Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2510.06147

Non-iid hypothesis testing: from classical to quantum

Giacomo De Palma, Marco Fanizza, Connor Mowry, Ryan O’Donnell

Published Oct 7, 2025• 0 citations
arXivHTMLPDF

Abstract

We study hypothesis testing (aka state certification) in the non-identically distributed setting. A recent work (Garg et al. 2023) considered the classical case, in which one is given (independent) samples from $T$ unknown probability distributions $p_1, \dots, p_T$ on $[d] = \{1, 2, \dots, d\}$, and one wishes to accept/reject the hypothesis that their average $p_{\mathrm{avg}}$ equals a known hypothesis distribution $q$. Garg et al. showed that if one has just $c = 2$ samples from each $p_i$, and provided $T \gg \frac{\sqrt{d}}{ε^2} + \frac{1}{ε^4}$, one can (whp) distinguish $p_{\mathrm{avg}} = q$ from $d_{\mathrm{TV}}(p_{\mathrm{avg}},q) > ε$. This nearly matches the optimal result for the classical iid setting (namely, $T \gg \frac{\sqrt{d}}{ε^2}$). Besides optimally improving this result (and generalizing to tolerant testing with more stringent distance measures), we study the analogous problem of hypothesis testing for non-identical quantum states. Here we uncover an unexpected phenomenon: for any $d$-dimensional hypothesis state $σ$, and given just a single copy ($c = 1$) of each state $ρ_1, \dots, ρ_T$, one can distinguish $ρ_{\mathrm{avg}} = σ$ from $D_{\mathrm{tr}}(ρ_{\mathrm{avg}},σ) > ε$ provided $T \gg d/ε^2$. (Again, we generalize to tolerant testing with more stringent distance measures.) This matches the optimal result for the iid case, which is surprising because doing this with $c = 1$ is provably impossible in the classical case. We also show that the analogous phenomenon happens for the non-iid extension of identity testing between unknown states. A technical tool we introduce may be of independent interest: an Efron-Stein inequality, and more generally an Efron-Stein decomposition, in the quantum setting.

DOI
10.48550/arxiv.2510.06147

Related papers

Generalized quantum Stein's lemma for mixed sources

Haruka Kanazawa, Hayata Yamasaki

The generalized quantum Stein's lemma characterizes the optimal asymptotic exponent of the type-II error in quantum hypothesis testing for an independent and identically distributed (IID) null hypothesis against a composite alternative hypothesis. Classically, a probabilistic mixture of IID sources arises as a natural generalization of IID sources, and, in the non-composite setting, the optimal type-II error exponent in hypothesis testing for such classical mixed sources is known to be characterized concisely by the worst-case component of the mixture. In this work, we extend these foundational results to composite quantum hypothesis testing where the null hypothesis is a mixed source, i.e., a probabilistic mixture of IID quantum states, and the alternative hypothesis is composite as in the generalized quantum Stein's lemma. When the type-I error vanishes asymptotically, we characterize the optimal type-II error exponent of this composite quantum hypothesis testing problem in terms of the worst-case component of the mixture, by developing techniques for the non-commutative quantum setting inspired by the classical information-spectrum analysis. We also show that the analogous characterization does not hold in general for a fixed nonzero type-I error threshold, by providing a counterexample beyond the vanishing type-I error regime. These results clarify the applicability of the generalized quantum Stein's lemma to highly non-IID null hypotheses arising from arbitrary finite probabilistic mixtures of IID quantum states.

