{"id":"c3b82cf7-3037-49f0-959d-db900ce41cb2","arxiv_id":"2608.07996","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For c-uniformly convex potentials V, the log-Sobolev and Talagrand deficits control L1 distance to extremizers with exponent 1/19, improved to the optimal 1/2 for radial densities.","lead":"This paper proves quantitative stability estimates for the Bakry-Emery log-Sobolev and Talagrand inequalities, with a universal deficit exponent of 1/19 and an optimal exponent of 1/2 in the radial setting. The results tell analysts how close near-extremizers must be to explicit extremizers and characterize all equality cases.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The limiting step in the proof of Theorem 1.1 can be vacuous: the bound (3.5) is infinite for admissible log-concave f, so l'Hospital cannot yield (3.9).","rationale":"The central claim of the paper is the quantitative deficit estimate (1.2), and its proof hinges on passing the Prékopa–Leindler stability bound (3.4) to the limit λ→0. The passage is performed through the upper bound (3.5) and the l'Hospital limit (3.9). I find that this passage is not justified for the stated class of admissible functions. The linearization step g_λ ≤ g + λ/(2c(1−λ))|∇g|^2 uses only concavity, but the resulting exponential integrand can be non-integrable: the example g_t(x)=−e^{t x} yields f_t in the theorem's hypotheses with δ_LSI(f_t)→0, yet the integral in (3.5) diverges for every λ>0. Consequently, the ratio in (3.5) is +∞, and the l'Hospital computation in (3.6)–(3.9) is not meaningful; (3.9) does not follow from the displayed inequalities. The same λ→0 protocol is used in the radial case (Theorem 1.2) and the Talagrand case (Theorems 1.3–1.4), although in the Talagrand case the convexity control on g may prevent this particular divergence. The defect is therefore load-bearing for the log-Sobolev stability theorems. The reader's weakest assumption noted the lack of a written dominated-convergence justification; my analysis shows a stronger obstruction: there are admissible functions for which the integrand is infinite, so no local dominated convergence argument can succeed without additional hypotheses. Because the main theorems may still be true and the gap may be repairable by an approximation argument (e.g., mollifying g or truncating |∇g|), I recommend a conditional acceptance rather than rejection. A concrete computational check with the one-dimensional family f_t would settle whether the proof's current route can be repaired.","tokens_in":25383,"tokens_out":16282,"duration_ms":149397,"concrete_test":"For n=1, μ_V the standard Gaussian (V=x^2/2+log√(2π), c=1), and f_t(x)=exp(−e^{t x}/2), compute F_t(λ)=∫_R exp(−e^{t x}+λ/(2(1−λ))t^2e^{2t x}) dμ_V(x). Show that F_t(λ)=+∞ for every t,λ>0, while the hypotheses of Theorem 1.1 hold and δ_LSI(f_t)→0 as t→0. If the paper's (3.9) were valid, the proof would require F_t(λ) to be finite and differentiable near λ=0; the divergence contradicts this and demonstrates the gap. A second check: verify whether the authors' l'Hospital argument can be rerun with the actual g_λ (not the linearization) to obtain the same limit.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The proof of Theorem 1.1 reduces to the identity (3.9), obtained by l'Hospital from the upper bound (3.5). The upper bound comes from the pointwise linearization g(z+λ/(1−λ)h) ≤ g(z)+λ/(1−λ)⟨∇g(z),h⟩ for concave g, which yields g_λ(z) ≤ g(z)+λ/(2c(1−λ))|∇g(z)|^2. Replacing g_λ by this linearized version and exponentiating gives the integrand exp(g + λ/(2c(1−λ))|∇g|^2). For a general positive log-concave f∈W^{1,2}, this integrand need not be μ_V-integrable for any λ>0. Example: on R with standard Gaussian μ (c=1), take g_t(x)=−e^{t x}, f_t=e^{g_t/2}. Then f_t is positive, log-concave, in W^{1,2}, and δ_LSI(f_t)→0 as t→0, so the theorem's small-deficit hypothesis is met for small t. But for every λ>0 the integrand exp(−e^{t x}+λ/(2(1−λ))t^2e^{2t x}) grows like exp(δ e^{2t x}) as x→∞ and its integral against dμ is +∞. Thus (3.5) gives ∞≤∞, and the l'Hospital computation in (3.6)–(3.9) has no finite quantity to differentiate; (3.9) is not a consequence of (3.5). Since (3.9) is the bridge from Prékopa–Leindler stability to the deficit estimate (1.2), and the radial and Talagrand proofs repeat the same limiting step, the central argument is incomplete as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies