{"id":"fc9b8144-1d20-466f-b15b-b20ac3fd38c3","arxiv_id":"1908.03711","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For functions of measures on Riemannian manifolds, the intrinsic and Lions derivatives coincide and equal the gradient of the extrinsic derivative, giving a direct limit formula for computing them.","lead":"This paper proves that four different definitions of derivatives of functions of measures, introduced independently in the literature, agree on Riemannian manifolds for a large class of functions. It provides a simple formula to compute the intrinsic or Lions derivative as the ordinary gradient of the extrinsic derivative.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.1(3) is false as stated: on R with p=2, f(η)=∫(1−cos x²)/2 dη lies in C^{E,1,1}_B but is not L-differentiable, so D^L=∇D^E fails for the stated class.","rationale":"The reader identified the regularity of C^{E,1,1}_B as the delicate point, which is the right location, but did not identify the actual failure mode. The density approximation in Theorem 2.1(3)(a) is also terse and the 'inf ρ>0' assertion is problematic on noncompact manifolds, but that issue is probably fixable and is not decisive. The decisive problem is the dominated-convergence step in part (b): the growth assumption in Definition 1.1(5) bounds ∇D^E f by C(1+ρ^p), whose square is not η-integrable for general η∈M_p, so the DCT is not justified. The counterexample above is a concrete, minimal instance of the failure: f is a linear function of the measure with a smooth bounded h, so all extrinsic regularity conditions hold, yet the second derivative of h is unbounded and prevents Fréchet L-differentiability in the sense of Definition 1.4(2). This invalidates the paper's central formula D^L=∇D^E for the stated class and affects Corollary 2.2(3). I therefore recommend REJECT rather than a conditional acceptance, because the central theorem is false as stated; a corrected version would need strictly stronger hypotheses and a substantially different proof.","tokens_in":14510,"tokens_out":22971,"duration_ms":251355,"concrete_test":"Run the one-dimensional counterexample: set M=R, p=2, h(x)=(1−cos x²)/2, f(η)=∫h dη, x_n=(π/2+2πn)^{1/2}, a_n=x_n^{-4}, η=∑ a_n δ_{x_n}, and v_n with value c_n=x_n²/(1+2x_n²) at x_n and 0 elsewhere. Verify f∈C^{E,1,1}_B exactly as above, then compute Q_n=[f(η∘φ_{v_n}^{-1})−f(η)−∫h′v_n dη]/‖v_n‖_{L²(η)}. The second-order Taylor term gives Q_n→1/2 while ‖v_n‖_{L²(η)}→0, so Definition 1.4(2) fails and the claimed L-differentiability is disproved.","verdict_should_be":"REJECT","load_bearing_attack":"The proof of Theorem 2.1(3)(b) claims L-differentiability of f∈C^{E,1,1}_B by applying dominated convergence to the squared remainder normalized by ‖v‖². The only available domination uses Definition 1.1(5), which gives |∇D^E f(γ)|(x) ≤ C_L(1+ρ_o(x)^p); after squaring, this is not integrable against arbitrary η∈M_p for p>0. This is not just a missing detail: the implication is false. Take M=R, p=2, h(x)=(1−cos x²)/2 and f(η)=∫h dη. Then D^E f=h, ∇D^E f=h′=x sin x², and |h′|≤|x|, so |∇D^E f|∈L²(η) for every η∈M_2 and the growth condition in Definition 1.1(5) holds; hence f∈C^{E,1,1}_B(M_2). Let x_n²=π/2+2πn, a_n=x_n^{-4}, η=∑ a_n δ_{x_n}∈M_2, and let v_n∈L²(B(TM);η) have value c_n=x_n²/(1+2x_n²) at x_n and vanish elsewhere. Then ‖v_n‖_{L²(η)}=c_n√a_n→0, while Taylor expansion gives [f(η∘φ_{v_n}^{-1})−f(η)−D^L_{v_n}f(η)]/‖v_n‖_{L²(η)}→1/2, contradicting Definition 1.4(2). Thus Theorem 2.1(3) and Corollary 2.2(3) fail for the stated class; additional control on the second derivative of D^E f is needed, not merely on its gradient.