---
title: Nussbaum–Szkoła Distributions in Quantum Testing
url: https://www.emergentmind.com/topics/nussbaum-szkola-distributions
type: topic
---

# Nussbaum–Szkoła Distributions in Quantum Testing

Searching arXiv for recent and foundational papers on Nussbaum–Szkoła distributions to ground the article.
Nussbaum–Szkoła distributions are a canonical classical representation associated with a pair of quantum states. Given spectral decompositions of two states, they assign probability weights on pairs of eigen-indices by combining eigenvalues with basis-overlap intensities. Their defining significance is that key quantum distinguishability quantities become exactly classical for the resulting distributions. In particular, for two states \(\rho\) and \(\sigma\), the Chernoff overlap \(\operatorname{Tr}(\rho^s\sigma^{1-s})\) is reproduced by the classical overlap \(\sum_{i,j} P_{ij}^s Q_{ij}^{1-s}\), and more generally a broad class of Petz-type quantum \(f\)-divergences equals the corresponding classical \(f\)-divergence of the Nussbaum–Szkoła pair [1508.06624] [2308.02929]. This construction has become a central analytical device in quantum hypothesis testing, asymptotic error analysis, and the study of quantum divergences, with extensions from finite-dimensional matrix algebras to separable Hilbert spaces and semifinite von Neumann algebras [2606.06246] [2604.19853].

## 1. Definition and basic construction

For two finite-dimensional quantum states \(\rho\) and \(\sigma\) with spectral decompositions
\[
\rho=\sum_i \lambda_i |v_i\rangle\langle v_i|,\qquad
\sigma=\sum_j \mu_j |w_j\rangle\langle w_j|,
\]
the Nussbaum–Szkoła distributions are defined on index pairs \((i,j)\) by
\[
P_{ij}:=\lambda_i |\langle v_i|w_j\rangle|^2,\qquad
Q_{ij}:=\mu_j |\langle v_i|w_j\rangle|^2.
\]
The same construction is described in equivalent notation in later works, for example
\[
P_{ij}:=r_i |\langle \psi_i|\phi_j\rangle|^2,\qquad
Q_{ij}:=s_j |\langle \psi_i|\phi_j\rangle|^2
\]
for spectral decompositions \(\rho=\sum_i r_i |\psi_i\rangle\langle \psi_i|\) and \(\sigma=\sum_j s_j |\phi_j\rangle\langle \phi_j|\) [2308.02929] [2604.06908].

Normalization follows from the orthonormality of the eigenbases:
\[
\sum_{i,j} P_{ij}=1,\qquad \sum_{i,j} Q_{ij}=1,
\]
for normalized states. In a more general trace-class setting, one has \(\sum_{i,j}P_{ij}=\operatorname{Tr}[\rho]\) and \(\sum_{i,j}Q_{ij}=\operatorname{Tr}[\sigma]\) [2308.02929] [2606.06246].

The overlap matrix
\[
W_{ij}:=|\langle \psi_i|\phi_j\rangle|^2
\]
is doubly stochastic:
\[
\sum_j W_{ij}=1,\qquad \sum_i W_{ij}=1.
\]
This matrix encodes the relative orientation of the eigenbases. Degeneracies are handled by choosing any orthonormal basis inside each degenerate eigenspace. The resulting NS distributions may depend on that basis choice, but several divergence values computed from them do not [2308.02929].

In separable Hilbert spaces, the same construction remains discrete because trace-class density operators admit countable spectral decompositions. Continuous spectra do not arise for trace-class density operators in that setting [2308.02929].