quantitative stability for the Bakry--Emery log-Sobolev inequality and Talagrand transport inequality on R^n under the curvature condition ∇²V − cI_n ≥ 0. The main theorems claim that, for log-concave functions (or measures) with sufficiently small deficit, the L¹ distance to an explicit translate of the reference density is controlled by a power of the deficit, with universal exponent 1/19, improved to the optimal 1/2 in the radial case. The proofs use Prékopa--Leindler stability theorems of Böröczky--De and Figalli--Ramos through a Maurey-type argument, and the results are applied to equality cases and to the Hopf--Lax hypercontractivity deficit.","tokens_in":25802,"tokens_out":14357,"duration_ms":169676,"significance":"The results, if fully established, would be a significant advance: they move quantitative L¹ stability beyond the Gaussian setting to general Bakry--Emery measures, identify equality cases algebraically via Ker(∇²V − cI_n), and give optimal radial exponents with explicit test families. The use of external Prékopa--Leindler stability results and the explicit optimality computations are strengths. However, the main L¹ estimate for the log-Sobolev inequality depends on a limiting step that is not justified as written; this must be repaired before the central claims can be considered proven.","major_comments":[{"comment":"The l'Hospital limit in (3.9) is not justified, and in fact the upper bound in (3.5) is infinite for admissible data. Take n=1, c=1, V(x)=x²/2+const (standard Gaussian), and for t>0 set g_t(x)=−e^{tx}, f_t=e^{g_t/2}. Then f_t is positive and log-concave, belongs to W^{1,2}(R,μ_V), and δ_LSI(f_t)→0 as t→0, so for small t it satisfies the hypothesis of Theorem 1.1. However, for every λ>0 the integral on the right-hand side of (3.5) equals ∫ exp(−e^{tx}+λ t²/(2(1−λ)) e^{2tx}) dμ_V(x), which is +∞ because the exponent tends to +∞ as x→∞. Thus (3.5) reads ∞≤∞; there is no finite function of λ to differentiate, and (3.9) does not follow from the preceding estimates. Since (3.9) is the bridge from Prékopa–Leindler stability to the deficit estimate (1.2), the proof of Theorem 1.1 is incomplete as written.","section":"§3.1, Eqs. (3.5)–(3.9)"},{"comment":"The radial proof explicitly repeats the argument of §3.1, applying Corollary 2.1 instead of Proposition 2.1. It therefore inherits the same invalid λ→0 limiting step. The improved exponent 1/2 in (1.3) is not established by the current proof; a repair of the limiting step in §3.1, or a separate radial argument, is needed.","section":"§3.2, proof of Theorem 1.2/(i)"},{"comment":"The differentiation under the integral in (3.6) is asserted as a 'simple computation' with no domination argument. For the admissible functions g_t in the previous comment, no such domination exists: the λ-dependent integral is infinite for every λ>0. Even if one restricted to functions for which the integral is finite, the manuscript gives no reason that the derivative can be interchanged with the integral for merely log-concave g∈W^{1,2}. The theorem's hypotheses therefore need either a different upper bound, a regularization argument, or an additional structural assumption on g.","section":"§3.1, Eq. (3.6)"}],"minor_comments":[{"comment":"The computation of lim_{λ→0} ε_λ/τ_λ in (4.4) is stated without details; since g is only known to satisfy that x↦(c/2)|x|²+g(x) is convex, a uniform domination argument for derivatives near λ=0 should be supplied.","section":"§4.1, Eq. (4.4)"},{"comment":"The passage from (3.8) to the l'Hospital computation uses the dominated convergence theorem, but no dominating function is exhibited for e^{g/(1−λ)} as λ→0+. For concave g that can take positive values, this is not automatic from f∈W^{1,2}.","section":"§3.1, Eq. (3.8)"},{"comment":"The identity after (4.13) is written 'for every x∈R^n' but the preceding integration is over x with respect to μ_V; it should say 'for μ_V-a.e. x' or be understood in the integrated sense.","section":"§4.2, Eq. (4.13)"},{"comment":"In the optimality proof, the displayed quantity after 'Dividing (3.12) by ε' labels the left-hand side as LHS and then asserts an equality with a liminf; since only a liminf inequality is needed, the equality sign should be replaced by ≥ to be precise.