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies four notions of derivatives for functions on the space of finite measures over a Riemannian manifold: the extrinsic derivative D^E, the intrinsic derivative D^I, the L-derivative D^L, and the linear functional derivative D^F. The main theorem (Theorem 2.1) claims that, for a class of functions called C^{E,1,1}_B, the intrinsic and L-derivatives coincide and are both given by the spatial gradient of the extrinsic derivative, D^I f(η)(x)=D^L f(η)(x)=∇{D^E f(η)}(x); and that for C^{E,1}_K functions the extrinsic derivative is a linear functional derivative. Corollary 2.2 extends these statements to probability measures, and Lemma 3.4 gives a derivative formula for laws of random variables. The paper also proves that L-differentiability implies intrinsic differentiability (Theorem 2.1(1)) and derives a formula for ∇f(η+sδ_·) from the L-derivative (Theorem 2.1(4)).","tokens_in":14930,"tokens_out":18073,"duration_ms":173783,"significance":"If the main theorem were correct, it would provide a simple and practically useful bridge between several notions of derivatives in measure that appear in mean-field games, McKean–Vlasov analysis, and measure-valued diffusions. The paper is clearly written, the definitions are carefully set out, and the proof of Theorem 2.1(1) and some of the auxiliary lemmas (e.g., Lemma 3.1 for discrete measures) are sound. However, the central regularity class C^{E,1,1}_B is insufficient for the claimed L-differentiability, and the paper gives a false statement as a main theorem. One auxiliary assertion about linear functional derivatives on the space of finite measures is also false. Consequently, the paper's principal contribution is not established.","major_comments":[{"comment":"Theorem 2.1(3)(b) is false as stated. Let M=R, p=2, o=0, and set h(x)=x^2/2 - (1/2)cos(x^2), f(η)=∫ h dη. Then D^E f(η)(x)=h(x), ∇D^E f(η)(x)=h'(x)=x+x sin(x^2), and |h'(x)| ≤ 2(1+|x|^2), so f∈C^{E,1,1}_B(M_2). Let x_n=√(2π n), a_n=x_n^{-4}, and η=∑ a_n δ_{x_n}∈M_2. For N≥1 define v_N by v_N(x_n)=1/x_n for n≥N and v_N=0 otherwise. Then ‖v_N‖_{L^2(η)}^2=∑_{n≥N} x_n^{-6}→0 as N→∞. The weak L-derivative candidate is D^L f(η)=h', so D^L_{v_N}f(η)=∑_{n≥N} a_n. Taylor expansion of h at x_n gives h(x_n+c)=h(x_n)+c x_n+(1/2)c^2(1+2x_n^2)+O(c^2/x_n^2), hence f(η∘φ_{v_N}^{-1})-f(η)=∑_{n≥N} a_n(2+O(x_n^{-2})). Therefore the L-differentiability remainder equals approximately ∑_{n≥N} a_n, while ‖v_N‖=(∑_{n≥N} x_n^{-6})^{1/2}; both quantities are of order N^{-1}, so the ratio does not tend to 0. Thus f is not L-differentiable at η, contradicting Theorem 2.1(3)(b). The proof fails at the dominated convergence step: the only bound available from Definition 1.1(5) is |∇D^E f|≤C(1+ρ_o^p), whose square is not η-integrable for p>0. Additional control on the second derivative of D^E f is required.","section":"Theorem 2.1(3)(b)"},{"comment":"Theorem 2.1(2) is also false as stated for finite measures. Take M=R, p≥0, and f(η)=η(M)^2. Then D^E f(η)(x)=2η(M), which is continuous in (x,η); for any compact K⊂M_p the total mass η(M) is bounded on K, so f∈C^{E,1}_K(M_p). However, the linear functional derivative condition (1.3) in Definition 1.5 requires sup_{η(ρ_o^p)≤L}|D^F f(η)(y)|≤C(1+ρ_o^p(y)). For this f the natural candidate D^F f(η)(y)=2η(M) fails because the set {η:η(ρ_o^p)≤L} is not bounded in total mass (e.g., η=nδ_o has η(ρ_o^p)=0 but η(M)=n). Thus (1.3) cannot hold, and the claimed statement f∈C^{E,1}_K ⇒ f has a linear functional derivative is false on M_p. The proof of Theorem 2.1(2), which only cites Lemma 3.2, does not address the growth condition (1.3).","section":"Theorem 2.1(2)"},{"comment":"The density reduction in the proof of Theorem 2.1(3)(a) is not valid on noncompact manifolds. The proof states that any η∈M_p can be approximated by measures of the form (4.1), η(dx)=ρ(x)dx with ρ∈C_b^∞(M) and inf ρ>0. On a noncompact manifold (such as R), a bounded positive function with strictly positive infimum cannot be integrable with respect to the volume