--- title: Finite Free Information Theory url: https://www.emergentmind.com/topics/finite-free-information-theory type: topic --- # Finite Free Information Theory to=exa_search code 天天爱彩票怎么json {"query":"site:arxiv.org \"Finite Free Information Theory\" arXiv 2026 2602.15822 2604.12212 2604.11922 2605.12536", "numResults": 5}:-------------2$, the Hermite pair itself violates the proposed inequality; for $p<2$, the numerically extremal configurations bifurcate into non-matching pairs with bimodal root structure, converging back to the Hermite diagonal as $p\to 2^-$ [2604.11922]. ## 3. Chronological entropy, matrix microstates, and free information geometry A second major program develops a finite-$n$, microstate-based “Finite Free Information Theory” for noncommutative stochastic processes. Its core object is a new multivariate free entropy $\chi_{\chron}^{U}$, defined from matrix microstates tested by chronological formulas. These formulas belong to a filtered metric language $L_{\operatorname{proc}}$ with domains $D_{t,r}$, algebra operations, a metric, trace components, and constants $z_{s,t,j}$ coding increments of a non-selfadjoint free Brownian motion compatible with the filtration. Chronological formulas are built from quantifier-free trace-polynomial expressions in resolvents and Brownian increments, then closed under continuous connectives, partial suprema and infima in chronological order, and the heat-shift $S_t$ [2604.12212]. For $x\in M_0^m$, $y\in M_0^{m'}$, a finite set $\Phi$ of restricted chronological formulas, and $\varepsilon>0$, the finite-$n$ microstate space is $$ \Gamma^{(n)}(x \mid Y^{(n)} \rightsquigarrow y; \Phi, \varepsilon) := \left\{ X \in \mathbb{M}_n^m : \max_{\varphi \in \Phi} \big| \Lambda_\varphi^{(n)}(X,Y^{(n)}) - \varphi^{M}(x,y) \big| < \varepsilon \right\}. $$ The Gaussian chronological entropy is $$ \tilde{\chi}_{\chron}^{U}(x \mid y) := \sup_{\Phi, \varepsilon} \lim_{n \to U} \Big( -\frac{1}{n^2}\log \sigma^{(n)}(\Gamma^{(n)}(x \mid Y^{(n)} \rightsquigarrow y; \Phi, \varepsilon)) \Big), $$ and the Lebesgue version is $$ \chi_{\chron}^{U}(x \mid y) := m \log(2\pi) + \tfrac{1}{2}\|x\|_2^2 - \tilde{\chi}_{\chron}^{U}(x \mid y). $$ The construction is based on finite-$n$ pointwise functionals $\Lambda_\varphi^{(n)}$ and an ultrafiber quotient $\pi:M\to Q:=M^{/E,V}$ that turns a random matrix ultraproduct into a $\mathrm{II}_1$ factor with filtration and free Brownian motion [2604.12212]. The information-geometric content is expressed on the space of conditional chronological types $$ \mathbb{S}_{\chron,m}(M/\!y) := \{\, \mathrm{tp}_{\chron}^M(x,y): x \in M_0^m \,\}, $$ equipped with the free Wasserstein distance $$ d_W(\mu_0,\mu_1) := \inf\big\{ \|x_0 - x_1\|_2 : \mathrm{tp}_{\chron}(x_j,y)=\mu_j \big\}. $$ Optimal couplings exist, and there is a chronological Monge–Kantorovich duality with convex chronologically definable predicates of the form $\varphi_j=\psi_j+q$, where $q(x)=\frac12\|x\|_2^2$. If $(x_0,x_1)$ is an optimal coupling of $\mu_0,\mu_1$ and $x_t:=(1-t)x_0+t x_1$, then $$ t\mapsto \chi_{\chron}^{U}(\mu_t) \quad\text{is concave on }[0,1]. $$ Thus the new entropy is displacement-concave along free Wasserstein geodesics [2604.12212]. The heat semigroup is defined on chronologically definable predicates by $$ (P_t\varphi)(x,y):=(S_t\varphi)(x+z_{0,t},y). $$ Its evolution satisfies the metric EVI$_0$ inequality $$ \frac{1}{2}\Big[ d_W(\nu_t,\sigma)^2 - d_W(\nu_s,\sigma)^2 \Big] \le (t-s)\Big[ \chi_{\chron}^{U}(\nu_t) - \chi_{\chron}^{U}(\sigma) \Big],\qquad 0\le s\le t, $$ so heat evolution is the Wasserstein gradient flow of $-\chi_{\chron}^{U}$ in the metric sense. The corresponding minimizing-movement scheme is the free JKO iteration $$ \mu_{k+1} \in \operatorname{argmin}_\mu \Big\{ \frac{1}{2\tau} d_W(\mu,\mu_k)^2 - \chi_{\chron}^{U}(\mu)\Big\}. $$ This places the framework squarely inside an information-geometric and optimal-transport paradigm [2604.12212]. The same theory also proves a true chain rule under iterated conditioning, $$ \chi_{\chron}^{U}(x,y \mid w) = \chi_{\chron}^{U}(x \mid y,w) + \chi_{\chron}^{U}(y \mid w), $$ invariance under definable closure of the conditioning variable, and a stochastic-control representation. For $\varphi\in F_{\chron,m+m'}^0$, the ultralimit pressure is $$ \mathcal{P}^{U}(\varphi)(y) := \inf_{\alpha} \bigg[ \varphi^{Q}\Big(\pi(z_1)+\int_0^1 \alpha_t\,dt, y\Big) + \frac{1}{2}\int_0^1 \|\alpha_t\|_2^2\,dt \bigg], $$ with adapted bounded controls $\alpha$. The Gaussian chronological entropy admits the variational formula $$ \tilde{\chi}_{\chron}^{U}(x \mid y) = \sup_{\varphi \in \overline{F}_{\chron,m+m'}^0} \Big[ \mathcal{P}^{U}(\varphi)(y) - \varphi^{Q}(x,y) \Big]. $$ This gives the theory a finite-$n$/control-theoretic bridge analogous, in the paper’s terms, to Borell/Boué–Dupuis and Schrödinger-bridge/Benamou–Brenier structures [2604.12212]. ## 4. Free Fisher information, distributional regularity, and operator-algebraic rigidity A related strand studies the consequences of finite free Fisher information itself. In the tracial, non-microstates setting, if $X=(X_1,\dots,X_n)$ has finite $\Phi^*(X)$ and $P$ is a selfadjoint noncommutative polynomial of degree $d\ge 1$, then the cumulative distribution function $F_{P(X)}$ is Hölder continuous with explicit exponent $$ \alpha(d)=\frac{2}{3(2^d-1)}. $$ If $X$ admits Lipschitz conjugate variables, the exponent improves to $$ \alpha(d)=\frac{1}{2^d-1}. $$ For linear polynomials, the exponent $2/3$ under finite $\Phi^*(X)$ is optimal, while Lipschitz conjugate variables yield Lipschitz continuity and hence absolute continuity with bounded density. These regularity estimates imply finite logarithmic energy and therefore finite non-microstates free entropy $\chi^*(P(X))>-\infty$ for every selfadjoint nonconstant polynomial $P$, partially resolving a conjecture of Charlesworth–Shlyakhtenko under the stronger assumption $\Phi^*(X)<\infty$ [1809.11153]. The same work supplies an explicit route from weak convergence to rates in Kolmogorov distance. If the limiting law has Hölder CDF with exponent $\beta$ and the Cauchy transforms satisfy suitable strip estimates, then $$ d_K(\mu_n,\nu)\le D\,\varepsilon_n^{\frac{\beta}{2+k+(2-\beta)l}}, $$ with a compact-support refinement $$ d_K(\mu_n,\nu)\le D\,\varepsilon_n^{\frac{\beta}{k+\beta}}. $$ Applications include convergence in Kolmogorov distance for polynomial eigenvalue laws of Gibbs ensembles and explicit GUE rates such as $$ d_K(\overline{\mu}_{X^{(N)}},\mu_S)\le D\,N^{-4/35} $$ in the block-GUE semi-flat case, and $$ d_K(\overline{\mu}_{P(X^{(N)})},\mu_{P(S)})\le D\,N^{-\frac{1}{13\cdot 2^{d+2}-60}} $$ for selfadjoint polynomial GUE models [1809.11153]. In the non-tracial setting, finite free Fisher information for eigenvectors of a modular operator has much stronger structural consequences. If $M$ is generated by a finite selfadjoint set $G=G^*$ of eigenoperators of $\sigma^\varphi$ with finite free Fisher information, then $$ (M^\varphi)'\cap M=\mathbb{C}. $$ In particular, $M^\varphi$ is a $\mathrm{II}_1$ factor, and if $H<\mathbb{R}_+^\times$ is the closed subgroup generated by the eigenvalues of $G$, then $M$ is a factor of type $\mathrm{III}_1$, $\mathrm{III}_\lambda$ ($0<\lambda<1$), or $\mathrm{II}_1$ according as $H=\mathbb{R}_+^\times$, $H=\lambda^\mathbb{Z}$, or $H=\{1\}$. The same hypotheses imply that $M^\varphi$ does not have property $\Gamma$, and if $M$ is type $\mathrm{III}_\lambda$ with $0<\lambda<1$, then $M$ is full [1604.03900]. These results rely on $\mu$-modular