {a,c}, G = id, and ρ_y = 0.2δ_a + 0.3δ_b + 0.5δ_c. Every feasible pushforward has the form p_λ = λδ_a + (1−λ)δ_c, and for the total-variation divergence φ(t) = |t−1| one computes D_TV(p_λ || ρ_y) = |λ−0.2| + |0.5−λ| + 0.3, which is minimized for every λ ∈ [0.2, 0.5]. The conditional ρ_{y|R}^c corresponds to λ = 2/7, but λ = 0.3 is another minimizer whose pushforward is not the conditional. Thus the equality for an arbitrary minimizer is false. The theorem should be corrected by assuming φ is strictly convex (or by stating that there exists a minimizer with the claimed pushforward, rather than every minimizer).","section":"Section 3, proof of Theorem 3.2"},{"comment":"The projection P_G(y) = argmin_{y∈R} d(y,y) is used as a single-valued map in Theorem 4.2, but the argmin is generally set-valued. When the projection is not unique on a set of positive ρ_y-mass, the conclusion G#ρ_x^* = P_G#ρ_y cannot hold for an arbitrary minimizer. For example, if R is a circle and ρ_y is a Dirac mass at the center, every measure supported on R is a minimizer of W_p, not only P_G#ρ_y for a fixed selection. The theorem needs an explicit hypothesis that the metric projection is single-valued ρ_y-almost everywhere (or a statement that the conclusion holds only for a suitably constructed minimizer). In addition, the compactness assumption on R in Definition 4.1 is not restated in Theorem 4.2, although the proof relies on the projection being well-defined and measurable.","section":"Section 4, Definition 4.1 and Theorem 4.2"},{"comment":"The '⊇' direction of equation (4) and the feasibility assertion in Theorem 4.2 require that every probability measure supported on R can be written as G#ρ_x for some ρ_x ∈ P(Θ). This holds only if G admits a measurable right-inverse (a selection h: R → Θ with G(h(y)) = y for y ∈ R), which is what the paper informally calls a 'left-inverse function of G'. For a general nonlinear, non-injective map G, such a measurable selection may not exist. If it fails, the candidate measures ρ_{y|R}^c and P_G#ρ_y may lie outside the feasible set, and the minimizer may differ from the claimed ones. This is a missing hypothesis, not merely a gap in the proof; it should be stated explicitly, for example by requiring the existence of a measurable selection on R (or at least ρ_y-a.e. on R).","section":"Equation (4) and proof of Theorem 4.2"}],"minor_comments":[{"comment":"The abstract and the introductory bullet list state the recovery results without the qualifications needed by the theorems: strict convexity of φ, single-valuedness of the projection, and existence of a measurable selection. The abstract's phrase 'for any φ-divergence' is therefore misleading when read together with the proof.","section":"Abstract and Introduction"},{"comment":"The conditional distribution ρ_{y|R}^c is defined only when ρ_y(R) > 0. The paper does not discuss the case ρ_y(R) = 0, where the theorem statement would involve a division by zero or a degenerate conditional. A short remark on this edge case would be useful.","section":"Definition 3.1"},{"comment":"The statement that non-absolutely-continuous ρ'_y are 'irrelevant' because D_φ 'achieves the maximum value' is imprecise: for KL and χ^2 the divergence is infinite, while for TV the value depends on the definition and may not be a maximum in the usual sense. This does not affect the main proof, but the wording should be corrected.","section":"Footnote 2"},{"comment":"The map called a 'left-inverse function of G' in the proofs is in standard terminology a right-inverse or a measurable selection of G, since it satisfies G(h(y)) = y on R. Using standard terminology would avoid confusion.","section":"Terminology"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short, readable note with a worthwhile conceptual dichotomy, but the stated theorems are overbroad. The fixes are local: add strict convexity to Theorem 3.2, add a single-valued-projection hypothesis (and compactness) to Theorem 4.2, and state the measurable-selection condition for feasibility. With those changes, the results would be correct and likely publishable. I do not see a novelty concern; the authors acknowledge prior related work. The main risk is that the non-uniqueness examples may be considered obvious to readers, so the revision must respond to them explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a clean conceptual note. The genuinely new part is the characterization for general nonlinear G: φ-divergence minimizers push forward to the conditional of ρ_y on the range, Wasserstein minimizers push forward to the projection of ρ_y onto the range. The linear case was in their earlier work; the extension here is modest but real, and the measure-disintegration / projection framing is right.\n\nWhat's good: the proofs are short, the Jensen argument is basically correct for strictly convex φ, and the Wasserstein argument is standard and correct when the projection is single-valued. The paper is honest: it says the linear case appeared before, explicitly assumes existence of minimizers, and even invites reader criticism. It does not overfit or hide derivations.