quant-pharXiv:2605.20776•May 21, 2026•0 citations

On Adaptivity in Classical and Quantum Learning

Aadil Oufkir

L'Adaptativité dans l'Apprentissage Classique et Quantique Apprendre à partir de données est un problème fondamental en statistique qui a des applications dans presque tous les aspects de notre vie. Il existe deux classes naturelles de stratégies pour apprendre à partir de données. Les stratégies non adaptatives collectent toutes les données nécessaires, puis traitent toutes ces données en une seule fois, tandis que les stratégies adaptatives peuvent utiliser les informations accumulées pour décider si davantage d'échantillons doivent être collectés. De plus, avec les méthodes adaptatives, la façon dont les nouvelles informations sont obtenues peut également être adaptée en fonction de ce qui a été observé lors des étapes précédentes. Étant donné que notre objectif est de minimiser la quantité de données requises, dans cette thèse, nous étudions le nombre optimal de ressources pour les deux approches pour différents problèmes d'apprentissage classiques et quantiques.Pour tester les distributions classiques, nous montrons que les stratégies adaptatives surpassent leurs analogues non adaptatives d'un facteur de quatre lorsque la taille de l'alphabet est petite. De plus, les stratégies adaptatives peuvent s'arrêter plus tôt lorsque les distributions testées sont très différentes les unes des autres. Cet avantage est également présent même pour les grands alphabets.En ce qui concerne le test des états quantiques, nous présentons des situations où les stratégies adaptatives ont un avantage significatif par rapport aux stratégies non adaptatives. Celles-ci comprennent le test d'hypothèses (binaires), le test d'identité et de proximité des états quantiques.Cependant, dans certaines situations, nous avons plusieurs façons de générer les données. Ces situations peuvent être modélisées par des canaux quantiques qui produisent des données après avoir reçu une entrée choisie par la stratégie d'apprentissage. Ces entrées peuvent également être adaptées aux observations précédentes par les stratégies adaptatives, tandis que les stratégies non adaptatives doivent déterminer toutes les entrées à l'avance. Nous déterminons la complexité optimale pour tester si un canal est parfait ou non. De plus, nous caractérisons le nombre optimal de ressources pour apprendre un canal général ou un bruit de type Pauli dans le cadre non adaptatif. En outre, nous réduisons l'écart entre les stratégies adaptatives et non adaptatives pour l'apprentissage du bruit de type Pauli.

•Sep 14, 2023•0 citations

Online Shadow Tomography Matching the Classical Bounds

Sitan Chen, Ryan O'Donnell, Angelos Pelecanos +1 more

In Online Shadow Tomography, we are given copies of an unknown $d$-dimensional quantum state $ρ$, an adversary (adaptively) proposes a sequence of bounded observables $A^{(1)},\ldots,A^{(m)}$, and after each $A^{(t)}$ is given we must estimate $\mathrm{Tr}(A^{(t)}ρ)$ to within $\pm ε$. This is the direct quantum generalization of the classical problem of Adaptive Data Analysis. Prior results for online Shadow Tomography were suboptimal in all three parameters $m, d, ε$, lagging behind the best known and classical rates, for which there is some evidence of optimality. In this work, we finally close this gap, giving a pair of algorithms matching the classical rates. Our first algorithm is the first to achieve $o(\log^2 m)$-dependence together with $\mathrm{poly}(\log(d)/ε)$; moreover, it improves all three exponents even in the Offline Shadow Tomography setting. Our second algorithm is known to be optimal among bounds independent of $d$, and improves the best prior result by a $\sqrt{m} \log m$ factor. The key to our proof is a new framework for quantifying post-measurement damage, based on the quantum Efron-Stein decomposition.

quant-pharXiv:2607.29686•Aug 3, 2026•0 citations

Adversarial Hypothesis Testing and a Quantum Stein’s Lemma for Restricted Measurements

Fernando G. S. L. Brandão, Aram W. Harrow, James R. Lee +1 more

Recall the classical hypothesis testing setting with two sets of probability distributions P and Q. One receives either n i.i.d. samples from a distribution p E P or from a distribution q E Q and wants to decide from which set the points were sampled. It is known that the optimal exponential rate at which errors decrease can be achieved by a simple maximum-likelihood ratio test which does not depend on p or q, but only on the sets P and Q. We consider an adaptive generalization of this model where the choice of p E P and q E Q can change in each sample in some way that depends arbitrarily on the previous samples. In other words, in the kth round, an adversary, having observed all the previous samples in rounds 1, . . . , k - 1, chooses pk E P and qk E Q, with the goal of confusing the hypothesis test. We prove that even in this case, the optimal exponential error rate can be achieved by a simple maximum-likelihood test that depends only on P and Q. We then show that the adversarial model has applications in hypothesis testing for quantum states using restricted measurements. For example, it can be used to study the problem of distinguishing entangled states from the set of all separable states using only measurements that can be implemented with local operations and classical communication (LOCC). The basic idea is that in our setup, the deleterious effects of entanglement can be simulated by an adaptive classical adversary. We prove a quantum Stein's Lemma in this setting: In many circumstances, the optimal hypothesis testing rate is equal to an appropriate notion of quantum relative entropy between two states. In particular, our arguments yield an alternate proof of Li and Winter's recent strengthening of strong subadditivity for von Neumann entropy.

