{"id":"12426982-181e-4372-97b2-77127d271c88","arxiv_id":"2506.19064","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For compactly supported probability measures, the logarithmic potential of their free additive convolution equals the minimum of an explicit variational energy, with explicit forms for semicircle and Marchenko-Pastur convolutions.","lead":"This math paper proves a formula that computes the logarithmic potential of the free convolution of two probability measures from simpler ingredients: the R-transform of one measure and the logarithmic potential of the other. The result turns a difficult random-matrix determinant quantity into a one-dimensional minimization, with explicit versions for semicircle and Marchenko-Pastur laws.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified","rationale":"The reader's weakest_assumption correctly identified the bridge between the real-analytic R-transform and the standard holomorphic/formal R-transform as the least self-contained part of the proof. I agree that Lemmas 2.1 and 2.2, together with the inclusion (2.18), are load-bearing. In my own reading, however, I did not find a concrete failure in that bridge: the uniqueness of the maximal real analytic continuation is standard, the additivity on D_mu cap D_nu follows from equality of power series at 0, and the inclusion (2.18) is used only on its positive half for the main theorem, where it follows directly from (2.17). One minor imprecision is that (2.17), as printed, gives only the positive-endpoint half of (2.18); the negative-endpoint half would require the reflected inequality, but the proof of Theorem 1.1 does not use that half. I also checked the derivative computation (4.1), the fixed-point equivalences (1.18), the sign identity (4.6), and the monotonicity argument in Lemma 4.4; I found no algebraic or logical gap. The external results from Mingo-Speicher are quoted accurately in substance, and the auxiliary Theorem 6.1 is explicitly separated from the main proof. For these reasons I do not see a load-bearing concern that would change the reader's ACCEPT verdict.","tokens_in":31499,"tokens_out":35394,"duration_ms":383425,"concrete_test":"Verify Lemma 5.2 from the source: state [10, Theorems 26 and 28] with their exact tangent-disk domains, construct the intersection of each tangent disk with the R-transform function domain from [10, Theorem 17] for compactly supported measures, and check that equality on that open set forces equality of power series at 0 and hence the real-analytic identity R_rho = R_mu + R_nu on D_mu cap D_nu. If any inclusion or equality fails for some compactly supported mu, nu, then Lemma 2.2 and the subsequent fixed-point identification (1.18) would fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I traced the central argument of Theorem 1.1: the unique-minimizer statement reduces to Lemma 4.4, and the value identity reduces to Corollary 4.6, both of which depend on the fixed-point and sign analysis in Lemma 4.2. That analysis in turn uses the inclusion (2.18), the positivity R'_mu(-g)>0 on the relevant interval, and the additivity of the real R-transform from Lemma 2.2. I checked the algebra in (1.16)-(1.18), the sign bookkeeping in Lemma 4.2, the change of variables in Theorem 1.3, and the derivation of Lemmas 2.1-2.2 in Section 5. I could not find a concrete step where the argument fails. The least self-contained point is the bridge to Mingo-Speicher: Lemma 5.2 relies on [10, Theorems 26 and 28] and on equality of power series near 0; the proof is terse, but the required domain intersections and analytic-continuation steps are standard and appear valid. The auxiliary Theorem 6.1 is explicitly not used in the proof of Theorem 1.1 or Theorem 1.3, so I did not treat it as load-bearing. No significant objection identified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper establishes a variational formula for the logarithmic potential U_{μ⊞ν}(z) when μ,ν are compactly supported probability measures on R, μ non-degenerate, for all real z below the left edge z*_{μ⊞ν} of the support of μ⊞ν. The formula expresses U_{μ⊞ν}(z) as the infimum, over g in (0,g*_{μ⊞ν}), of an energy E_{μ,ν,z}(g) built from the real R-transform of μ and