\n\nSoft spots, roughly in order.\n\nFirst, Theorem 3.2 is overclaimed as written. Equality in Jensen gives that the conditional is a minimizer, not that every minimizer equals the conditional. For non-strictly convex φ, e.g. total variation, the set of minimizers can be an interval. The paper needs either strict convexity or an existential quantifier—'there exists a minimizer with G#ρ*_x = ...' is what the proof actually establishes. The same quantifier issue appears in Theorem 4.2 when P_G is set-valued; only a selection of P_G#ρ_y is guaranteed.\n\nSecond, both proofs use an unstated measurable-selection/left-inverse assumption. Equation (4) claims every measure supported on R is representable as G#ρ_x; that requires a measurable right-inverse of G on R, or at least a measurable selection of preimages. For general nonlinear non-injective G this is not automatic. This is fixable by adding a hypothesis, and it does not kill the main idea.\n\nThird, existence of the minimizer is assumed. The paper acknowledges this in footnote 1, so it is not a hidden flaw, but it does mean the theorems are conditional on well-posedness.\n\nThe abstract says 'the reconstructed G#ρ*_x recovers...' which inherits the uniqueness problem. Still, the structural insight—divergence choice selects conditional vs. marginal recovery—is sound and useful.\n\nBottom line: this deserves peer review. A referee should ask for a clean statement of the measurable-selection and strict-convexity/uniqueness assumptions, and a careful set-valued version of the Wasserstein claim. The mathematical core is fine for a note.\n\nDisagree with the stress-test note? I checked the TV example; it lands. The same non-uniqueness is already visible in the Jensen proof, so I count this as a real gap, not a manufactured one.","headline":"A clean conceptual note on how divergence choice selects between conditional and projected recovery; the main theorem needs a uniqueness fix for non-strictly convex φ.","tokens_in":6198,"tokens_out":1616,"would_cite":true,"duration_ms":16694,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","62B10","28A50"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a least-squares problem over probability measures, the choice of divergence decides what the optimal pushforward reproduces: φ-divergences recover the conditional distribution of the data on the range of the map, while Wasserstein…","keywords":["probability measure space","variational problem","phi-divergence","Wasserstein distance","conditional distribution","marginal distribution","pushforward measure","inverse problem"],"falsifier":"Construct a Borel map G whose range R admits no measurable right-inverse, such as a map whose image is a non-Borel analytic set, and take ρ_y with positive mass outside R; then ρ^c_{y|R} is not in {G#ρ_x : ρ_x∈P(Θ)}, so if the true minimizer's pushforward differs from ρ^c_{y|R}, the conclusion of Theorem 3.2 fails. A simpler check for the Wasserstein case: with non-compact R, the projection P_G may not exist, so P_G#ρ_y is undefined and the theorem's statement needs modification.","tokens_in":5023,"feed_emoji":"🎯","tokens_out":7543,"duration_ms":78265,"temperature":0.7,"pith_summary":"The paper studies the variational problem of finding a probability measure ρ_x whose pushforward under a forward map G best matches a given data distribution ρ_y, measured by a divergence D. It claims that the choice of D is not cosmetic: if D is any φ-divergence, the optimal pushforward equals the conditional distribution of ρ_y restricted to G's range; if D is any Wasserstein distance W_p, it equals the pushforward of ρ_y under projection onto that range. This mirrors the classical fact that in Euclidean least squares the choice of norm changes the minimizer, but here the distinction is between conditional and marginal aspects of the target distribution. The consequences matter for inverse problems where one wants a distributional fit rather than a point estimate.","feed_headline":"Choice of divergence picks conditional or marginal recovery","feed_subtitle":"For φ-divergences the optimal pushforward matches the conditional distribution; for Wasserstein, the projected marginal.","key_machinery":"For the φ-divergence case, the engine is the disintegration of ρ_y with respect to the two-valued map T(y)=1 if y∈R, 0 otherwise. Writing ρ_y = ρ_y(R) ρ^c_{y|R} + ρ_y(R^n\\R) ρ^c_{y|R^c}, the divergence splits into a constant off-range term plus an integral over R; Jensen's inequality on the convex φ gives a lower bound that is attained exactly at ρ^c_{y|R}. For the Wasserstein case, the projection P_G(y)=argmin_{z∈R} d(z,y) converts any coupling cost into a lower bound d(P_G(y),y)^p, yielding W_p(G#ρ_x,ρ_y)^p ≥ W_p(P_G#ρ_y,ρ_y)^p, with equality achieved when G#ρ_x = P_G#ρ_y. Both proofs share a feasibility step: any measure supported on R must be representable as G#ρ_x, which is asserted through a left-inverse of G.","core_discovery":"Two theorems give