## 2. Chernoff identity and the original hypothesis-testing role

The defining identity behind the hypothesis-testing applications is
\[
\sum_{i,j} P_{ij}^s Q_{ij}^{1-s}=\operatorname{Tr}(\rho^s \sigma^{1-s}),\qquad 0\le s\le 1.
\]
As a consequence, the classical Chernoff information of \((P,Q)\),
\[
C_{\mathrm{cl}}(P,Q):=-\log \min_{0\le s\le 1}\sum_{i,j}P_{ij}^sQ_{ij}^{1-s},
\]
exactly matches the quantum Chernoff quantity
\[
-\log \min_{0\le s\le 1}\operatorname{Tr}(\rho^s\sigma^{1-s}).
\]
This identity is the classical backbone of the Nussbaum–Szkoła lower-bound technique in quantum hypothesis testing [1508.06624].

The same point is expressed in later formulations through the equality of quantum and classical Chernoff distances:
\[
\xi_{QC}(\rho,\sigma):=-\log \inf_{0\le s\le 1}\operatorname{Tr}[\rho^s\sigma^{1-s}]
\]
and
\[
\xi_C(P,Q):=-\log \inf_{0\le s\le 1}\sum_{i,j}P_{ij}^sQ_{ij}^{1-s},
\]
with \(\xi_{QC}(\rho,\sigma)=\xi_C(P,Q)\) [2606.06246].

This exact transfer from quantum to classical Chernoff overlaps explains why NS distributions entered the subject through quantum state discrimination. Classical large-deviation and Chernoff machinery can be applied to \((P,Q)\), and the resulting statements then yield quantum lower bounds or exact asymptotic exponents. In the binary i.i.d. setting, this underlies the quantum Chernoff bound. In the multi-hypothesis setting, it supports the passage from pairwise exponents to the multiple quantum Chernoff distance [1508.06624].

A related property emphasized in more recent work is that NS distributions preserve Petz Rényi divergences for \(\alpha\in[0,1]\), and this supports the transfer of Hoeffding- and Stein-type asymptotic analyses to the classical side as well [2606.06246].

## 3. Multiple-state discrimination and the Nussbaum–Szkoła conjecture

For an arbitrary finite ensemble \(\{\rho_1,\ldots,\rho_r\}\) on a finite-dimensional Hilbert space and an arbitrary prior \(\{p_1,\ldots,p_r\}\) independent of \(n\), one tests the i.i.d. hypotheses \(\{\rho_1^{\otimes n},\ldots,\rho_r^{\otimes n}\}\). The optimal average error probability satisfies
\[
P_e(\{p_i \rho_i^{\otimes n}\}_{i=1}^r)=\exp\{-\xi n+o(n)\},
\]
with
\[
\xi=\min_{i\neq j} C(\rho_i,\rho_j),
\qquad
C(\rho_i,\rho_j):=\max_{0\le s\le 1}\{-\log \operatorname{Tr}(\rho_i^s\rho_j^{1-s})\}.
\]
This resolves the conjecture of Nussbaum and Szkoła that the optimal asymptotic error exponent for multiple quantum hypotheses equals the minimum over pairwise Chernoff distances [1508.06624].

In that proof, NS distributions enter through a lower-bound mechanism: each quantum pair \((\rho_i,\rho_j)\) is mapped to classical NS distributions \((P^{(i,j)},Q^{(i,j)})\), classical Chernoff bounds are applied to those pairs, and appropriate pairwise contributions are aggregated. The paper crystallizes this through a one-shot lower bound
\[
P_*(\{A_1,\ldots,A_r\})\ge \frac{1}{2(r-1)}\sum_{i<j}\sum_{k,\ell}\min\{\lambda_{ik},\lambda_{j\ell}\}\operatorname{Tr}(Q_{ik}Q_{j\ell}),
\]
for spectral decompositions \(A_i=\sum_{k=1}^{T_i}\lambda_{ik}Q_{ik}\). For density matrices with rank-one spectral projectors, the overlap term reduces to \(|\langle v_{ik}|w_{j\ell}\rangle|^2\), and the summand matches the classical quantity appearing in the NS construction [1508.06624].