the logarithmic potential of ν; the unique minimizer is identified as g = G_{μ⊞ν}(z). Explicit semicircle and Marchenko-Pastur special cases are derived, and Theorem 1.3 gives formulas for z*_{μ⊞ν} and g*_{μ⊞ν} in terms of the auxiliary function F_{μ,ν}(h) = R_μ(-G_ν(h)) + h. The proof combines a derivative computation for E_{μ,ν,z}, a sign analysis of the fixed-point equation g = G_ν(z - R_μ(-g)), monotonicity of the inverse Stieltjes transform, and a comparison of derivatives and limits to identify E_{μ,ν,z}(G_{μ⊞ν}(z)) with U_{μ⊞ν}(z). An auxiliary complex invertibility theorem for Stieltjes transforms is included but explicitly not used in the proofs of the main results.","tokens_in":31635,"tokens_out":17016,"duration_ms":177427,"significance":"If the main theorem holds, this is a clean and useful addition to the free-probability toolkit: it gives a variational description of the logarithmic potential of free additive convolutions, with direct applications to determinant asymptotics for sums of independent random matrices and to random landscape problems. The paper is honest about the presence of spurious critical points outside (0,g*_{μ⊞ν}), and it carefully restricts the infimum to the interval where the minimizer is unique. The explicit semicircle and Marchenko-Pastur formulas, together with the edge-characterization Theorem 1.3, are concrete and likely to be used. The real-analytic R-transform framework is developed carefully, and the connection to the standard formal/holomorphic theory via Mingo-Speicher is handled in a self-contained section. The auxiliary Theorem 6.1 is a bonus that strengthens a known invertibility result; the manuscript correctly states that it is not load-bearing for the main theorems.","major_comments":[],"minor_comments":[{"comment":"Equation (4.20) as printed states sign(E'_{μ,ν,z}(g)) = sign(G_{μ⊞ν}(z) - g), which is opposite to the sign obtained from Lemma 4.2(a)-(b) together with R'_μ(-g)>0 on (0,g*_{μ⊞ν}); the correct identity is sign(E'_{μ,ν,z}(g)) = sign(g - G_{μ⊞ν}(z)). The subsequent display (4.21) is consistent with the corrected sign, so this appears to be a sign typo that should be fixed.","section":"Section 4, Lemma 4.4(b), Eq. (4.20)"},{"comment":"The displayed inequality (2.17) only bounds the left-edge Stieltjes value G_{μ⊞ν}(supp_- μ⊞ν), but the claimed inclusion Dhat_{μ⊞ν} ⊂ Dhat_μ ∩ Dhat_ν also requires the corresponding right-edge inequality. That right-edge half follows by applying the same bound to reflected measures, or by citing the symmetric statement in [8, Lemma 6.1]; since (2.18) is used in Lemma 4.2(c) and Lemma 4.4, this one-line justification should be added.","section":"Section 2.3, Eq. (2.18)"},{"comment":"The sentence 'G_μ(z)≠0 for any z such that |z|<∞, since G_μ is injective' is not by itself a complete proof of non-vanishing; injectivity on B does not rule out a zero unless 0 is known to occur elsewhere in the image. The conclusion is true because Re G_μ(z) has a definite sign for Re z outside the support, so the sentence should be rephrased accordingly.","section":"Section 6, proof of Theorem 6.1(b)"},{"comment":"There are a few minor grammatical slips, for example 'two compactly supported probability measure' in the abstract and 'the inf. resp. sup.' in Section 1; these should be corrected in a final polish.","section":"Abstract and Section 1"}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this with some care and came out more confident than I expected. The paper proves Theorem 1.1, a genuine variational formula for U_{mu push nu}(z) as a one-dimensional infimum built from R_mu and U_nu, with the unique minimizer identified as G_{mu push nu}(z). The semicircle and Marchenko-Pastur special cases (1.6)-(1.7) are explicit and ready for use, and Theorem 1.3 pins down the left edge via F(h). None of that is in the cited literature: [6, Lemma 4.3] is a special upper bound, and [11] uses the fixed-point equation without a variational characterization.