Quantum Donsker–Varadhan Representation

• Quantum Donsker–Varadhan Representation is a quantum analog of classical variational formulas that replaces complex entropy or free-energy calculations with optimizations over tailored auxiliary quantum objects.
• It is applied in diverse areas such as approximate von Neumann entropy estimation, Born-rule generative modeling of measurement distributions, and mean-field large-deviation analysis in spin systems.
• By constraining the optimization domain (through rank limitations, scalar critics, or state-space duality), the framework provides practical approximations with provable error bounds and convergence guarantees.

Searching arXiv for papers on "Quantum Donsker-Varadhan Representation" and closely related "quantum Varadhan theorem" to ground the article in cited sources. Quantum Donsker–Varadhan representation denotes a family of variational constructions that import the classical Donsker–Varadhan or Varadhan philosophy into noncommutative settings. In recent arXiv literature, the expression is used in three distinct senses: as an approximate variational characterization of von Neumann entropy over rank-constrained density operators for quantum neural estimation, as a classical Donsker–Varadhan formula applied to Born-rule measurement distributions in quantum generative modeling, and as an equilibrium quantum Varadhan theorem for mean-field spin systems with a state-space duality involving quantum relative entropy (Shin et al., 2023, Wilde, 2 Dec 2025, Dorlas, 3 Nov 2025). The shared structural idea is a replacement of a difficult information-theoretic or free-energy quantity by a supremum or infimum over auxiliary test objects—scalar critics, Hermitian operators, or quantum states—together with an entropy or pressure penalty.

1. Terminological scope and classical antecedent

The classical antecedent is the Donsker–Varadhan variational formula for Kullback–Leibler divergence,

$$D_{\mathrm{KL}}(p\|q)=\sup_{T:\Omega\to\mathbb R}\left\{\mathbb E_p[T]-\log\left(\mathbb E_q[e^T]\right)\right\},$$

which underlies MINE when $p=p_{XY}$ and $q=p_Xp_Y$, so that the divergence becomes mutual information (Shin et al., 2023). Recent quantum work preserves the same variational architecture while changing the optimization domain and the object being represented.

Usage in the literature	Variational target	Optimization variable
QDVR for entropy estimation	von Neumann entropy $S(\rho)$	Hermitian operators or rank-$r$ density matrices
Born-rule DV training	classical $D(p\|q_\gamma)$	scalar critic $T$ or $T_w$
Quantum Varadhan theorem	limiting pressure/free-energy density	scalar means $u$ or states $\varphi$

In the entropy-estimation line, the representation is explicitly presented as a “quantum analog of the classical Donsker-Varadhan representation,” but the represented quantity is von Neumann entropy rather than quantum relative entropy (Shin et al., 2023). In the Born-rule generative-modeling line, the paper explicitly states that the construction is not a genuinely new quantum information quantity: the variational formula remains classical and is applied after measurement to the induced output law $p=p_{XY}$0 (Wilde, 2 Dec 2025). In the large-deviation line, the closest analogue is a quantum Varadhan theorem or quantum Varadhan lemma for mean-field quantum spin systems, with a variational representation over states involving quantum relative entropy (Dorlas, 3 Nov 2025).

2. Entropy-based QDVR in quantum information

The formal starting point for the entropy-estimation literature is the Gibbs variational principle. For a density matrix $p=p_{XY}$1, define

$p=p_{XY}$2

The exact statement is

$p=p_{XY}$3

with the infimum taken over Hermitian $p=p_{XY}$4 (Shin et al., 2023). This is then reduced in stages. First, shift invariance holds: $p=p_{XY}$5 Consequently, the infimum may be restricted to positive Hermitian matrices without changing the value (Shin et al., 2023).

The representation called the quantum Donsker–Varadhan representation in this line of work is a reduced-domain approximation. If $p=p_{XY}$6 is an $p=p_{XY}$7-rank density matrix, define

$p=p_{XY}$8

For

$p=p_{XY}$9

the paper states

$q=p_Xp_Y$0

where the infimum is taken over $q=p_Xp_Y$1-dimensional rank-$q=p_Xp_Y$2 density matrices $q=p_Xp_Y$3 (Shin et al., 2023). The uncommon-information paper states the same approximation in $q=p_Xp_Y$4-dimensional form as

$q=p_Xp_Y$5

provided $q=p_Xp_Y$6 is an $q=p_Xp_Y$7-rank density matrix and

$q=p_Xp_Y$8

(Ji et al., 8 Jul 2025).

This formulation is not merely a restatement of the Gibbs principle. Its distinctive move is to replace the unrestricted Hermitian optimization by an optimization over rank-constrained density matrices. In the appendix of the QMINE paper, the approximating operator is constructed from an eigendecomposition

$q=p_Xp_Y$9

by choosing

$S(\rho)$0

with $S(\rho)$1, thereby yielding both the approximation guarantee and the trace bound (Shin et al., 2023). This places the construction squarely in the category of approximate variational entropy representations rather than exact operator-level Donsker–Varadhan theorems.