IEEE Transactions on Information Theory•Mar 10, 2020•29 citations

Robust generalized quantum Stein's lemma

Giulia Mazzola, David Sutter, Renato Renner

The generalized quantum Stein's lemma provides an explicit expression for the optimal error exponent when distinguishing many independent and identically distributed (iid) copies of a given bipartite state from the set of separable bipartite states. Here we prove that this result is robust, in the sense that the iid assumption can be relaxed to almost-iid. In particular, our result shows that the original argument of Brandão and Plenio, which contains a logical gap, can be made rigorous. Our proof relies on a novel continuity bound for the relative entropy of entanglement with respect to the quantum Wasserstein distance. Combined with a recent insight that almost-iid states and their exact iid counterparts are asymptotically close in this distance, the bound implies that their relative entropies of entanglement coincide asymptotically.

quant-pharXiv:2605.16500•May 19, 2026•0 citations

On the Power of Adaptivity in Testing Quantum States in Fidelity

Jan Seyfried, Sayantan Sen, Marco Tomamichel

We study the problems of quantum state certification, equivalence testing and independence testing. In certification, given samples of an unknown quantum state $ρ$ and the description of a state $σ$, the goal is to test whether $ρ=σ$, or whether $ρ$ and $σ$ are far in a given distance measure. In equivalence testing, $σ$ is also unknown and only accessible via samples. Independence testing decides whether $ρ_{AC}=ρ_A\otimesρ_C$, or is far from being a product. The sample complexities of these problems are now well-understood for a decision gap $\varepsilon$ in trace distance: in the single-copy measurement setting with $d$-dimensional states, all three tasks can be solved using the same non-adaptive approach, which uses $Θ(d^{3/2}/\varepsilon^2)$ samples and is optimal in general, even without adaptivity. In this work, we consider decision gaps expressed in fidelity and study possible separations between these problems and how adaptivity can help. We prove that certification with respect to fidelity for a state $σ$ of rank $r$ does not benefit from adaptivity and requires $\widetildeΘ(r^{3/2}/\varepsilon)$ samples. For equivalence testing and independence testing, we provide adaptive algorithms using $\widetilde{O}(\min\{d^{3/2}/\varepsilon^2,d^{9/4}/\varepsilon\})$ and $\widetilde{O}(\min\{(d_Ad_C)^{3/2}/\varepsilon^2,d_A^{9/4}d_C^{3/4}/\varepsilon\})$ samples, for $d_A\geq d_C$, respectively. Our main technique is a framework that uses partial learning and a reduction to testing in $\ell_2$-distance, adapted from the distribution testing literature. We show that adaptivity matters for equivalence testing in fidelity by proving that $\widetildeΩ(1/\varepsilon^2)$ samples are necessary in the non-adaptive case even for qubits, showing a separation from certification.