The same line of development was later sharpened in arbitrary separable Hilbert spaces. A dimension-free one-shot upper bound in terms of pairwise errors shows
\[
\operatorname{Err}^*(A_1,\ldots,A_M)\le 4\sum_{1\le i<j\le M}\operatorname{Err}^*(P^{ij},Q^{ij}),
\]
where \((P^{ij},Q^{ij})\) is the NS pair for \((A_i,A_j)\). In the i.i.d. regime this yields
\[
\operatorname{Err}^*(p_1\rho_1^{\otimes n},\ldots,p_M\rho_M^{\otimes n})
\le 4(M-1)e^{-n C(\rho_1,\ldots,\rho_M)},
\]
with
\[
C(\rho_1,\ldots,\rho_M)=\min_{i\neq j}\xi_{QC}(\rho_i,\rho_j).
\]
This removes the dimension-dependent prefactor present in earlier finite-dimensional results and establishes achievability of the multiple Chernoff distance in arbitrary separable Hilbert spaces [2606.06246].

These results jointly situate NS distributions as the pairwise classical objects governing the asymptotics of multiple quantum testing: they provide the lower-bound exponent, guide the structure of pairwise decompositions, and remain effective beyond finite dimensions [1508.06624] [2606.06246].

## 4. Equality for quantum \(f\)-divergences

A major generalization of the NS framework shows that, for a very broad class of quantum \(f\)-divergences, the quantum quantity equals the classical \(f\)-divergence of the corresponding NS distributions. Let \(f:(0,\infty)\to\mathbb{R}\) be convex or concave, and define the classical Csiszár \(f\)-divergence by
\[
D_f(P\Vert Q)=\sum_{x\in X}Q(x)\,f\!\left(\frac{P(x)}{Q(x)}\right),
\]
with the standard conventions for zero entries. For density operators \(\rho,\sigma\), the quantum \(f\)-divergence considered in [2308.02929] is the Petz-type quasi-entropy defined through the relative modular operator \(\Delta_{\rho,\sigma}\):
\[
D_f(\rho\Vert\sigma)=\int_{0^+}^{\infty} f(\lambda)\,
\langle \sqrt{\sigma},\,\xi^{\Delta_{\rho,\sigma}}(d\lambda)\,\sqrt{\sigma}\rangle_2
+f(0)\operatorname{Tr}(\sigma \Pi_\rho^\perp)
+f'(\infty)\operatorname{Tr}(\rho \Pi_\sigma^\perp).
\]
The main theorem states that
\[
D_f(\rho\Vert\sigma)=D_f(P\Vert Q),
\]
where \(P,Q\) are the NS distributions of \(\rho,\sigma\) [2308.02929].

This equality holds in finite and infinite dimensions, without faithfulness assumptions, and support mismatches are accounted for by boundary terms on both sides. The equivalence
\[
P\ll Q \quad\Longleftrightarrow\quad \operatorname{supp}(\rho)\subseteq \operatorname{supp}(\sigma)
\]
determines when the divergence reduces to the pure-ratio form without boundary terms [2308.02929].

The mechanism behind the equality is spectral. The relative modular operator acts on rank-one operators \(|e_i\rangle\langle f_j|\) with eigenvalues \(\lambda_i/\mu_j\), and the spectral measure weight tested against \(\sqrt{\sigma}\) is precisely \(\mu_j |\langle e_i|f_j\rangle|^2\). Thus the operator functional calculus reduces to the scalar summation defining the classical \(f\)-divergence of the NS pair [2308.02929].

