Von Neumann Entropy Estimation

• Von Neumann entropy is a spectral measure that quantifies the information content of a quantum state via its eigenvalue distribution.
• Techniques range from full state tomography and statistical regularization to observer-based dynamic tracking for accurate entropy estimation.
• Recent variational and sublinear quantum query approaches provide precise entropy estimates under model-specific and operational constraints.

Von Neumann entropy estimation is the problem of inferring the spectral functional

$S(\rho)=-\operatorname{Tr}(\rho\log\rho)$

for an unknown quantum state $\rho$, or of estimating closely related quantities such as relative entropy, entanglement entropy, and model-specific entropy densities. The difficulty is intrinsic: $S(\rho)$ depends nonlinearly on the eigenvalues of $\rho$, so it is not directly accessible from a single expectation value and, in general, requires either substantial state information, a structural prior, or a specialized operational framework. The recent literature suggests several distinct regimes: state-reconstruction-based estimation, observer-based online estimation, variational and sublinear quantum algorithms, analytic estimation under ensemble or model assumptions, and domain-specific surrogates in graphs, optics, and many-body systems (Balas et al., 2023, Gur et al., 2021).

1. Formulation of the estimation problem

For a density operator $\rho\in\mathcal S\subset\mathbb C^{d\times d}$, the spectral decomposition

$$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$

reduces von Neumann entropy to the Shannon entropy of the eigenvalue distribution,

$S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$

Accordingly, $S(\rho)=0$ iff $\rho$ is pure, and $0\le S(\rho)\le \log d$. In estimation problems, this basic reduction is simultaneously useful and obstructive: it shows that entropy is purely spectral, but it also shows that any method must recover, approximate, or bypass the eigenvalue distribution itself (Balas et al., 2023).

The literature treats several target tasks. One is direct estimation of $\rho$0 for a fixed but unknown state. Another is estimation of $\rho$1, where support conditions and possible divergence to $\rho$2 make the problem more delicate. A third is multiplicative approximation, where one seeks $\rho$3 such that $\rho$4 for some $\rho$5. In the purified quantum query access model, such multiplicative approximation is possible in sublinear query complexity only when the entropy is bounded away from zero; the paper proves that no polynomial query algorithm can multiplicatively approximate the entropy of distributions with arbitrarily low entropy (Gur et al., 2021).

A different regime arises when the state is not arbitrary but drawn from an ensemble or constrained by a model. In such cases one may estimate typical entropies, fluctuations, or asymmetry of the entropy distribution from structural parameters alone, such as subsystem dimensions, Rényi data, or one-particle symbols. This suggests that “von Neumann entropy estimation” is not a single protocol class but a family of inference problems distinguished by access model, prior structure, and error criterion.

2. Reconstruction, regularization, and observer-based estimation

The most direct route is full quantum state tomography: measure an informationally complete POVM, reconstruct $\rho$6, diagonalize it, and compute $\rho$7. This is conceptually straightforward but measurement- and computation-intensive, scales badly with dimension, and is usually offline. In the comparison drawn in the observer-based work, the general overhead is described as $\rho$8 and therefore poorly suited to continuous dynamical tracking (Balas et al., 2023).

A statistical alternative is to estimate the density matrix itself under structural regularization. In the regression model

$\rho$9

the entropy-penalized estimator is

$S(\rho)$0

This is a convex program over density matrices, and the analysis yields oracle inequalities in $S(\rho)$1, Kullback–Leibler, Hellinger, and trace norms. The method is framed as low-rank matrix estimation on the state space, with guarantees that adapt to low-rank structure while remaining compatible with density-matrix constraints (Koltchinskii, 2010).

A more dynamical formulation replaces offline reconstruction by a quantum state observer. For a finite-dimensional closed system evolving under

$S(\rho)$2

with measurement outputs $S(\rho)$3, vectorization yields a linear system $S(\rho)$4. If $S(\rho)$5 is linearly observable, then a Luenberger-type observer

$S(\rho)$6

can be designed so that $S(\rho)$7. Because the linear observer need not preserve positivity or unit trace, the estimate is projected onto the convex state space $S(\rho)$8 by a Hilbert metric projection, which is non-expansive. The entropy estimate

$S(\rho)$9

is then shown to converge essentially exponentially to $\rho$0. Moreover, if $\rho$1 is full-rank, the relative entropy $\rho$2 converges essentially exponentially to zero. The same framework proves a minimal measurement requirement: if the closed quantum system is linearly observable, then $\rho$3 (Balas et al., 2023).

