Quantum Exponential Families

This section is under development.

Mathematical Framework

A quantum exponential family is defined by:

\[\rho(\theta) = \exp\left(\sum_a \theta_a F_a - \psi(\theta)\right)\]

where:

  • \(\rho(\theta)\) is a density matrix (positive semidefinite, trace 1)

  • \(F_a\) are Hermitian operators (generators)

  • \(\theta = (\theta_1, \ldots, \theta_n)\) are natural parameters

  • \(\psi(\theta) = \log \text{Tr}[\exp(\sum_a \theta_a F_a)]\) is the log-partition function

Properties

  • Convexity: \(\psi(\theta)\) is strictly convex

  • Duality: Expectation parameters \(\eta_a = \langle F_a \rangle\)

  • Fisher metric: \(G_{ab} = \text{Cov}(F_a, F_b)\) where covariance uses BKM inner product

Why Duhamel Integrals Appear

If you are used to classical exponential families, the appearance of operator-valued Duhamel integrals in the quantum setting can seem mysterious. In the classical case, sufficient statistics \(T_i(x)\) commute with each other, and

\[p_\theta(x) = \exp\Bigl(\sum_i \theta_i T_i(x) - \psi(\theta)\Bigr)\]

leads directly to

\[\frac{\partial}{\partial \theta_i} p_\theta(x) = \bigl(T_i(x) - \mathbb{E}_\theta[T_i]\bigr)\,p_\theta(x),\]

so derivatives of the log-partition function and the Fisher metric can be expressed using ordinary covariances.

In the quantum case, the sufficient statistics are Hermitian operators \(F_i\) and in general do not commute with the Hamiltonian \(K(\theta) = \sum_i \theta_i F_i\). Differentiating the matrix exponential \(\exp(K(\theta))\) therefore produces the Wilcox/Duhamel formula

\[\frac{\partial}{\partial \theta_i} e^{K(\theta)} = \int_0^1 e^{(1-s)K(\theta)} F_i e^{sK(\theta)} \,\mathrm{d}s,\]

and, after centering \(F_i\) and normalising, the derivative of the density matrix

\[\partial_i \rho(\theta) = \int_0^1 \rho(\theta)^{1-s}\,\bigl(F_i - \mu_i(\theta)\bigr)\, \rho(\theta)^{s}\,\mathrm{d}s\]

is an operator-ordered integral rather than a simple product. This is the origin of the Kubo–Mori / BKM metric and higher cumulants: the inner products and covariances in quantum information geometry are defined with respect to this non-commutative kernel, not the classical pointwise product.

What Is Special About Our Geometry?

Two structural choices make the Duhamel machinery both tractable and geometrically natural in this project:

  • Lie-closed operator bases: we choose \(\{F_a\}\) to form a Lie algebra \([F_a, F_b] = i \sum_c f_{abc} F_c\). Then the Heisenberg-evolved operators \(e^{-sK} F_i e^{sK}\) stay in the linear span of the \(F_a\), so the Duhamel kernel becomes a finite-dimensional linear operator \(K_\rho = f(\mathrm{ad}_H)\) on this Lie algebra (with \(f(z) = (e^z - 1)/z\)). In other words, the Duhamel integral does not disappear, but it is encoded as a matrix function of the adjoint representation rather than an intractable operator integral.

  • Categorical forcing of unitarity: using the categorical framework of Parzygnat and the GENERIC-like decomposition, we know a priori that the entropy-conserving (antisymmetric) sector of the flow must be implemented by unitary conjugation, hence has von Neumann form \(\dot{\rho}_{\mathrm{rev}} = -i[H_{\mathrm{eff}}, \rho]\). The Lie-closed exponential-family coordinates then provide a concrete way to express the effective Hamiltonian \(H_{\mathrm{eff}}\) in terms of the antisymmetric tensor \(A_{ab}\) and the structure constants \(f_{abc}\), with the Duhamel/BKM kernel absorbed into the finite-dimensional map that relates \(A\) to the Hamiltonian coefficients \(\eta_a(\theta)\).

Compared to standard presentations of quantum information geometry—which often start from arbitrary density matrices and modular theory—our framework keeps the exponential-family viewpoint in the foreground. This makes the role of natural parameters, the BKM metric, and the Duhamel kernel transparent, and it explains why Lie-closed coordinates are particularly well adapted to the categorical/unitary structure of the quantum inaccessible game.

See Also