derivations, conjugate variables entire for the modular group, Dirichlet forms on the centralizer, contraction resolvents, and non-tracial $L^2$-homology estimates. In this line of work, finite free Fisher information is not merely a regularity parameter; it becomes a rigidity hypothesis controlling diffuseness, factoriality, and type classification [1604.03900]. ## 5. Rectangular finite free probability and singular-value information Another formulation replaces eigenvalues by singular values of rectangular matrices. For $A\in\mathbb{R}^{m\times d}$ with $m\ge d$, the basic polynomial is $$ p(x,y)=\det\!\begin{bmatrix} yI_m & A \\ A^T & xI_d \end{bmatrix} = y^{m-d}\det(xyI_d-A^TA). $$ Equivalently, if $p$ is a univariate polynomial with nonnegative roots, its rectangular polynomial extension of order $m-d$ is $p(x,y)=y^{m-d}p(xy)$. This encodes the squared singular values of $A$ directly at finite dimension [2208.09768]. The finite rectangular additive convolution is defined by averaging over bi-orthogonal rotations: $$ [p \boxplus_{d,\lambda} q](x,y) := \iint \det\!\begin{bmatrix} yI_m & A+QBR\\ (A+QBR)^T & xI_d \end{bmatrix} \,dQ\,dR, $$ where $Q\in O_m$, $R\in O_d$ are Haar and $\lambda=d/m$. Setting $y=1$ yields a univariate real-rooted polynomial with nonnegative roots. The operation is bilinear, associative, and preserves real-rootedness. The theory also gives an explicit coefficient formula and a differential-operator expression for the convolution [2208.09768]. This program introduces finite analogues of classical rectangular free-probability transforms. A $d$-point random variable $T^{(m,d)}_{Sp}$ is associated to a polynomial $p$ so that its power sums are fixed linear functionals of the coefficients, and the finite rectangular $R$-transform is defined by $$ R^{(\operatorname{finite})}_{d,\lambda;Sp}(s) = \log E\!\left[e^{-T^{(m,d)}_{Sp}s^{md}}\right] \mod[s^{d+1}]. $$ It linearizes finite rectangular additive convolution: $$ R^{(\operatorname{finite})}_{d,\lambda;S[p\boxplus_{d,\lambda} q]}(s) = R^{(\operatorname{finite})}_{d,\lambda;Sp}(s) + R^{(\operatorname{finite})}_{d,\lambda;Sq}(s). $$ A modified finite $R$-transform converges, as $d\to\infty$, to the classical rectangular free $R$-transform, so the finite theory converges to asymptotic rectangular free probability [2208.09768]. The same framework produces explicit finite-dimensional LLN and CLT analogues for polynomials. After $1/N$ rescaling, iterated finite rectangular convolution converges to the zero polynomial, while after $1/\sqrt N$ rescaling it converges, up to scaling, to a generalized Laguerre polynomial. In transform terms, $$ T^{(m,d)}_{Sp}\equiv \sigma^2 \iff R^{(\operatorname{finite})}_{d,\lambda;Sp}(s)=m\sigma^2 s \iff p\sim L_d^{(m-d)}(\sigma^2 x). $$ This identifies generalized Laguerre polynomials as the Gaussian analogues of the theory [2208.09768]. Because the roots of the convolved polynomial approximate squared singular values, the construction interfaces directly with spectral information functionals. For a nonnegative spectral law $F$, the Shannon transform is $$ S_F(\gamma)=\int \log(1+\gamma x)\,dF(x), $$ and for a $d\times d$ positive matrix it is $(1/d)\sum_i \log(1+\gamma\lambda_i)$. In the finite rectangular framework, one computes $[p\boxplus_{d,\lambda} q](x,1)$, extracts its nonnegative roots $\{r_k\}$, and evaluates $$ S_F(\gamma)=\frac{1}{d}\sum_{k=1}^d \log(1+\gamma r_k). $$ The paper explicitly notes that a finite multiplicative convolution or finite $S$-transform is not introduced there, so multiplicative problems remain asymptotic in the current rectangular theory [2208.09768]. ## 6. Physical and finite-horizon extensions In free scalar quantum field theory at finite temperature, Finite Free Information Theory refers to the mutual