measure, so no such η belongs to M_p. Consequently, the reduction to (4.2) is only established for a class of infinite measures (or for compact manifolds), and the limit argument does not cover the stated class. This leaves the intrinsic-derivative formula (2.1) without a complete proof as stated.","section":"Proof of Theorem 2.1(3)(a)"}],"minor_comments":[{"comment":"There are a number of typographical errors, e.g., 'funtions' in the introduction, 'Off course' in Remark 1.1(a), and 'the the' in the introduction. These should be corrected in any revision.","section":"Minor"}],"recommendation":"reject","confidential_remarks":"The reader's report recommended acceptance with high confidence, but the stress-test counterexample is valid and, in my assessment, directly falsifies Theorem 2.1(3)(b) and the corresponding part of Corollary 2.2(3). The additional counterexample to Theorem 2.1(2) shows that the linear functional derivative assertion also fails for finite measures. Together these are load-bearing errors in the main theorem. I recommend rejection; a substantially revised version with stronger regularity assumptions (e.g., L^2 control on ∇D^E f or second-order control) and a corrected proof of the intrinsic part could be resubmitted as a new paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the headline identity D^L = ∇D^E is not true for the function class C^{E,1,1}_B. The proof of Theorem 2.1(3)(b) has a domination gap, and the theorem is genuinely false. The paper is still useful in parts, but it needs major revision.\n\nWhat I think is solid: the definitions are set out cleanly, and the paper does clarify existing terminological confusion. Theorem 2.1(1) — L-differentiability implies intrinsic differentiability with equal derivatives — is essentially a one-line observation about the two flows having the same tangent at s=0, and it is right. Theorem 2.1(2) and Lemma 3.2 give the extrinsic derivative as the linear functional derivative under C^{E,1}_K; that looks correct. The intrinsic-derivative half of Theorem 2.1(3), formula (2.1), also appears correct, because the test vector fields are smooth and compactly supported, so the unbounded-growth problems do not arise.\n\nThe soft spot is the L-differentiability half. The proof derives an expression for I_v with a squared integrand and then applies dominated convergence. The only available bound is |∇D^E f| ≤ C(1+ρ_o^p), and its square is not η-integrable for arbitrary η∈M_p when p>0. That is not a minor missing detail; the statement is false. On R with p=2, set h(x)=(1−cos x²)/2 and f(η)=∫h dη. Then D^E f=h, ∇D^E f=h'=x sin x², and |h'|≤|x|, so f∈C^{E,1,1}_B(M_2). Take η=Σ a_n δ_{x_n} with x_n=√(2πn), a_n=x_n^{-5}; this has finite second moment. Let v_n be c_n at x_n and 0 elsewhere, c_n=√x_n. Then ‖v_n‖_{L²(η)} = √(a_n)c_n = x_n^{-2}→0, but f(η∘φ_{v_n}^{-1})−f(η) ≈ (1/2)a_n h''(x_n)c_n² = a_n x_n² c_n², and dividing by ‖v_n‖ gives ≈ 1. So the L-differentiability limit fails. (The stress-test note's own construction used a_n=x_n^{-4} and accidentally produced infinite second moment; the corrected weights above remove that slip.)\n\nSo the paper's advertised formula (2.2) needs stronger hypotheses, e.g. Lipschitz ∇D^E f, or η(ρ_o^{2p})<∞, or bounded second derivative. The authors also need to fix Corollary 2.2(3) accordingly.