the exact form of the minimizer. Theorem 3.2: assuming a minimizer ρ*_x exists for min D_φ(G#ρ_x || ρ_y), the pushforward satisfies G#ρ*_x = ρ^c_{y|R}, where ρ^c_{y|R}(B)=ρ_y(B∩R)/ρ_y(R) is the conditional distribution of ρ_y on the range R = G(Θ). This holds for every convex φ with φ(1)=0, including KL and χ². Theorem 4.2: assuming a minimizer exists for min W_p(G#ρ_x, ρ_y), the pushforward satisfies G#ρ*_x = P_G#ρ_y, where P_G projects each y onto the nearest point of (compact) R. Thus the divergence class entirely determines which aspect of the data distribution is reproduced: off-range data mass contributes a constant penalty under φ-divergences, while under Wasserstein only the projected locations matter.","pith_inferences":["A testable extension: for Sinkhorn divergences, the recovered pushforward may interpolate between the conditional and the projected marginal as the entropic regularization parameter is varied.","If a measurable right-inverse for G on R does not exist, the candidate measures ρ^c_{y|R} and P_G#ρ_y can fall outside the attainable set, so the dichotomy likely needs an explicit reachability hypothesis to hold for arbitrary nonlinear maps.","The same proof strategy (disintegration for convex functionals, projection for metric costs) may yield analogous conditional-versus-marginal dichotomies for other statistical divergences, such as Bregman divergences or the maximum mean discrepancy.","One could use these identities as a diagnostic in inverse problems: comparing the minimizer's pushforward against ρ^c_{y|R} versus P_G#ρ_y tells a practitioner which divergence is actually encoded in their data-fitting objective."],"forward_implications":["All φ-divergences give the same recovered pushforward, so within that class the specific divergence (KL, χ², total variation) does not change the answer.","Wasserstein recovery ignores the shape of ρ_y off the range except through each point's nearest range point, so the off-range mass acts only through its projected distribution.","When G is linear, the two theorems reduce to the familiar linear-space statement: the conditional distribution is recovered under φ-divergences and the marginal under W_p, matching the earlier linear-operator result in the paper's references.","The results give a user a clear criterion: pick a φ-divergence to match the data on the range, or pick W_p to match the projected data.","Because the minimizer remains a probability measure even with unlimited data, the formulation is a genuine distributional least-squares fitting, not a posterior that collapses to a point."],"supporting_citations":[{"why":"Supplies the Measure Disintegration Theorem used to decompose ρ_y into conditional distributions on R and its complement in the proof of Theorem 3.2.","marker":"[AGS08]"},{"why":"Earlier work by the authors that reported the conditional-versus-marginal recovery distinction for the special case of an overdetermined linear operator G=A.","marker":"[LOWY24]"},{"why":"Earlier work reporting a gradient-flow optimization algorithm for the KL-divergence case, situating the present general theorem.","marker":"[LWY23]"},{"why":"Used to support the discussion distinguishing the measure-theoretic optimization formulation from Bayesian inversion.","marker":"[BBE11]"}],"fun_headline_variants":["Divergence choice decides conditional or marginal fit","φ-divergence recovers conditional, Wasserstein marginal","The divergence you pick fixes which distribution you get","Measure-space least squares: divergence class dictates answer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper needs every probability measure supported on the range R to be expressible as G#ρ_x for some ρ_x, which requires a measurable way to pick one preimage of each y in R; for a general nonlinear or non-injective map this selection may not exist.","fun_headline_variants_meta":{"raw":{"variants":["Divergence choice decides conditional or marginal fit","φ-divergence recovers conditional, Wasserstein marginal","The divergence you pick fixes which distribution you get","Measure-space least squares: divergence class dictates answer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000716,"raw_usage":{"total_tokens":3200,"prompt_tokens":911,"completion_tokens":2289,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":2228}},"tokens_in":527,"tokens_out":2289,"duration_ms":17275,"temperature":1.0,"reasoning_tokens":2228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:11:13.880826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a Borel map G whose range R admits no measurable right-inverse, such as a map whose image is a non-Borel analytic set, and take ρ_y with positive mass outside R; then ρ^c_{y|R} is not in {G#ρ_x : ρ_x∈P(Θ)}, so if the true minimizer's pushforward differs from ρ^c_{y|R}, the conclusion of Theorem 3.2 fails. A simpler check for the Wasserstein case: with non-compact R, the projection P_G may not exist, so P_G#ρ_y is undefined and the theorem's statement needs modification.","supporting_citations":[],"review_version":1}