3. Quantum neural parameterizations and information-theoretic applications

The reduced-domain QDVR becomes algorithmically useful after parameterizing the optimizing density matrix. In the uncommon-information paper, for a rank-$S(\rho)$2 state $S(\rho)$3, one introduces a parameterized unitary $S(\rho)$4 and defines

$S(\rho)$5

with $S(\rho)$6 and $S(\rho)$7 (Ji et al., 8 Jul 2025). Substituting this into the QDVR objective gives

$S(\rho)$8

The first term is evaluated on a quantum computer, the second classically, gradients with respect to $S(\rho)$9 are computed classically, gradients with respect to $r$0 are evaluated using the parameter-shift rule, and optimization is performed with hybrid quantum-classical gradient descent (Ji et al., 8 Jul 2025).

The QMINE paper gives the same architecture in the setting of entropy and quantum mutual information estimation. Writing

$r$1

it obtains

$r$2

and after basis rotation by a parameterized circuit $r$3,

$r$4

(Shin et al., 2023). The paper uses

$r$5

and enforces the simplex constraint on the $r$6 by the angular parametrization

$r$7

For the circuit parameters, the paper states the parameter-shift rule

$r$8

and for the spectral parameters,

$r$9

(Shin et al., 2023).

These entropy estimators are then composed into estimators for larger information-theoretic quantities. The QMINE work estimates quantum mutual information through

$D(p\|q_\gamma)$0

and also via the equivalent identity

$D(p\|q_\gamma)$1

(Shin et al., 2023). The uncommon-information work uses QDVR to estimate entropies appearing in the bounds

$D(p\|q_\gamma)$2

with

$D(p\|q_\gamma)$3

and

$D(p\|q_\gamma)$4

(Ji et al., 8 Jul 2025).

The empirical evidence reported in these papers is modest-scale but specific. The QMINE paper reports QMI error rates between $D(p\|q_\gamma)$5 and $D(p\|q_\gamma)$6 for random density matrices on $D(p\|q_\gamma)$7 qubits, and its rank-analysis simulations support the claim that choosing the parameter rank $D(p\|q_\gamma)$8 is sufficient and empirically best (Shin et al., 2023). The uncommon-information paper reports stabilization after about $D(p\|q_\gamma)$9 optimization steps for the loose upper bound $T$0 on $T$1 qubits, around $T$2 steps for $T$3- and $T$4-qubit systems, roughly $T$5 steps for the loose lower bound $T$6, about $T$7 steps for the tight upper bound $T$8, and roughly $T$9 steps for the tight lower bound $T_w$0 in the special decomposition setting it studies (Ji et al., 8 Jul 2025).

4. Born-rule generative modeling and the measurement-level DV reformulation

A different use of Donsker–Varadhan appears in Born-rule generative modeling. The target is a classical distribution $T_w$1 over a finite alphabet $T_w$2, approximated by a model distribution

$T_w$3

where $T_w$4 is a POVM and $T_w$5 is a parameterized quantum state (Wilde, 2 Dec 2025). The training objective is the classical relative entropy

$T_w$6

The paper invokes the classical Donsker–Varadhan representation

$T_w$7

and also the sampling-friendly variant

$T_w$8

where $T_w$9 is the class of all functions $u$0 (Wilde, 2 Dec 2025). This yields the minimax training problem

$u$1

The paper emphasizes three reasons for this reformulation: $u$2 can be estimated unbiasedly from data samples from $u$3, $u$4 can be estimated unbiasedly from samples from the quantum model $u$5, and the objective is concave in the critic function $u$6 (Wilde, 2 Dec 2025).

The quantum generator in the main result is the evolved quantum Boltzmann machine,

$u$7

with

$u$8

(Wilde, 2 Dec 2025). The critic term can be rewritten as a quantum expectation value through

$u$9

so that the generator-side gradient reduces to differentiating an observable expectation under the EQBM ansatz (Wilde, 2 Dec 2025).

The paper gives explicit derivatives with respect to the quantum parameters and the critic parameters, and uses the evolved quantum Boltzmann gradient estimator to construct unbiased estimators for the required quantum quantities (Wilde, 2 Dec 2025). It then proposes four hybrid quantum-classical minimax algorithms: Extragradient, Two-timescale gradient descent-ascent, Follow-the-ridge, and HessianFR. The convergence statements are differentiated carefully: Extragradient and two-timescale GDA have no general guarantee to local minimax points in fully nonconvex-nonconcave settings, whereas Follow-the-ridge and HessianFR are stated to have provable convergence to a local minimax point through the cited results of Wang et al. and Gao et al. (Wilde, 2 Dec 2025).

The crucial conceptual point is explicit in the paper: no genuinely new quantum DV principle is introduced. The formula is classical, the critic is classical, and the “quantum” aspect enters through the generator family, the POVM, and the quantum gradient-estimation machinery (Wilde, 2 Dec 2025).