\n\nThe proof is structurally sound. The derivative identity (4.1), the fixed-point equivalence (1.18), the sign bookkeeping in (4.6), and the uniqueness in (0,g*) via monotonicity all hang together. The way the paper handles spurious critical points, by explicitly restricting the infimum to (0,g*), is honest and necessary. There are no fitted parameters and no circular dependence on the authors’ own work; their [4] appears only as an application. I traced the algebra in Lemnas 4.2 and 4.4, and the derivation of Corollary 4.6, and could not find a concrete failure.\n\nThe soft spots are real but not damning. The bridge between the real-analytic R-transform and the Mingo–Speicher formalism (Lemmas 2.1–2.2, Section 5) is terse and load-bearing, and it imports (2.18) from [8, Lemma 6.1]. A careful referee should check the domain inclusions and the analytic-continuation steps there; that is the part I would focus on. The auxiliary Theorem 6.1 is explicitly not used in the main proof, so it adds length without adding support; it is a genuine result, but its presence is a slight distraction. Minor typos in inline math do not impede reading.\n\nWho gets value: random-matrix and probability people working on determinant asymptotics, Kac–Rice landscapes, spin glass complexity, and empirical-risk surfaces. The explicit formulas for semicircle and Marchenko-Pastur are likely to be cited. The paper deserves a serious referee—the proof is long and uses nontrivial external results, so it needs verification, not a desk rejection. My recommendation: send it to peer review with a referee who knows free probability and will spend time on Section 5.","headline":"A clean, useful variational formula for the logarithmic potential of free additive convolutions, with a long but coherent proof; the main soft spot is the analytic-continuation bridge, not the core argument.","tokens_in":32272,"tokens_out":1619,"would_cite":true,"duration_ms":18506,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46L54","31A10","60B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"For any compactly supported $\\mu,\\nu$ with $\\mu$ non-degenerate, the logarithmic potential of $\\mu\\boxplus\\nu$ below the left edge of its support is the infimum of an explicit one-variable functional, attained uniquely at…","keywords":["free additive convolution","logarithmic potential","R-transform","variational formula","Stieltjes transform","random matrix determinant","semicircle law","Marchenko-Pastur law"],"falsifier":"Take an explicit pair with known R-transform and a two-point $\\nu$, for instance the Figure 1c example $R_\\mu(t)=t+t^2/6$, $\\nu=(\\delta_{-0.17}+\\delta_{-1.17})/2$, $z=-2.77$; compute $U_{\\mu\\boxplus\\nu}(z)$ from the Stieltjes transform of the convolution via the identity chain (1.18), and compare with the infimum of $E_{\\mu,\\nu,z}$ over $E_{\\mu,\\nu,z}\\cap(0,g^*)$ using high-precision quadrature. Any disagreement beyond solver tolerance would falsify Theorem 1.1.","tokens_in":31216,"feed_emoji":"🧮","tokens_out":22241,"duration_ms":183247,"temperature":0.7,"pith_summary":"The paper establishes a variational formula for the logarithmic potential $U_{\\mu\\boxplus\\nu}(z)=\\int\\log|z-\\lambda|\\,(\\mu\\boxplus\\nu)(d\\lambda)$ of the free additive convolution of two compactly supported probability measures on $\\mathbb{R}$. For every real $z$ below the left edge $z^*_{\\mu\\boxplus\\nu}$ of the support of $\\mu\\boxplus\\nu$, the potential is shown to equal the infimum, over $g\\in(0,g^*_{\\mu\\boxplus\\nu})$, of an explicit functional $E_{\\mu,\\nu,z}(g)$ built from the $R$-transform of $\\mu$ and the logarithmic potential of $\\nu$; the unique minimizer is the Stieltjes transform $G_{\\mu\\boxplus\\nu}(z)$. When $\\mu$ is the semicircle law or the Marchenko-Pastur law, the formula reduces to explicit one-line expressions. Logarithmic potentials of additive convolutions control the exponential asymptotics of determinants of sums of independent random matrices, which appear in random-landscape problems such as spin-glass complexity and neural-network loss surfaces.","feed_headline":"Minimizing one function gives the log potential of a free convolution","feed_subtitle":"Below the left edge, the log potential equals an explicit minimum; the minimizer is the Stieltjes transform.","key_machinery":"The central object is the real $R$-transform, defined as the unique maximal real-analytic extension $R_{\\mu}$ of $\\widehat R_{\\mu}(t)=G_{\\mu}^{-1}(-t)-t^{-1}$ on