This yields direct quantum versions of many classical inequalities. Representative examples transferred in [2308.02929] include:
\[
D(\rho\Vert\sigma)\le \log\!\big(1+\chi^2(\rho\Vert\sigma)\big),
\]
\[
D(\rho\Vert\sigma)\le \tfrac12\big(V(\rho\Vert\sigma)+\chi^2(\rho\Vert\sigma)\big),
\]
and
\[
\mathscr{H}^2(\rho\Vert\sigma)\le V(\rho\Vert\sigma)^2
\le \mathscr{H}(\rho\Vert\sigma)\sqrt{2-\mathscr{H}^2(\rho\Vert\sigma)}.
\]
A Pinsker-type continuity statement is also obtained: for strictly convex \(f\) with \(f(1)=0\), there exists \(\psi_f\) with \(\lim_{x\downarrow0}\psi_f(x)=0\) such that
\[
V(\rho\Vert\sigma)\le \psi_f\big(D_f(\rho\Vert\sigma)\big),
\]
and therefore \(D_f(\rho_n\Vert\sigma_n)\to 0\) implies \(V(\rho_n\Vert\sigma_n)\to 0\) [2308.02929].

A later extension proves the same reduction for normal states on a semifinite von Neumann algebra. If \(M\) is semifinite with faithful normal semifinite trace \(\tau\), and \(\varphi,\psi\) are normal states, then there exist a \(\sigma\)-finite measured space \((X,\Sigma,\nu)\) and nonnegative measurable functions \(f_\varphi,f_\psi\) such that
\[
S_f(\varphi\Vert\psi)=D_f(f_\varphi d\nu\Vert f_\psi d\nu)
\]
for every convex \(f:(0,\infty)\to\mathbb{R}\). In finite dimensions this recovers the familiar discrete formulas \(p_{ij}=\lambda_i|\langle v_i|w_j\rangle|^2\) and \(q_{ij}=\mu_j|\langle v_i|w_j\rangle|^2\) [2604.19853].

## 5. Structural and geometric perspectives

The NS mapping has also been used beyond the standard \(f\)-divergence class. A 2026 work introduces a quantum relative-\(\alpha\)-entropy
\[
S_\alpha(\rho\Vert\sigma)
=
\frac{\alpha}{1-\alpha}\log \operatorname{Tr}(\rho\,\sigma^{\alpha-1})
-\frac{1}{1-\alpha}\log \operatorname{Tr}(\rho^\alpha)
+\log \operatorname{Tr}(\sigma^\alpha),
\qquad \alpha>0,\ \alpha\neq 1,
\]
with the convention \(S_\alpha(\rho\Vert\sigma):=+\infty\) whenever \(\operatorname{supp}(\rho)\nsubseteq \operatorname{supp}(\sigma)\) [2604.06908].

For the NS-type distributions
\[
P_{ij}:=r_i |\langle \psi_i|\phi_j\rangle|^2,\qquad
Q_{ij}:=s_j |\langle \psi_i|\phi_j\rangle|^2,
\]
the paper proves the exact correspondence
\[
S_\alpha(\rho\Vert\sigma)=J_\alpha(P\Vert Q),
\]
where
\[
J_\alpha(P\Vert Q)
=
\frac{\alpha}{1-\alpha}\log \sum_{i,j} P_{ij}Q_{ij}^{\alpha-1}
-\frac{1}{1-\alpha}\log \sum_{i,j} P_{ij}^\alpha
+\log \sum_{i,j} Q_{ij}^\alpha.
\]
The work emphasizes that this divergence lies outside the quantum \(f\)-divergence class, yet still admits an exact classical reduction through NS-type distributions [2604.06908].

Within that framework, the overlap matrix \(W_{ij}=|\langle \psi_i|\phi_j\rangle|^2\) is interpreted as encoding the relative geometry of the eigenbases. The divergence depends on the eigenvalues and on \(W\), is invariant under simultaneous unitary conjugation, and is additive under tensor products because the corresponding NS distributions factorize:
\[
P^{(12)}_{(i,k),(j,\ell)}=P^{(1)}_{ij}P^{(2)}_{k\ell},\qquad
Q^{(12)}_{(i,k),(j,\ell)}=Q^{(1)}_{ij}Q^{(2)}_{k\ell}.
\]
This suggests that the NS formalism is not limited to the traditional Petz-type setting, but can also expose geometric content in divergences whose classical analogues are not Csiszár \(f\)-divergences [2604.06908].