quant-pharXiv:2609.08733•Sep 9, 2026•0 citations
We study hypothesis testing (aka state certification) in the non-identically distributed setting. A recent work (Garg et al. 2023) considered the classical case, in which one is given (independent) samples from TTT unknown probability distributions p1,…,pTp_1, \dots, p_Tp1​,…,pT​ on [d]={1,2,…,d}[d] = \{1, 2, \dots, d\}[d]={1,2,…,d}, and one wishes to accept/reject the hypothesis that their average pavgp_{\mathrm{avg}}pavg​ equals a known hypothesis distribution qqq. Garg et al. showed that if one has just c=2c = 2c=2 samples from each pip_ipi​, and provided T≫dε2+1ε4T \gg \frac{\sqrt{d}}{ε^2} + \frac{1}{ε^4}T≫ε2d​​+ε41​, one can (whp) distinguish pavg=qp_{\mathrm{avg}} = qpavg​=q from dTV(pavg,q)>εd_{\mathrm{TV}}(p_{\mathrm{avg}},q) > εdTV​(pavg​,q)>ε. This nearly matches the optimal result for the classical iid setting (namely, T≫dε2T \gg \frac{\sqrt{d}}{ε^2}T≫ε2d​​). Besides optimally improving this result (and generalizing to tolerant testing with more stringent distance measures), we study the analogous problem of hypothesis testing for non-identical quantum states. Here we uncover an unexpected phenomenon: for any ddd-dimensional hypothesis state σσσ, and given just a single copy (c=1c = 1c=1) of each state ρ1,…,ρTρ_1, \dots, ρ_Tρ1​,…,ρT​, one can distinguish ρavg=σρ_{\mathrm{avg}} = σρavg​=σ from Dtr(ρavg,σ)>εD_{\mathrm{tr}}(ρ_{\mathrm{avg}},σ) > εDtr​(ρavg​,σ)>ε provided T≫d/ε2T \gg d/ε^2T≫d/ε2. (Again, we generalize to tolerant testing with more stringent distance measures.) This matches the optimal result for the iid case, which is surprising because doing this with c=1c = 1c=1 is provably impossible in the classical case. We also show that the analogous phenomenon happens for the non-iid extension of identity testing between unknown states. A technical tool we introduce may be of independent interest: an Efron-Stein inequality, and more generally an Efron-Stein decomposition, in the quantum setting.
The generalized quantum Stein's lemma characterizes the optimal asymptotic exponent of the type-II error in quantum hypothesis testing for an independent and identically distributed (IID) null hypothesis against a composite alternative hypothesis. Classically, a probabilistic mixture of IID sources arises as a natural generalization of IID sources, and, in the non-composite setting, the optimal type-II error exponent in hypothesis testing for such classical mixed sources is known to be characterized concisely by the worst-case component of the mixture. In this work, we extend these foundational results to composite quantum hypothesis testing where the null hypothesis is a mixed source, i.e., a probabilistic mixture of IID quantum states, and the alternative hypothesis is composite as in the generalized quantum Stein's lemma. When the type-I error vanishes asymptotically, we characterize the optimal type-II error exponent of this composite quantum hypothesis testing problem in terms of the worst-case component of the mixture, by developing techniques for the non-commutative quantum setting inspired by the classical information-spectrum analysis. We also show that the analogous characterization does not hold in general for a fixed nonzero type-I error threshold, by providing a counterexample beyond the vanishing type-I error regime. These results clarify the applicability of the generalized quantum Stein's lemma to highly non-IID null hypotheses arising from arbitrary finite probabilistic mixtures of IID quantum states.
L'Adaptativité dans l'Apprentissage Classique et Quantique Apprendre à partir de données est un problème fondamental en statistique qui a des applications dans presque tous les aspects de notre vie. Il existe deux classes naturelles de stratégies pour apprendre à partir de données. Les stratégies non adaptatives collectent toutes les données nécessaires, puis traitent toutes ces données en une seule fois, tandis que les stratégies adaptatives peuvent utiliser les informations accumulées pour décider si davantage d'échantillons doivent être collectés. De plus, avec les méthodes adaptatives, la façon dont les nouvelles informations sont obtenues peut également être adaptée en fonction de ce qui a été observé lors des étapes précédentes. Étant donné que notre objectif est de minimiser la quantité de données requises, dans cette thèse, nous étudions le nombre optimal de ressources pour les deux approches pour différents problèmes d'apprentissage classiques et quantiques.Pour tester les distributions classiques, nous montrons que les stratégies adaptatives surpassent leurs analogues non adaptatives d'un facteur de quatre lorsque la taille de l'alphabet est petite. De plus, les stratégies adaptatives peuvent s'arrêter plus tôt lorsque les distributions testées sont très différentes les unes des autres. Cet avantage est également présent même pour les grands alphabets.En ce qui concerne le test des états quantiques, nous présentons des situations où les stratégies adaptatives ont un avantage significatif par rapport aux stratégies non adaptatives. Celles-ci comprennent le test d'hypothèses (binaires), le test d'identité et de proximité des états quantiques.Cependant, dans certaines situations, nous avons plusieurs façons de générer les données. Ces situations peuvent être modélisées par des canaux quantiques qui produisent des données après avoir reçu une entrée choisie par la stratégie d'apprentissage. Ces entrées peuvent également être adaptées aux observations précédentes par les stratégies adaptatives, tandis que les stratégies non adaptatives doivent déterminer toutes les entrées à l'avance. Nous déterminons la complexité optimale pour tester si un canal est parfait ou non. De plus, nous caractérisons le nombre optimal de ressources pour apprendre un canal général ou un bruit de type Pauli dans le cadre non adaptatif. En outre, nous réduisons l'écart entre les stratégies adaptatives et non adaptatives pour l'apprentissage du bruit de type Pauli.