a maximal interval $D_{\\mu}$; the free additive convolution is the law whose real $R$-transform equals $R_{\\mu}+R_{\\nu}$ on $D_{\\mu}\\cap D_{\\nu}$. The carrying identity is the derivative formula $$E'_{\\mu,\\nu,z}(g)=R'_{\\mu}(-g)\\bigl(g-G_{\\nu}(z-R_{\\mu}(-g))\\bigr),$$ whose fixed-point equation is equivalent to $z=G_{\\mu\\boxplus\\nu}^{-1}(g)$ for $g\\in(0,g^*)$. Because $R'_{\\mu}(-g)>0$ on that interval, the sign of the derivative is controlled by whether $g$ lies below or above $G_{\\mu\\boxplus\\nu}(z)$, making the unique critical point the global minimum inside $(0,g^*)$. The companion function $F_{\\mu,\\nu}(h)=R_{\\mu}(-G_{\\nu}(h))+h$ encodes the edges of the support through $g^*=G_{\\nu}(h^*)$ and $z^*=F_{\\mu,\\nu}(h^*)$, with $h^*$ the leftmost critical point of $F$ or the right endpoint of its domain.","core_discovery":"Theorem 1.1 states that for compactly supported probability measures $\\mu,\\nu$ on $\\mathbb{R}$ with $\\mu$ non-degenerate and for every real $z<z^*_{\\mu\\boxplus\\nu}$, $$U_{\\mu\\boxplus\\nu}(z)=\\inf_{g\\in\\widehat E_{\\mu,\\nu,z}\\cap(0,g^*) }E_{\\mu,\\nu,z}(g)=\\inf_{g\\in E_{\\mu,\\nu,z}\\cap(0,g^*) }E_{\\mu,\\nu,z}(g),$$ where $E_{\\mu,\\nu,z}(g)=\\int_0^g sR'_{\\mu}(-s)\\,ds+\\int\\log(\\lambda-z+R_{\\mu}(-g))\\,\\nu(d\\lambda)$; here $\\widehat E$ uses the real-analytic $R$-transform on its natural domain and $E$ uses its maximal real-analytic extension. The unique minimizer of both infima is $g=G_{\\mu\\boxplus\\nu}(z)$. The proof identifies this point as the unique critical point of $E_{\\mu,\\nu,z}$ in $(0,g^*)$ by showing that $g=G_{\\mu\\boxplus\\nu}(z)$ solves the fixed-point equation $G_{\\nu}(z-R_{\\mu}(-g))=g$, and that the derivative changes sign from negative to positive at this point. A companion characterization (Theorem 1.3) expresses the edge parameters as $g^*=G_{\\nu}(h^*)$ and $z^*=F_{\\mu,\\nu}(h^*)$ with $F_{\\mu,\\nu}(h)=R_{\\mu}(-G_{\\nu}(h))+h$, so the formula can be used without first constructing the full convolution.","pith_inferences":["Going beyond the paper, the one-dimensional nature of the minimization suggests a numerical scheme for determinants of sums of random matrices: evaluate $U_{\\mu\\boxplus\\nu}(z)$ by solving the scalar fixed-point equation, avoiding the cost of constructing the full convolution; the paper does not propose this algorithm.","The real-$R$-transform bridge used in Section 5 requires only finite variance for part of the construction, so the variational formula may extend to non-compactly supported measures with finite variance if the domain inclusions (2.18) can be re-established in that class; the paper does not address this extension.","The edge formula $z^*=\\sup_h F_{\\mu,\\nu}(h)$ reads like a variational characterization of a phase boundary: changes in the number of solutions of $F_{\\mu,\\nu}(h)=z$ as $z$ crosses $z^*$ could serve as a stability criterion in random-matrix determinant problems; this dynamical reading is not made by the authors.","Theorem 6.1, the stronger invertibility of $G_{\\mu}$ on $\\{z:\\operatorname{Re}z\\notin\\operatorname{supp}\\mu\\}$, is stated as independently useful; a concrete test would be to see whether it simplifies other free-probability arguments that currently rely on large-disk invertibility."],"forward_implications":["For $\\mu=\\mu_{sc,\\beta}$, Theorem 1.1 yields (1.6): $U_{\\mu\\boxplus\\nu}(z)=\\inf_{g\\in(0,g^*)}\\{\\beta^2g^2/2+\\int\\log(\\lambda-z-\\beta^2g)\\,\\nu(d\\lambda)\\}$, restricted to $\\beta^2g<\\operatorname{supp}_-\\nu-z$, valid for $z<z^*$.","For $\\mu=\\mu_{MP,\\beta}$, the same theorem yields (1.7): $U_{\\mu\\boxplus\\nu}(z)=\\inf_{g\\in(0,g^*)}\\{\\beta\\log(1+g)-\\beta g/(1+g)+\\int\\log(\\lambda-z+\\beta/(1+g))\\,\\nu(d\\lambda)\\}$, with constraint $z-\\operatorname{supp}_-\\nu<\\beta/(1+g)$.","In both special cases $F_{\\mu,\\nu}$ is strictly concave, so $h^*$ is the global maximum and (1.9) gives explicit formulas for $z^*$ and $g^*$ in terms of $G_{\\nu}$; the critical-point trichotomy (1.8) lets one locate the edge by inspecting $E_{\\mu,\\nu,z}$.","For general $\\mu,\\nu$, Theorems 1.1 and 1.3 together give a way to compute $U_{\\mu\\boxplus\\nu}(z)$ below the edge by solving a one-dimensional