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the l'Hospital step in §3.1. The example in my report is not contrived: it is a smooth, bounded, log-concave function satisfying the small-deficit hypothesis, and it makes the displayed upper bound infinite for every λ>0. The authors should either prove a finite upper bound for a suitable class or rework the transfer from Prékopa–Leindler stability to LSI deficit. I believe the manuscript is likely fixable, so I recommend major revision rather than rejection; however, without a repair, the main LSI theorem and its radial variant are unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper genuinely extends L1 stability for log-Sobolev and Talagrand inequalities from the Gaussian measure to general Bakry–Emery measures, with a universal 1/19 exponent, and proves optimal 1/2 in the radial case plus new equality characterizations. That is real new work, not a repackaging of [5]. But the proof of the main theorems has a technical hole at the central λ→0 limit that is not a cosmetic detail.\n\nWhat works well: the architecture is sound—use Boroczky–De and Figalli–Ramos stability, feed it into a Maurey-type argument, then extract deficits via Kantorovich duality and HWI. The optimality constructions in §3.3 and §4.4 are explicit and check out. The equality characterization in §4.2, tracking through Donsker–Varadhan, is a nice addition. The citation pattern is honest: the co-author paper [5] is the prior Gaussian case, not a load-bearing input.\n\nWhere the soft spot is: in §3.1, the bound (3.5) is derived from gλ(z) ≤ g(z)+ λ/(2c(1−λ))|∇g(z)|^2. For a general log-concave f ∈ W^{1,2}, the exponential of that linearization need not be integrable. The stress-test example is real: on R with standard Gaussian, g_t(x) = −e^{t x} gives f_t ∈ W^{1,2}, small deficit for small t, but ∫ exp(g_t + λ/(2c(1−λ))|∇g_t|^2)dμ = ∞ for every λ>0. So the right-hand side of (3.5) is infinite, and the l'Hospital computation in (3.6)–(3.9) has no finite function to differentiate. This is not an omitted dominated-convergence justification; it's a vacuous bound. The same limit is used in the radial theorem and in the Talagrand proof, so the gap propagates.\n\nProportionately: this is likely fixable. One expects a truncation/approximation argument (e.g., localize where |∇g| is bounded, or approximate g by concave functions with bounded gradient) to recover the same exponent. But it needs to be written out. As it stands, the central proof is incomplete.\n\nWho this is for: anyone working on stability of functional inequalities, especially LSI/Talagrand beyond Gaussian. It deserves a serious referee: the results are substantial enough that the gap should be fixed rather than the paper shelved. My recommendation: send to peer review, and make the authors address the finiteness and interchange of limits in the λ→0 step explicitly.","headline":"Genuine non-Gaussian stability results with an optimal radial exponent, but the key λ→0 limit in the proofs is not justified—the bounding integrand can be infinite for admissible f.","tokens_in":26333,"tokens_out":4370,"would_cite":false,"duration_ms":42468,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35A23","35R45","35B35"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that near-extremizers of the Bakry–Émery log-Sobolev and Talagrand inequalities are quantitatively L1-close to explicit translated densities, with optimal radial exponents.","keywords":["log-Sobolev inequality","Talagrand transport inequality","Bakry–Émery condition","Prékopa–Leindler stability","quantitative stability","hypercontractivity","equality cases"],"falsifier":"Take a sequence of positive log-concave functions with deficits δ_k → 0 and compute their L1 distance to the nearest density $e^{{V(x)-V(x-x0)}}$. Finding a sequence whose distance decays slower than $δ_k^{{1/19}}$ (or slower than $δ_k^{{1/2}}$ for a radial example) would refute the theorem; the paper's own optimality calculation for the 1/2 exponent provides the matching example where the distance decays exactly like $δ^{{1/2}}$.","tokens_in":25204,"feed_emoji":"📐","tokens_out":9426,"duration_ms":88293,"temperature":0.7,"pith_summary":"The paper establishes quantitative stability for the Bakry–Émery log-Sobolev inequality and the Talagrand transport inequality on R^n. It shows that if a positive log-concave function has a very small log-Sobolev deficit, then its normalized square is close in L1 to an explicit density of the form $e^{{V(x)-V(x-x0)}}$; a parallel statement holds for