information across a spatial bipartition. For a free real scalar in $3+1$ dimensions, the mutual information between a region $A$ and its complement $B$ satisfies an area law $$ I(A:B)\simeq \kappa(\mu,T)\,\operatorname{Area}(\partial A)+\text{subleading terms}, $$ with $\kappa(\mu,T)$ computable in an inverse-mass expansion. The high-temperature limit is finite and purely classical in origin: $$ \lim_{T\to\infty} I(A:B)\neq 0. $$ For general coupled harmonic systems, the high-$T$ expansion has no $1/T^2$ term, and in the field-theory setting this cancellation is recovered once a uniform angular cutoff is imposed. The low-temperature correction is exponentially small, proportional to $e^{-m_{\operatorname{eff}}/T}$, while the $T\to\infty$ limit matches the classical thermal calculation [1907.08508]. A complementary numerical study of free scalar theory on concentric spherical shells established the same area-law behavior, reporting in the massless $3$-dimensional spherical setup $$ I \approx 0.59\, (R^2/a^2)\quad \text{at }T=0, \qquad I \approx 0.38\, (R^2/a^2)\quad \text{as }T\to\infty, $$ thereby exhibiting the finite classical remnant directly in the lattice computation [1907.04817]. A conceptually different finite-horizon usage defines information as deviation from a constrained Maximum-Caliber ensemble. For a one-step transition network, if $\mu$ is the MaxCal input marginal $$ \mu(x^t)=e^{h(x^t)}/\kappa, $$ then information is $$ \psi(\rho\times P)=\mathcal{D}(\rho\|\mu). $$ For longer horizons, large-deviation theory yields $$ \psi_{\operatorname{ldp}}(\nu)=\mathcal{D}(\rho\|\pi)+T\,\mathcal{D}(\nu\|\rho\times P). $$ Within this framework, Integrated Information Theory repertoires are re-derived from constrained MaxEnt/MaxCal posteriors, and partition-based integration is defined by $$ \phi^{\mathcal P}(\rho)=\psi(\rho)-\psi^{\mathcal P}(\rho), \qquad \phi(\rho)=\min_{\mathcal P}\phi^{\mathcal P}(\rho). $$ The same paper states a duality to active-inference free-energy functionals and gives CLT- and LDP-based predictive-coding reductions in which $\psi$ becomes a Bayesian-surprise or transition-accuracy term [2605.12536]. The open-problem landscape is correspondingly plural. In the chronological microstate program, open directions include ultrafilter independence of $\mathcal{P}^U(\varphi)$, a chronological free Fisher functional, and EVI$_\lambda$ for free Ornstein–Uhlenbeck semigroups [2604.12212]. In the polynomial program, the main unresolved issues are proof of the dyadic singular-value conjecture for $E_n|_W$, proof of the $p$-Stam inequality for all $p\in(1,2]$, uniqueness of the Hermite equality case at $p=2$, and characterization of the $p<2$ bifurcating extremizers [2604.11922]. In the rectangular program, the missing finite multiplicative analogue remains explicit [2208.09768]. In the non-tracial Fisher-information line, an important question is how much of the factoriality and fullness theory survives without an eigenoperator generating set [1604.03900]. In the MaxCal/IIT/FEP program, open questions include explicit non-equilibrium current formulations, non-Gaussian regimes beyond the CLT, and empirical validation relating $\psi$, $\phi$, and PCI [2605.12536]. Taken together, these programs show that Finite Free Information Theory currently names a family of finite analogues rather than a single invariant package. What unifies them is a shared strategy: replace asymptotic free-information objects by exact finite structures—zeros, microstate sets, singular-value polynomials, path ensembles, or finite-temperature Gaussian subsystems—and recover entropy, Fisher information, transport, or mutual-information phenomena before passing to a limit. Source: https://www.emergentmind.com/topics/finite-free-information-theory