3. Direct quantum algorithms

Variational quantum algorithms attempt to estimate entropy without reconstructing the full state. In the quantum neural framework, the starting point is the variational identity

$\rho$4

together with a variational formula for measured relative entropy. A Hermitian operator $\rho$5 is parameterized by a parameterized quantum circuit $\rho$6 and a classical neural network $\rho$7, with the circuit specifying an eigenbasis and the neural network specifying eigenvalues. Measurement statistics after applying $\rho$8 to copies of $\rho$9 convert the variational objective into an empirically optimizable functional. The same architecture extends to von Neumann entropy, Rényi entropy, measured relative entropy, and measured Rényi relative entropy, and the reported noiseless simulations show accurate estimates for the examples tested (Goldfeld et al., 2023).

A distinct line pursues sublinear quantum query complexity. In the purified quantum query access model, there is a quantum algorithm that outputs a $\rho\in\mathcal S\subset\mathbb C^{d\times d}$0-multiplicative approximation of $\rho\in\mathcal S\subset\mathbb C^{d\times d}$1 using

$\rho\in\mathcal S\subset\mathbb C^{d\times d}$2

queries, provided

$\rho\in\mathcal S\subset\mathbb C^{d\times d}$3

The algorithm combines projected unitary encodings, quantum singular value estimation, quantum singular value transformation, and quantum amplitude estimation. Technically, it treats the eigenvalue distribution by splitting it into “heavy” and “light” parts, approximating $\rho\in\mathcal S\subset\mathbb C^{d\times d}$4 on the heavy region by power functions $\rho\in\mathcal S\subset\mathbb C^{d\times d}$5, and estimating the low-eigenvalue mass separately. The same work proves lower bounds

$\rho\in\mathcal S\subset\mathbb C^{d\times d}$6

for $\rho\in\mathcal S\subset\mathbb C^{d\times d}$7-multiplicative estimation of Shannon and von Neumann entropies, and proves that no polynomial query algorithm can multiplicatively approximate entropy when it can be arbitrarily close to zero (Gur et al., 2021).

These two algorithmic traditions are complementary. The variational method is heuristic, model-flexible, and directly tied to circuit ansätze and optimization landscapes; the sublinear-query method is asymptotic and complexity-theoretic, with explicit guarantees but a more restrictive oracle model. Together they show that direct entropy estimation can be framed either as variational optimization or as spectral query estimation, rather than only as tomography.

4. Analytic estimation under ensemble or model assumptions

When the state is treated as random or constrained by an exactly solvable model, entropy estimation becomes an analytic problem about distributions of eigenvalues. In Page’s bipartite model for Haar-random pure states on $\rho\in\mathcal S\subset\mathbb C^{d\times d}$8 with $\rho\in\mathcal S\subset\mathbb C^{d\times d}$9, the reduced density matrix eigenvalues follow the fixed-trace ensemble and the entanglement entropy

$$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$0

is a random variable on $$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$1. Exact formulas are available for the first three cumulants. The first cumulant is Page’s formula,

$$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$2

the second cumulant gives the variance, and the third cumulant gives the skewness. In the high-dimensional regime $$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$3 with $$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$4, the variance scales as $$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$5, the third cumulant as $$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$6, and the standardized skewness as $$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$7, so the entropy distribution becomes sharply concentrated and nearly Gaussian. For finite $$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$8, a Gram–Charlier type-A expansion using the first three cumulants improves the approximation of the full entropy distribution, especially for the left-skewed small-dimensional regime (Wei, 2019).

For Bures–Hall random states the same program has recently been extended. Closed-form expressions are available for the mean, variance, and third cumulant of the von Neumann entropy, again in terms of polygamma functions. The third cumulant produces a skewness-corrected approximation to the entropy distribution, and the asymptotics again give $$\rho=Q\Lambda Q^\dagger,\qquad \Lambda=\operatorname{diag}(\lambda_1,\dots,\lambda_d)$$9, $S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$0, and standardized skewness $S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$1. The analysis depends on a Pfaffian point-process representation, Meijer-$S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$2 kernel integrals, and a large collection of summation identities for anomalous polygamma sums (Wei et al., 7 Jun 2025).

A different model-specific route appears in shift-invariant quasi-free Fermionic lattice systems. There the entropy density $S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$3 is connected to integer-order Rényi entropy densities $S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$4 through a Laplace-transform representation

$S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$5

The paper proves that the mean von Neumann entropy density is completely determined by the set of integer-order mean Rényi entropies, and constructs explicit linear combinations of $S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$6 with uniform error bounds. This gives a rigorous reconstruction scheme for von Neumann entropy from replica-accessible Rényi data in a quasi-free Fermionic setting (Fannes et al., 2012).

5. Specialized estimators and surrogates

Some estimation problems admit exact or controlled surrogates because the state space is strongly constrained. For a pure entangled state of two two-level atoms, the reduced-state von Neumann entropy is determined exactly by the magnitude of the mean spin vector of a single atom. If $S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$7 denotes that magnitude, then the reduced density matrix eigenvalues are

$S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$8

and the entropy is

$S(\rho)=-\sum_{k=1}^d \lambda_k\log\lambda_k.$9

This yields an exact local observable proxy for entanglement entropy in that two-qubit pure-state setting (Deb, 2020).