probability measures with small Talagrand deficit. The closeness is controlled by a universal power of the deficit, exponent 1/19, and in the radial case the exponent improves to 1/2, which is optimal. These results imply elementary characterizations of equality cases, and they give a lower bound on the hypercontractivity deficit of the Hopf–Lax semigroup. The proof combines stability of the Prékopa–Leindler inequality with a limiting argument of the kind used to derive these functional inequalities from it.","feed_headline":"A 1/19 power law puts near-extremizers L1-close to translates","feed_subtitle":"Universal exponent controls both inequalities; radial case sharpens to optimal 1/2 and fixes equality cases.","key_machinery":"The mechanism is a limiting argument built on stability theorems for the Prékopa–Leindler inequality. For each λ∈(0,1) the proof assembles three densities uλ(x)=$e^{{g(x)/(1-λ)-V(x)}}$, v(y)=$e^{{-V(y)}}$, and wλ(z)=$e^{{gλ(z)-V(z)}}$, where g is the logarithm of f² and gλ is a sup-convolution; they satisfy wλ((1-λ)x+λy) ≥ uλ(x)^{1-λ} v(y)^λ. A stability theorem for the Prékopa–Leindler inequality bounds the normalized L1 distance between uλ and a translate of v by C_n (ελ/τλ)^{1/19}, with ελ the deficit of the inequality. A limiting step sends λ→0+; the ratio ελ/τλ is converted, by l'Hospital's rule and dominated convergence, into the log-Sobolev deficit (or, in the Talagrand case, into the Talagrand deficit), and compactness of the translating parameters gives the stated estimates. In the radial case the same construction is fed through an improved stability theorem whose 1/2 exponent is inherited.","core_discovery":"The paper's central discovery is that the deficit in either inequality controls an explicit L1 distance to a translated reference density. For every positive log-concave f in $W^{{1,2}}$(R^n, μV) with δ_LSI < $C_n^{{-19}}$, there is x0 with ∫ |f²/α - $e^{{V(x)-V(x-x0)}}$| dμV ≤ C_n $δ_LSI^{{1/19}}$; and for every probability ν ≪ μV with δ_Tal < $C_n^{{-19}}$, the density built from the optimal transport potential and its inf-convolution satisfies an analogous L1 bound. In the radial cases the exponent becomes 1/2 and is shown optimal. A corollary is the equality characterization: equality in the log-Sobolev inequality forces f(x)=a $e^{{⟨x0,x⟩}}$ with x0 in the kernel of ∇²V - cI, and equality in the Talagrand inequality forces ν to be a translate of μV by such a direction. The hypercontractivity deficit of the Hopf–Lax semigroup is bounded below by an integrated log-Sobolev deficit, which yields its equality case.","pith_inferences":["The universal 1/19 exponent appears to inherit the exponent from the external Prékopa–Leindler stability theorem rather than from the geometry of the log-Sobolev inequality itself; if sharper Prékopa–Leindler stability becomes available, the same limiting skeleton should improve the exponent 1/19 while keeping the L1 distance.","The kernel condition x0 ∈ Ker(∇²V - cI) means the extremizers are Gaussian factors only along the directions where V is exactly quadratic; non-Gaussian potentials with flat quadratic directions still exhibit the same stability and equality theory, so the result is not a Gaussian-only phenomenon.","A natural next test is whether the optimal 1/2 exponent extends beyond radial symmetry under, say, a Poincaré or spectral-gap assumption; the current proof needs radiality only to invoke the improved Prékopa–Leindler stability, so the obstacle appears technical rather than conceptual."],"forward_implications":["Functions with log-Sobolev deficit below C_n^{-19} are, up to a small L1 error, one of the densities e^{V(x)-V(x-x0)}; this makes the qualitative rigidity of near-extremizers quantitative.","Probability measures with Talagrand deficit below the same threshold are L1-close to a translate of μV, with the distance-to-translate controlled by δ^{1/19}.","In the radial setting, both deficits have the optimal 1/2 exponent; no absolute exponent larger than 1/2 can work for the whole family of Bakry–Émery potentials.","Equality in the log-Sobolev inequality holds exactly for f(x)=a e^{⟨x0,x⟩} with x0 in Ker(∇²V - cI); equality in the Talagrand inequality holds exactly for translated copies of μV in these same directions.","The Hopf–Lax hypercontractivity deficit is at least c∫_0^t q(τ)^{-2} δ_LSI(e^{q(τ)Qτu/2}) dτ, and its equality case forces u to be affine with slope