Elyxuné: Sharp Exchange Laws for Lorentzian Measures and a Negative Answer to the Anari-Liu-Oveis Gharan-Vinzant-Vuong Polynomial Approximate-Exchange Conjecture
Year: 2026
Type: article
Abstract: This work gives a negative answer to the polynomial approximate-exchange conjecture posed by Nima Anari, Kuikui Liu, Shayan Oveis Gharan, Cynthia Vinzant and Thuy-Duong Vuong in their STOC 2021 paper, Log-Concave Polynomials IV: Approximate Exchange, Tight Mixing Times, and Near-Optimal Sampling of Forests. The precise target is the unnumbered conjecture in Section 4, immediately after Proposition 25, using the exchange property defined in Definition 24 of arXiv:2004.07220v2, printed page 15. arxiv.org The conjecture concerns a fundamental quantitative question about weighted matroid bases: does log-concavity guarantee symmetric exchange with a loss bounded by a polynomial in the rank? Given positive-weight bases \(S,T\) and an element \(i\in S\setminus T\) prescribed in advance, one seeks \(j\in T\setminus S\) satisfying \[ \mu(S)\mu(T)\leq \alpha\,\mu(S-i+j)\mu(T+i-j). \] The original paper established an exponential bound and conjectured polynomial rank dependence. arxiv.org Elyxuné proves that exponential loss is unavoidable and determines the exact optimal universal constant \[ \boxed{C_r=r\,2^{r-1}.} \] Here \(C_r\) is the smallest constant valid for every rank-\(r\) homogeneous multi-affine log-concave generating polynomial, every finite ground set, every positive-weight pair of bases, and every prescribed exchange element. In this setting, the log-concave and Lorentzian classes coincide. The result admits a sharper formulation in terms of the exchange distance \(d=|S\setminus T|\). Writing \[ P=\mu(S)\mu(T),\qquad E_j=\mu(S-i+j)\mu(T+i-j), \] the manuscript establishes \[ \boxed{ P\leq \frac{2^{d-1}}{d}\left(\sum_j\sqrt{E_j}\right)^2 \leq 2^{d-1}\sum_jE_j \leq d\,2^{d-1}\max_jE_j. } \] Every displayed constant is optimal. Thus the relevant parameter is the distance between the bases, with no additional dependence on ground-set size. The manuscript also determines the complete sharp exchange profile \[ P\leq K_{d,q}\|E\|_q,\qquad \boxed{K_{d,q}=2^{d-1}d^{\,1-1/q}}, \qquad \frac12\leq q\leq\infty, \] where \(\|E\|_q=(\sum_jE_j^q)^{1/q}\) for finite \(q\), and \(\|E\|_\infty=\max_jE_j\). For \(q<1\), this expression is a quasi-norm. Optimality is understood as a supremum over admissible measures; finite attainment is not asserted. The lower bound is constructive. For rational \(1<\lambda<2\), explicit strictly positive rational weights on every basis of the uniform matroid \(U_{r,2r}\) produce a designated pair and prescribed element for which every available exchange has the same ratio \[ r\,\lambda^{r-1}. \] Fixing \(\lambda=3/2\) gives an exponential obstruction exceeding every polynomial in \(r\). Letting \(\lambda\) approach \(2\) produces matching near-extremizers for all the sharp constants simultaneously. These examples lie in the relative interior of the fixed-support Lorentzian stratum and have nonsingular active quadratic Hessians. The obstruction therefore persists with full support and strictly positive coefficients: it reflects an intrinsic limitation of log-concavity, without requiring zero weights or boundary degeneracies. The universal upper bound follows by contracting the common part of the bases, restricting to their union, and aggregating the remaining variables into a weighted ternary polynomial. Quadratic Hessian inequalities telescope to an endpoint estimate; optimization over the aggregation weights yields the square-root exchange bound. The proof covers arbitrary admissible matroid supports, including zero coefficients. Normalization and positive external fields preserve the exchange ratios. The importance of the result is its exact description of the limits of exchange under log-concavity. Lorentzian polynomials connect matroid theory, discrete convex analysis and negative dependence. Annals of Mathematics The theorem settles the polynomial-versus-exponential question for prescribed-element exchange and supplies its matching optimal inequality. Consequently, any polynomial guarantee for this same pointwise exchange rule requires additional structural assumptions. The explicit interior examples provide reproducible tests for proposed improvements, while the sharp constants give precise estimates for mathematical arguments using exchange ratios. This conclusion is compatible with rapid mixing of the corresponding down-up walks. An obstruction to this pointwise exchange inequality does not establish slow mixing or a computational lower bound for an optimization algorithm. Exchange formulations that allow both exchanged elements to be selected have different quantifiers and fall outside the negative conclusion. Additional results include an exact two-ratio transfer region for ternary Lorentzian seeds and a necessary-and-sufficient criterion for extensions \(h(z)+u\,g(z)\). Under the stated fixed-support and nonvanishing hypotheses, the extension criterion becomes a system of linear matrix inequalities in the coefficients of \(h\). The accompanying release contains the manuscript and source, exact-arithmetic verification programs, machine-readable results, and 4,340 structured research records: 12 claim records, 3,816 rational coefficient records and 512 sharp exchange-profile records. Written proofs establish the all-rank results; finite symbolic and exact-arithmetic checks supplement them. Historical priority remains provisional; independent peer review and proof-assistant certification are not claimed.
Language: English
Subfield: Discrete Mathematics and Combinatorics
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7233989412