\n\nWho this is for: someone working on mean-field games or McKean-Vlasov equations who wants a practical way to compute Lions derivatives on manifolds. They should not take (2.2) as stated. The paper does contain a usable treatment of the intrinsic derivative and the Eulerian/extrinsic relation, so it is worth engaging with. If I were the editor I would send it to a careful referee, but with the expectation of major revision; I would not accept it in its current form.","headline":"The paper's advertised identity D^L = ∇D^E is false under the stated C^{E,1,1}_B hypotheses; the L-differentiability proof has a real domination gap, and a corrected counterexample kills Theorem 2.1(3)(b).","tokens_in":15504,"tokens_out":11404,"would_cite":false,"duration_ms":105589,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B05","60B10","58C35"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that, for a broad and natural class of functions on the space of finite measures over a Riemannian manifold, the intrinsic and L-derivatives are exactly the Riemannian gradient of the extrinsic derivative, so the standard…","keywords":["intrinsic derivative","extrinsic derivative","L-derivative","linear functional derivative","Riemannian manifold","finite measure space","probability measure space","derivative formulas in measure"],"falsifier":"Repeat the paper's main identity for a measure with atoms, where the density-approximation step is nontrivial. On $M=\\mathbb{R}$, take $\\eta=\\delta_0+\\delta_1$ and $f(\\eta)=\\eta(h)^2$ with $h(x)=x^2$. The theorem predicts $D^L f(\\eta)(0)=0$ and $D^L f(\\eta)(1)=4$, since $\\nabla D^E f(\\eta)(x)=4x$; direct evaluation of the geodesic-flow difference quotient gives exactly these values. A reader who finds any other value, or a failure of the difference quotient to converge, has falsified the claim.","tokens_in":14282,"feed_emoji":"🧮","tokens_out":10638,"duration_ms":105399,"temperature":0.7,"pith_summary":"This paper sets out to clarify how four notions of derivative for functions of measures—extrinsic, linear functional, intrinsic, and L-derivative—are related on the space of finite measures over a complete Riemannian manifold. Its central claim is that, for functions whose extrinsic derivative is itself differentiable in the spatial variable, the intrinsic derivative and the L-derivative both equal the Riemannian gradient of the extrinsic derivative: $D^I f(\\eta)(x)=D^L f(\\eta)(x)=\\nabla\\{D^E f(\\eta)(\\cdot)\\}(x)$. When this holds, the often abstract intrinsic and L-derivatives become concrete: differentiate the function along a Dirac-mass perturbation, then take the spatial gradient. The paper also extends the formulas to probability measures via the centralised extrinsic derivative, and derives a chain-rule formula for the derivative of the law of a random variable. A sympathetic reader would care because the result turns the calculation of measure derivatives in measure-valued and mean-field problems into elementary calculus on the underlying manifold.","feed_headline":"Four measure derivatives collapse to one gradient formula","feed_subtitle":"On Riemannian manifolds, the intrinsic and L-derivatives equal the gradient of the extrinsic derivative for a natural class of functions.","key_machinery":"The load-bearing object is the identity $D^I f = D^L f = \\nabla D^E f$, where $\\nabla$ is the Riemannian gradient on $M$ and $D^E f(\\eta)(x)=\\lim_{s\\downarrow0}(f(\\eta+s\\delta_x)-f(\\eta))/s$ is the extrinsic derivative. The proof mechanism is a pair of variation lemmas. Lemma 3.1 expresses the change of $f$ under a density perturbation $(1+h_\\varepsilon)\\eta$ as an integral of $D^E f$ against the time-derivative of the density; Lemma 3.2 expresses the derivative along a linear interpolation $(1-r)\\eta+r\\gamma$ as the integral of $D^E f$ against $\\gamma-\\eta$. To obtain the gradient formula, the authors approximate arbitrary finite measures by measures with smooth positive densities, use the divergence theorem to move the derivative off the vector field and onto $D^E f$, and pass to the limit using continuity and growth conditions. For the L-derivative, geodesic flow and parallel transport reduce the difference quotient to an integral of $\\nabla D^E f$ against the vector field.","core_discovery":"The central discovery is a single identity linking two families of derivative