fixed-point equation rather than by first constructing the density of $\\mu\\boxplus\\nu$.","The identity $U_{\\mu\\boxplus\\nu}(z)=E_{\\mu,\\nu,z}(G_{\\mu\\boxplus\\nu}(z))$ combined with matching derivatives gives a direct variational proof of the logarithmic-potential formula, the input used for determinant asymptotics of sums of independent random matrices in the companion work."],"supporting_citations":[{"why":"Supplies the standard free-probability theory: R-transform power series, free additive convolution, analytic R-transform functions, and the large-disk invertibility of G used to prove Lemmas 2.1 and 2.2.","marker":"[10]"},{"why":"Gives inequality (2.17) through its Lemma 6.1, from which the domain inclusion (2.18) follows and which ensures the minimizer lies in the variational domains.","marker":"[8]"},{"why":"Defines the free additive convolution through R-transform power series, the foundational identity (1.4) on which the real-R-transform bridge rests.","marker":"[13]"},{"why":"Supplies the semicircle-law upper bound in its Lemma 4.3 that formula (1.6) extends and generalizes, and motivates the fixed-point equation.","marker":"[6]"}],"fun_headline_variants":["One minimization yields the free convolution log potential","Log potential of free convolution via a single minimization","Explicit variational formula for free convolution log potential","Log potential from minimizing an explicit energy functional","Variational formula for free additive convolution's log potential"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is the real-analytic bridge: every compactly supported non-degenerate probability measure has a real R-transform with a unique maximal real-analytic extension, and the free additive convolution is exactly represented by additivity of these real R-transforms on the overlap of their domains; if that bridge failed for some measure class, the variational domain $E_{\\mu,\\nu,z}$ and the fixed-point identification of the minimizer with $G_{\\mu\\boxplus\\nu}(z)$ would collapse.","fun_headline_variants_meta":{"raw":{"variants":["One minimization yields the free convolution log potential","Log potential of free convolution via a single minimization","Explicit variational formula for free convolution log potential","Log potential from minimizing an explicit energy functional","Variational formula for free additive convolution's log potential"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000433,"raw_usage":{"total_tokens":2220,"prompt_tokens":974,"completion_tokens":1246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":1176}},"tokens_in":590,"tokens_out":1246,"duration_ms":9826,"temperature":1.0,"reasoning_tokens":1176,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:39:08.130959+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an explicit pair with known R-transform and a two-point $\\nu$, for instance the Figure 1c example $R_\\mu(t)=t+t^2/6$, $\\nu=(\\delta_{-0.17}+\\delta_{-1.17})/2$, $z=-2.77$; compute $U_{\\mu\\boxplus\\nu}(z)$ from the Stieltjes transform of the convolution via the identity chain (1.18), and compare with the infimum of $E_{\\mu,\\nu,z}$ over $E_{\\mu,\\nu,z}\\cap(0,g^*)$ using high-precision quadrature. Any disagreement beyond solver tolerance would falsify Theorem 1.1.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the standard free-probability theory: R-transform power series, free additive convolution, analytic R-transform functions, and the large-disk invertibility of G used to prove Lemmas 2.1 and 2.2."},{"cited_title":"Guionnet, M","cited_arxiv_id":null,"evidence_quote":"Gives inequality (2.17) through its Lemma 6.1, from which the domain inclusion (2.18) follows and which ensures the minimizer lies in the variational domains."},{"cited_title":"Voiculescu,Addition of certain non-commuting random variables","cited_arxiv_id":null,"evidence_quote":"Defines the free additive convolution through R-transform power series, the foundational identity (1.4) on which the real-R-transform bridge rests."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the semicircle-law upper bound in its Lemma 4.3 that formula (1.6) extends and generalizes, and motivates the fixed-point equation."}],"review_version":1}