5. Quantum Varadhan theorem and entropy-pressure duality

A genuinely noncommutative analogue of the classical Varadhan/Donsker–Varadhan variational structure appears in the mean-field large-deviation setting. For a finite-dimensional single-site algebra $\varphi$0, self-adjoint matrices $\varphi$1, and empirical average

$\varphi$2

the one-variable theorem states

$\varphi$3

where $\varphi$4 is continuous and $\varphi$5 is the Legendre transform of

$\varphi$6

(Dorlas, 3 Nov 2025). Structurally, this is a Laplace principle for traces of exponentials of empirical observables.

The same quantity admits a state-space formulation. If

$\varphi$7

is the Gibbs state on $\varphi$8, then

$\varphi$9

and consequently

$p=p_{XY}$00

(Dorlas, 3 Nov 2025). This is the closest direct analogue of a Donsker–Varadhan formula in the paper: the optimizer is a quantum state, and the penalty is quantum relative entropy.

The multivariable extension covers self-adjoint $p=p_{XY}$01, continuous $p=p_{XY}$02, and the noncommutative cumulant generating function

$p=p_{XY}$03

with Legendre transform

$p=p_{XY}$04

The theorem then states

$p=p_{XY}$05

(Dorlas, 3 Nov 2025). The observables $p=p_{XY}$06 need not commute, so the rate function is not a classical joint large-deviation rate from a joint spectral measure; it is generated by the noncommutative pressure $p=p_{XY}$07.

The proof strategy is also explicitly different from the original Petz–Raggio–Verbeure argument. The paper constructs a complex-valued path measure $p=p_{XY}$08 on Skorokhod space, uses the Trotter product formula to obtain a Feynman–Kac-type path representation for traces, derives the upper bound through a large-deviation argument, and obtains the lower bound via entropy duality and product states (Dorlas, 3 Nov 2025). The resulting theorem is best characterized as a quantum Varadhan theorem or entropy-pressure variational principle for mean-field quantum spin systems, rather than a theory of quantum Markov occupation measures.

6. Conceptual distinctions, scope conditions, and recurrent misconceptions

A central misconception is that “Quantum Donsker–Varadhan representation” denotes a single standardized theorem. The recent literature instead uses the phrase for at least three non-equivalent constructions. In the QMINE line, it is an approximate variational representation of von Neumann entropy over rank-$p=p_{XY}$09 density matrices, derived from the Gibbs variational principle and used as a trainable loss for entropy and QMI estimation (Shin et al., 2023). In the Born-rule generative-modeling line, the paper explicitly states that the DV ingredient remains classical and is applied to the measured distribution $p=p_{XY}$10; the construction is quantum-adapted, but not an operator-level quantum DV theorem (Wilde, 2 Dec 2025). In the mean-field large-deviation line, the representation is genuinely noncommutative, but it is an equilibrium pressure theorem over finite-dimensional spin systems, not a Markov-generator or occupation-measure theory (Dorlas, 3 Nov 2025).

The scope conditions are correspondingly different. The entropy-estimation QDVR assumes rank information for $p=p_{XY}$11, depends on a scale parameter $p=p_{XY}$12, and yields an $p=p_{XY}$13-approximation rather than an exact reduced-domain identity (Shin et al., 2023). The uncommon-information application further requires structural inputs such as a common subspace $p=p_{XY}$14 for the upper bound or a decomposition $p=p_{XY}$15 for the lower bound; for generic tripartite states, the paper notes that no known general method produces the required decomposition (Ji et al., 8 Jul 2025). The QMINE paper also notes barren plateau issues, nonconvex optimization, and the absence of a full end-to-end complexity theorem proving its suggested low-copy scaling (Shin et al., 2023).

The Born-rule generative-modeling formulation has its own caveats. The generator side is generally nonconvex, a generic neural critic makes the minimax game nonconvex-nonconcave, convergence to local minimax points depends on the chosen algorithm and assumptions, practical performance remains to be tested numerically, and a full statistical analysis of approximation, estimation, and optimization error is left open (Wilde, 2 Dec 2025). The mean-field large-deviation theorem is restricted to finite-dimensional local algebras $p=p_{XY}$16, tensor-product spin systems, and empirical-average or symmetric-polynomial mean-field interactions; it does not address continuous-time quantum Markov semigroups or general noncommutative large deviations in infinite-dimensional von Neumann algebras (Dorlas, 3 Nov 2025).

Taken together, these usages show that the phrase identifies a shared variational strategy rather than a unique formal object. The common template is an entropy- or pressure-penalized extremization that replaces a difficult quantity by an optimization over auxiliary test structures. What varies from paper to paper is the represented quantity—von Neumann entropy, classical relative entropy after measurement, or mean-field limiting pressure—the optimization domain, and the precise sense in which the construction is genuinely quantum.