notions that were introduced by different routes. Theorem 2.1(3) states that if $f\\in C^{E,1,1}(M_p)$, then $f$ is intrinsically differentiable with $D^I f(\\eta)(x)=\\nabla\\{D^E f(\\eta)(\\cdot)\\}(x)$ for all $(x,\\eta)\\in M\\times M_p$; when $p\\in[0,2]$ and $f\\in C^{E,1,1}_B(M_p)$, the same gradient is the L-derivative. Theorem 2.1(4) adds the limiting formula $D^L f(\\eta)(x)=\\lim_{s\\downarrow0} s^{-1}\\nabla f(\\eta+s\\delta_\\cdot)(x)$, and Theorem 2.1(1) states that every L-differentiable function is intrinsically differentiable with $D^I=D^L$. For probability measures, the same links hold with the convex-combination extrinsic derivative $\\tilde D^E$ in place of $D^E$. Taken together, the four derivative notions coincide on a broad function class, and the identity gives a practical route to computing intrinsically defined derivatives from the simpler extrinsic one.","pith_inferences":["If the identity extends beyond the $C^{E,1,1}_B$ class, it would give a practical recipe for derivative-based numerical schemes on Wasserstein space over manifolds: approximate $D^E$ by finite differences in mass, then apply the manifold gradient.","The paper's gap between L-differentiability and extrinsic differentiability suggests a complementary direction: characterising the minimal spatial regularity of $D^E f$ under which the gradient formula still holds, or finding a counterexample at lower regularity.","The law-derivative formula can be read as differentiation under the expectation; a natural testable extension is to measure-dependent SDEs on manifolds, where this identity would yield Bismut-type formulas for the L-derivative without separate arguments.","On flat $\\mathbb{R}^d$ the result recovers known formulas; the new content is that curvature enters only through the Riemannian gradient and geodesic flow, so numerical implementations can treat the manifold as a black box with an exponential map."],"forward_implications":["For any cylindrical function $f(\\eta)=g(\\eta(h_1),\\dots,\\eta(h_n))$ with smooth $g,h_i$, the formula gives $D^L f(\\eta)(x)=\\sum_i (\\partial_i g)(\\dots)\\nabla h_i(x)$, so intrinsic and L-derivatives of such functions reduce to ordinary calculus.","For functions on probability measures, the corresponding formula uses the centralised extrinsic derivative $\\tilde D^E f(\\mu)(x)=D^E f(\\mu)(x)-\\mu(D^E f(\\mu))$, so the same gradient identity holds on $P_p$.","If a family of random variables $\\xi_s$ on $M$ has derivative $\\dot\\xi_0$ in $L^q$, then for $f$ in the appropriate class, $\\lim_{s\\downarrow0}(f(L_{\\xi_s})-f(L_{\\xi_0}))/s = \\mathbb{E}\\langle\\nabla\\{\\tilde D^E f(L_{\\xi_0})\\}(\\xi_0), \\dot\\xi_0\\rangle$, a chain rule for laws of random variables.","Every L-differentiable function is intrinsically differentiable and the two derivatives agree; together with the main gradient formula, the paper's regularity class $C^{E,1,1}_B(M_p)$ is contained in $C^{L,1}(M_p)$ for $p\\le2$.","The limiting formula $D^L f(\\eta)(x)=\\lim_{s\\downarrow0} s^{-1}\\nabla f(\\eta+s\\delta_\\cdot)(x)$ gives a direct way to compute the L-derivative without constructing geodesic flows."],"supporting_citations":[{"why":"Introduces the intrinsic derivative via push-forward by a vector-field flow, the definition whose relation to the other derivatives is the paper's subject.","marker":"[2]"},{"why":"Supplies the definition of the extrinsic derivative as a derivative along Dirac-measure perturbations, the starting point of the main formulas.","marker":"[8]"},{"why":"Introduces the L-derivative and linear functional derivative in the Wasserstein-space setting that the paper generalises to Riemannian manifolds.","marker":"[3]"},{"why":"Contains the comparison result (Proposition 5.48) that the paper extends from $P_2(\\mathbb{R}^d)$ to $M_p$, where $D^L=\\nabla D^F$ for $C^1$ linear functional derivatives.","marker":"[4]"},{"why":"Provides the standard calculus background on Wasserstein space that motivates the derivative notions and the formulation on $P_p$.","marker":"[1]"},{"why":"Gives a different characterisation of the L-derivative that helps locate the paper's result in the existing theory.","marker":"[5]"}],"fun_headline_variants":["Extrinsic, intrinsic, and L-derivatives coincide on manifolds","Gradient of extrinsic derivative equals intrinsic and L","One gradient formula links all measure derivatives","Measure derivatives unified by a single gradient identity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gradient identity is proved only for functions whose extrinsic derivative is itself once differentiable in the spatial variable with a continuous gradient, and, for the L-derivative version, with controlled growth and $p\\le2$; if that spatial regularity is absent, the identity can fail, and functions that are merely L-differentiable need not be extrinsically differentiable at all.","fun_headline_variants_meta":{"raw":{"variants":["Extrinsic, intrinsic, and L-derivatives coincide on manifolds","Gradient of extrinsic derivative equals intrinsic and L","One gradient formula links all measure derivatives","Measure derivatives unified by a single gradient identity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000286,"raw_usage":{"total_tokens":1746,"prompt_tokens":1070,"completion_tokens":676,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":615}},"tokens_in":686,"tokens_out":676,"duration_ms":7277,"temperature":1.0,"reasoning_tokens":615,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:04:58.216532+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the paper's main identity for a measure with atoms, where the density-approximation step is nontrivial. On $M=\\mathbb{R}$, take $\\eta=\\delta_0+\\delta_1$ and $f(\\eta)=\\eta(h)^2$ with $h(x)=x^2$. The theorem predicts $D^L f(\\eta)(0)=0$ and $D^L f(\\eta)(1)=4$, since $\\nabla D^E f(\\eta)(x)=4x$; direct evaluation of the geodesic-flow difference quotient gives exactly these values. A reader who finds any other value, or a failure of the difference quotient to converge, has falsified the claim.","supporting_citations":[{"cited_title":"Differential geometry o f Poisson spaces","cited_arxiv_id":null,"evidence_quote":"Introduces the intrinsic derivative via push-forward by a vector-field flow, the definition whose relation to the other derivatives is the paper's subject."},{"cited_title":"Analytic approach to Fleming-Viot processes with interactive selection","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of the extrinsic derivative as a derivative along Dirac-measure perturbations, the starting point of the main formulas."},{"cited_title":"Notes on mean field games","cited_arxiv_id":null,"evidence_quote":"Introduces the L-derivative and linear functional derivative in the Wasserstein-space setting that the paper generalises to Riemannian manifolds."},{"cited_title":"Probabilistic Theory of Mean Field Games with Applications I","cited_arxiv_id":null,"evidence_quote":"Contains the comparison result (Proposition 5.48) that the paper extends from $P_2(\\mathbb{R}^d)$ to $M_p$, where $D^L=\\nabla D^F$ for $C^1$ linear functional derivatives."},{"cited_title":"Gradient Flows in Metric Spaces and in the Spaces of Probability Measures","cited_arxiv_id":null,"evidence_quote":"Provides the standard calculus background on Wasserstein space that motivates the derivative notions and the formulation on $P_p$."},{"cited_title":"On differentiability in the Wasserstein spa ce and well-posedness for Hamilton-Jacobi equations","cited_arxiv_id":null,"evidence_quote":"Gives a different characterisation of the L-derivative that helps locate the paper's result in the existing theory."}],"review_version":1}