That same work derives generalized convexity statements on commuting subclasses using multiplicative mixtures
\[
M_{\rho,\sigma}^t=\frac{\rho^t \sigma^{1-t}}{\operatorname{Tr}(\rho^t \sigma^{1-t})},
\qquad t\in[0,1],
\]
and obtains a generalized convexity result for Petz–Rényi divergence for \(\alpha>1\), complementing the known convexity for \(\alpha<1\) [2604.06908].

## 6. Operational refinements in binary and asymmetric testing

In binary Bayesian testing, the NS map compares the optimal quantum Bayes error with the optimal classical Bayes error of the associated NS pair. For priors \(p\in(0,1)\), let \(A:=p\rho\) and \(B:=(1-p)\sigma\). The optimal quantum Bayes error is
\[
P_e^Q(\rho,\sigma)=\operatorname{Tr}[A\wedge B],
\]
while the optimal classical Bayes error for the NS distributions of \(A,B\) is
\[
P_e^C(P,Q)=\sum_{i,j}\min\{P_{ij},Q_{ij}\}.
\]
A dimension-free result shows
\[
\frac12 P_e^C(P,Q)\le P_e^Q(\rho,\sigma)\le 2 P_e^C(P,Q),
\]
for all separable Hilbert spaces and arbitrary priors. Nussbaum–Szkoła had established the lower bound \(\frac12 P_e^C\le P_e^Q\); the matching upper bound is the new contribution in [2606.06246].

The same paper introduces the Petz–Nussbaum–Szkoła trace harmonic mean
\[
HM_s(A,B):=\operatorname{Tr}[P !_s Q]
=\sum_{i,j}\frac{P_{ij}Q_{ij}}{sP_{ij}+(1-s)Q_{ij}},
\qquad s\in(0,1),
\]
where \(A !_s B\) denotes the Kubo–Ando weighted harmonic mean. It proves
\[
(s\wedge(1-s))\,HM_s(A,B)\le \operatorname{Tr}[A\wedge B]\le HM_s(A,B),
\]
and
\[
HM_s(A,B)\le \operatorname{Tr}[A^{1-s}B^s].
\]
At \(s=\tfrac12\), this yields a direct sandwich for the quantum Bayes error, and combining it with the scalar comparison between \(HM_{1/2}\) and \(\sum P\wedge Q\) gives the factor-of-two theorem [2606.06246].

For asymmetric hypothesis testing, a converse bound based on the NS mapping takes the form
\[
\beta_\alpha(\rho,\sigma)\ge s\,\beta_{\alpha/(1-s)}(P,Q),
\qquad 0\le s\le 1,\quad 0\le \alpha\le 1-s.
\]
Here \(\beta_\alpha(\rho,\sigma)\) is the optimal type-II error under a type-I constraint \(\alpha\), and \(\beta_{\alpha/(1-s)}(P,Q)\) is the classical Neyman–Pearson trade-off for the NS pair. This single one-shot inequality yields unified converses in the small-, moderate-, and large-deviation regimes by choosing \(s=s_n\) appropriately and importing classical asymptotic results for \(P^n,Q^n\) [2601.13970].

The transferred asymptotic forms include:
\[
D_h^\epsilon(\rho^{\otimes n}\Vert \sigma^{\otimes n})
\le n D(P\Vert Q)+\sqrt{nV(P\Vert Q)}\,\Phi^{-1}(\epsilon)+O(\log n)
\]
in the fixed-\(\epsilon\) regime when \(V(P\Vert Q)>0\),
\[
D_h^{\epsilon_n}(\rho^{\otimes n}\Vert \sigma^{\otimes n})
\le nD(P\Vert Q)-\sqrt{2V(P\Vert Q)}\,na_n+o(na_n)
\]
for \(\epsilon_n=e^{-na_n^2}\),
and the Hoeffding-type exponent bound
\[
\limsup_{n\to\infty}\frac1n D_h^{\epsilon_n}(\rho^{\otimes n}\Vert \sigma^{\otimes n})
\le
\sup_{0\le s\le 1}
\left\{
\frac{1}{s-1}\log \operatorname{Tr}[\rho^s \sigma^{1-s}]
+\frac{s}{s-1}r
\right\}
\]
when \(\epsilon_n\le e^{-nr}\) [2601.13970].