in the kernel and q to be affine."],"supporting_citations":[{"why":"Supplies the Prékopa–Leindler stability estimate with exponent 1/19 that the limiting argument converts into deficit bounds.","marker":"[17]"},{"why":"Provides the standard route from Prékopa–Leindler and Brunn–Minkowski to log-Sobolev and Talagrand inequalities that the proof adapts.","marker":"[15]"},{"why":"Supplies the improved radial stability estimate with exponent 1/2 used for the optimal radial theorems.","marker":"[33]"},{"why":"Provides the HWI (entropy–Wasserstein–Fisher information) inequality and Talagrand inequality, used to remove the log-concavity assumption in the equality characterization.","marker":"[51]"},{"why":"Provides the Hopf–Lax semigroup hypercontractivity framework and the Hamilton–Jacobi identity used in the application.","marker":"[12]"},{"why":"Provides the variational formula for relative entropy whose equality case identifies translated densities in the Talagrand equality proof.","marker":"[27]"},{"why":"Provides the classical deviation and limiting-argument template that the proof calls Maurey-type.","marker":"[47]"}],"fun_headline_variants":["L1 stability: 1/19 exponent universal, radial 1/2 optimal","Deficit exponent 1/19 bounds L1 distance in LSI and Talagrand","1/19 exponent yields L1-close translates; radial hits optimal 1/2","Equality in LSI and Talagrand forces translated densities","Near-extremizers drawn to translates via 1/19 power law"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing step is that the λ→0 limiting process is justified for merely log-concave functions, where only almost-everywhere differentiability is available; the proof invokes dominated convergence and l'Hospital's rule in this low-regularity setting without writing out the full justification.","fun_headline_variants_meta":{"raw":{"variants":["L1 stability: 1/19 exponent universal, radial 1/2 optimal","Deficit exponent 1/19 bounds L1 distance in LSI and Talagrand","1/19 exponent yields L1-close translates; radial hits optimal 1/2","Equality in LSI and Talagrand forces translated densities","Near-extremizers drawn to translates via 1/19 power law"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001002,"raw_usage":{"total_tokens":4227,"prompt_tokens":917,"completion_tokens":3310,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":3205}},"tokens_in":533,"tokens_out":3310,"duration_ms":25916,"temperature":1.0,"reasoning_tokens":3205,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:36:49.122346+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sequence of positive log-concave functions with deficits δ_k → 0 and compute their L1 distance to the nearest density $e^{{V(x)-V(x-x0)}}$. Finding a sequence whose distance decays slower than $δ_k^{{1/19}}$ (or slower than $δ_k^{{1/2}}$ for a radial example) would refute the theorem; the paper's own optimality calculation for the 1/2 exponent provides the matching example where the distance decays exactly like $δ^{{1/2}}$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Prékopa–Leindler stability estimate with exponent 1/19 that the limiting argument converts into deficit bounds."},{"cited_title":"Bobkov and M","cited_arxiv_id":null,"evidence_quote":"Provides the standard route from Prékopa–Leindler and Brunn–Minkowski to log-Sobolev and Talagrand inequalities that the proof adapts."},{"cited_title":"Figalli and J.P.G","cited_arxiv_id":null,"evidence_quote":"Supplies the improved radial stability estimate with exponent 1/2 used for the optimal radial theorems."},{"cited_title":"Otto and C","cited_arxiv_id":null,"evidence_quote":"Provides the HWI (entropy–Wasserstein–Fisher information) inequality and Talagrand inequality, used to remove the log-concavity assumption in the equality characterization."},{"cited_title":"Bobkov, I","cited_arxiv_id":null,"evidence_quote":"Provides the Hopf–Lax semigroup hypercontractivity framework and the Hamilton–Jacobi identity used in the application."},{"cited_title":"Donsker and S.R.S","cited_arxiv_id":null,"evidence_quote":"Provides the variational formula for relative entropy whose equality case identifies translated densities in the Talagrand equality proof."},{"cited_title":"Maurey, Some deviations inequalities.Geom","cited_arxiv_id":null,"evidence_quote":"Provides the classical deviation and limiting-argument template that the proof calls Maurey-type."}],"review_version":1}