In Online Shadow Tomography, we are given copies of an unknown ddd-dimensional quantum state ρρρ, an adversary (adaptively) proposes a sequence of bounded observables A(1),…,A(m)A^{(1)},\ldots,A^{(m)}A(1),…,A(m), and after each A(t)A^{(t)}A(t) is given we must estimate Tr(A(t)ρ)\mathrm{Tr}(A^{(t)}ρ)Tr(A(t)ρ) to within ±ε\pm ε±ε. This is the direct quantum generalization of the classical problem of Adaptive Data Analysis. Prior results for online Shadow Tomography were suboptimal in all three parameters m,d,εm, d, εm,d,ε, lagging behind the best known and classical rates, for which there is some evidence of optimality. In this work, we finally close this gap, giving a pair of algorithms matching the classical rates. Our first algorithm is the first to achieve o(log⁡2m)o(\log^2 m)o(log2m)-dependence together with poly(log⁡(d)/ε)\mathrm{poly}(\log(d)/ε)poly(log(d)/ε); moreover, it improves all three exponents even in the Offline Shadow Tomography setting. Our second algorithm is known to be optimal among bounds independent of ddd, and improves the best prior result by a mlog⁡m\sqrt{m} \log mm​logm factor. The key to our proof is a new framework for quantifying post-measurement damage, based on the quantum Efron-Stein decomposition.
Recall the classical hypothesis testing setting with two sets of probability distributions P and Q. One receives either n i.i.d. samples from a distribution p E P or from a distribution q E Q and wants to decide from which set the points were sampled. It is known that the optimal exponential rate at which errors decrease can be achieved by a simple maximum-likelihood ratio test which does not depend on p or q, but only on the sets P and Q. We consider an adaptive generalization of this model where the choice of p E P and q E Q can change in each sample in some way that depends arbitrarily on the previous samples. In other words, in the kth round, an adversary, having observed all the previous samples in rounds 1, . . . , k - 1, chooses pk E P and qk E Q, with the goal of confusing the hypothesis test. We prove that even in this case, the optimal exponential error rate can be achieved by a simple maximum-likelihood test that depends only on P and Q. We then show that the adversarial model has applications in hypothesis testing for quantum states using restricted measurements. For example, it can be used to study the problem of distinguishing entangled states from the set of all separable states using only measurements that can be implemented with local operations and classical communication (LOCC). The basic idea is that in our setup, the deleterious effects of entanglement can be simulated by an adaptive classical adversary. We prove a quantum Stein's Lemma in this setting: In many circumstances, the optimal hypothesis testing rate is equal to an appropriate notion of quantum relative entropy between two states. In particular, our arguments yield an alternate proof of Li and Winter's recent strengthening of strong subadditivity for von Neumann entropy.
The generalized quantum Stein's lemma provides an explicit expression for the optimal error exponent when distinguishing many independent and identically distributed (iid) copies of a given bipartite state from the set of separable bipartite states. Here we prove that this result is robust, in the sense that the iid assumption can be relaxed to almost-iid. In particular, our result shows that the original argument of Brandão and Plenio, which contains a logical gap, can be made rigorous. Our proof relies on a novel continuity bound for the relative entropy of entanglement with respect to the quantum Wasserstein distance. Combined with a recent insight that almost-iid states and their exact iid counterparts are asymptotically close in this distance, the bound implies that their relative entropies of entanglement coincide asymptotically.
We study the problems of quantum state certification, equivalence testing and independence testing. In certification, given samples of an unknown quantum state ρρρ and the description of a state σσσ, the goal is to test whether ρ=σρ=σρ=σ, or whether ρρρ and σσσ are far in a given distance measure. In equivalence testing, σσσ is also unknown and only accessible via samples. Independence testing decides whether ρAC=ρA⊗ρCρ_{AC}=ρ_A\otimesρ_CρAC​=ρA​⊗ρC​, or is far from being a product. The sample complexities of these problems are now well-understood for a decision gap ε\varepsilonε in trace distance: in the single-copy measurement setting with ddd-dimensional states, all three tasks can be solved using the same non-adaptive approach, which uses Θ(d3/2/ε2)Θ(d^{3/2}/\varepsilon^2)Θ(d3/2/ε2) samples and is optimal in general, even without adaptivity. In this work, we consider decision gaps expressed in fidelity and study possible separations between these problems and how adaptivity can help. We prove that certification with respect to fidelity for a state σσσ of rank rrr does not benefit from adaptivity and requires Θ~(r3/2/ε)\widetildeΘ(r^{3/2}/\varepsilon)Θ(r3/2/ε) samples. For equivalence testing and independence testing, we provide adaptive algorithms using O~(min⁡{d3/2/ε2,d9/4/ε})\widetilde{O}(\min\{d^{3/2}/\varepsilon^2,d^{9/4}/\varepsilon\})O(min{d3/2/ε2,d9/4/ε}) and O~(min⁡{(dAdC)3/2/ε2,dA9/4dC3/4/ε})\widetilde{O}(\min\{(d_Ad_C)^{3/2}/\varepsilon^2,d_A^{9/4}d_C^{3/4}/\varepsilon\})O(min{(dA​dC​)3/2/ε2,dA9/4​dC3/4​/ε}) samples, for dA≥dCd_A\geq d_CdA​≥dC​, respectively. Our main technique is a framework that uses partial learning and a reduction to testing in ℓ2\ell_2ℓ2​-distance, adapted from the distribution testing literature. We show that adaptivity matters for equivalence testing in fidelity by proving that Ω~(1/ε2)\widetildeΩ(1/\varepsilon^2)Ω(1/ε2) samples are necessary in the non-adaptive case even for qubits, showing a separation from certification.