In optical polarization, the state is effectively a $S(\rho)=0$0 coherency matrix $S(\rho)=0$1, reconstructed from spatially resolved Stokes parameters measured with an EMCCD array detector: $S(\rho)=0$2 The entropy

$S(\rho)=0$3

is then computed per pixel from the two eigenvalues of $S(\rho)=0$4, producing entropy maps across the beam profile. The method was applied to a diode laser, an LED, fluorescein fluorescence, and SPDC output, and the dependence of both Stokes parameters and entropy on pixel binning was studied explicitly (George et al., 2022).

Graph-theoretic analogues replace $S(\rho)=0$5 by a trace-normalized Laplacian. For $S(\rho)=0$6, the graph von Neumann entropy is

$S(\rho)=0$7

with $S(\rho)=0$8 now the eigenvalues of $S(\rho)=0$9. Because exact spectral computation is expensive, several surrogate estimators have been studied. One line uses quadratic approximations reducible to degree statistics, such as

$\rho$0

or, for the normalized Laplacian,

$\rho$1

These approximations can correlate well with exact entropies on Erdős–Rényi and some scale-free graphs, but can fail badly on highly structured small-world graphs because they miss path-length, clustering, and symmetry effects (Minello et al., 2018). A sharper surrogate is the structural information

$\rho$2

the Shannon entropy of the normalized degree sequence. For undirected unweighted graphs, the entropy gap

$\rho$3

satisfies

$\rho$4

so the degree-based estimator differs from the exact von Neumann graph entropy by at most $\rho$5 bits. The same work reports that this structural-information approximation is at least two orders of magnitude faster than SLaQ with comparable accuracy (Liu et al., 2021).

6. Operational, algebraic, and conceptual constraints

Entropy estimation is also shaped by structural theorems about when entropy changes, when it is operationally meaningful, and when it can be defined without finite-dimensional spectral calculus. For a bistochastic quantum operation $\rho$6, entropy preservation is characterized exactly by

$\rho$7

The same paper characterizes preservation of channel map entropy under composition: $\rho$8 These identities identify circumstances in which entropy can be propagated through a process without recomputation, and they convert entropy preservation into a fixed-point test for $\rho$9 (Zhang et al., 2011).

A more radical operational viewpoint characterizes von Neumann entropy through single-shot state convertibility. In a resource theory with unitaries, catalysts, and a dephasing environment on the catalyst, a transition $0\le S(\rho)\le \log d$0 is possible exactly when

$0\le S(\rho)\le \log d$1

for states of the same finite dimension and different spectra. This yields a transformability-based notion of entropy estimation: the entropy of $0\le S(\rho)\le \log d$2 is the threshold governing which target states are reachable under the allowed operations (Boes et al., 2018).

In algebraic QFT, local algebras are type III, so the usual reduced-density-matrix definition is unavailable. The operator-algebraic solution is to place, for $0\le S(\rho)\le \log d$3, a canonical intermediate type I factor $0\le S(\rho)\le \log d$4 between $0\le S(\rho)\le \log d$5 and $0\le S(\rho)\le \log d$6, and define

$0\le S(\rho)\le \log d$7

where $0\le S(\rho)\le \log d$8 is the vacuum density matrix restricted to $0\le S(\rho)\le \log d$9. The paper proves finiteness of this canonical entropy for the chiral conformal net generated by finitely many free Fermions, and finiteness of the lower entanglement entropy $\rho$00 for finitely many commuting $\rho$01-currents (Longo et al., 2019).

An even more general operational reconstruction dispenses with the spectral theorem itself. Using repeatability and reversibility, together with the second law, one can define a unique spectral entropy for weakly spectral states and show that the Groenewold–Ozawa information gain

$\rho$02

is monotone under a suitable ordering of instruments. In standard quantum theory this recovers the usual von Neumann entropy, but the construction is formulated in a broader operational setting (Minagawa et al., 2022).

Across these formulations, the unresolved issues are consistent. Observer-based methods still leave open the extension to open Lindblad dynamics and the relaxation of the full-rank assumption for relative entropy convergence (Balas et al., 2023). Variational estimators still lack full noise-robust and copy-complexity guarantees (Goldfeld et al., 2023). Operational characterizations still hinge on unresolved conjectures such as the catalytic entropy conjecture (Boes et al., 2018). Taken together, the literature indicates that von Neumann entropy estimation is best understood not as a single algorithmic task, but as a spectrum of inference problems whose solvability depends on access model, dynamical structure, symmetry, and the extent to which spectral information can be replaced by operational or model-specific surrogates.