These refinements strengthen the operational interpretation of NS distributions. They are not merely a vehicle for asymptotic exponents; they can bound the exact one-shot error trade-off and provide accurate finite-blocklength approximations [2601.13970] [2606.06246].

## 7. Special cases, caveats, and conceptual position

Several simplifying scenarios clarify the NS construction.

If \(\rho\) and \(\sigma\) commute, one can choose a common eigenbasis. Then
\[
P_{ij}=\lambda_i\delta_{ij},\qquad Q_{ij}=\mu_j\delta_{ij},
\]
and the NS distributions reduce to the ordinary classical eigenvalue distributions. In that case the Chernoff distance is classical,
\[
C(\rho,\sigma)=\max_{0\le s\le 1}\left\{-\log \sum_t \lambda_t^s \mu_t^{1-s}\right\},
\]
and classical finite-blocklength performance directly characterizes the quantum problem [1508.06624] [2308.02929].

For pure states \(\rho=|\psi\rangle\langle\psi|\) and \(\sigma=|\phi\rangle\langle\phi|\),
\[
\operatorname{Tr}(\rho^s \sigma^{1-s})=|\langle \psi|\phi\rangle|^2
\]
for all \(s\), hence
\[
C(\rho,\sigma)=-\log |\langle \psi|\phi\rangle|^2.
\]
The corresponding NS distributions have a single nonzero entry, and the classical Chernoff information equals the quantum one [1508.06624].

When supports are pairwise disjoint, overlap terms vanish and the NS construction becomes trivial in the sense that the relevant classical overlaps are zero. In such cases Gram–Schmidt-based constructions directly yield optimal measurements in multiple-state testing [1508.06624].

At the same time, several caveats recur across the literature.

First, NS distributions are not obtained by measuring the two states with a single POVM. Measured \(f\)-divergences are produced by a fixed measurement and typically lower bound the corresponding quantum divergence. NS distributions are different: they depend simultaneously on both states through their eigenbases and serve as an analytical construct yielding exact equalities for Petz-type \(f\)-divergences [2308.02929].

Second, degeneracies matter at the level of the distributions themselves. Different orthonormal bases within degenerate eigenspaces can lead to different NS pairs, even though the divergence values relevant to the main equalities remain basis-independent [2308.02929]. In asymmetric converse bounds, the validity of the inequalities does not depend on the basis choice, though numerical tightness may [2601.13970].

Third, the exact NS equality does not extend unchanged to all quantum divergences. For example, the equality \(D_\alpha(\rho\Vert\sigma)=D_\alpha(P\Vert Q)\) generally does not hold for sandwiched Rényi divergence; the exact correspondence applies to Petz-type Rényi divergences and to the Petz-type \(f\)-divergence class [2308.02929].

Finally, the domain of the theory has widened substantially. What began as a finite-dimensional tool for Chernoff-type quantum hypothesis testing now applies in separable Hilbert spaces for both hypothesis-testing exponents and \(f\)-divergence equalities, and in semifinite von Neumann algebras through a joint spectral representation of left and right multiplication operators [2606.06246] [2604.19853]. This suggests a durable structural role for Nussbaum–Szkoła distributions: they provide a precise classical model of the joint spectral mismatch between two quantum states, and in that model a wide range of quantum distinguishability problems become classical.

Source: https://www.emergentmind.com/topics/nussbaum-szkola-distributions