Quantum advantage in decentralized control of POMDPs: A control-theoretic view of the Mermin-Peres square

Venkat Anantharam

Quantum advantage in decentralized control of POMDPs: A control-theoretic view of the Mermin-Peres square

T0 review · 0 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper proves a strict quantum advantage in a dynamical decentralized POMDP by embedding the Mermin-Peres square into the per-step reward.

desk verdict First genuine dynamical quantum advantage in a decentralized POMDP, built on the Mermin-Peres square; a couple of typos but the construction holds up. read the letter →

arxiv 2501.16690 v2 pith:RZZV65FC submitted 2025-01-28 math.OC cs.ITcs.SYeess.SYmath.ITquant-ph

Venkat Anantharam This is my paper · ORCID

classification math.OCcs.ITcs.SYeess.SYmath.ITquant-ph MSC 90C4093E2081P40

keywords decentralizedPOMDPquantumadvantageMermin-Peressquarecommonrandomnesslong-termaveragerewardentanglementcontrolteamtheory

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper establishes the existence of a quantum advantage in a genuinely dynamical decentralized control problem. It exhibits a two-agent partially observable Markov decision process in which every classical strategy—even one with unlimited common randomness—has long-term average reward at most $1-2\delta$, while a strategy that consumes two fresh entangled qubit pairs at each step attains average reward exactly $1$. The advantage comes from a control-theoretic reinterpretation of the Mermin-Peres square, a $3\times 3$ array of qubit measurements whose row and column correlations are quantum-mechanically perfect but classically impossible. The paper also gives a companion example where a one-shot quantum advantage exists but the dynamical advantage disappears, showing that static advantage does not automatically survive in a dynamic setting.

What carries the argument

The load-bearing object is the Mermin-Peres square, a $3\times 3$ array of tensor products of Pauli matrices acting on two qubits. In each row the three entries commute, in each column the three entries commute, the product of the three entries in any row is the identity while the product in any column is minus the identity, and the matching row/column measurement outcomes multiply to $1$ when the two agents share two EPR pairs $\frac{1}{\sqrt2}(\lvert00\rangle+\lvert11\rangle)$. The paper's quantum strategy uses two fresh independent EPR pairs per time step: Alice measures her halves according to the row indexed by her observation $X_n$, Bob measures his halves according to the column indexed by $Y_n$, and the perfect correlation property turns the reward into $1$ at every step.

What would settle it

Run the prescribed Mermin-Peres strategy on a quantum device or simulator with controlled depolarizing noise level $p$ on each EPR pair and a fixed kernel satisfying the lower-bound condition; if the observed long-term average reward is not strictly above $1-2\delta$ for some $p>0$, the strict advantage claim fails for that noise level. More directly, any observed instance where Alice's and Bob's matching products differ at the same $(X_n,Y_n)$ disproves the exact correlation.

Watch

Extended reading notes

Core claim

The central claim is that in the specific decentralized POMDP with state space $\{1,2,3\}^2$, action sets $U=\{u:\{1,2,3\}\to\{\pm1\}:\prod_l u(l)=1\}$ and $V=\{v:\{1,2,3\}\to\{\pm1\}:\prod_k v(k)=-1\}$, reward $r(i,j,u,v)=u(j)v(i)$, and any transition kernel whose every entry exceeds $\delta<1/9$, the classical value under adapted strategies with common randomness obeys $\limsup_{N\to\infty} \frac{1}{N}\sum_{n=0}^{N-1} \mathbb{E}[r(X_n,Y_n,U_n,V_n)]\le 1-2\delta$. If instead each time step the agents receive two independent EPR pairs and Alice measures the row of the Mermin-Peres square indexed by $X_n$ while Bob measures the column indexed by $Y_n$, then $U_n^{(Y_n)}(X_n)V_n^{(X_n)}(Y_n)=1$ pointwise, so the average reward is exactly $1$. Since $1>1-2\delta$, this is a strict quantum advantage in the decentralized control of POMDPs.

Load-bearing premise

The load-bearing premise is that the two EPR pairs delivered at each time step are perfectly noiseless and independent, and that the Mermin-Peres measurements are ideal, mutually commuting projections; the paper states the resulting pointwise identity $U_n^{(Y_n)}(X_n)V_n^{(X_n)}(Y_n)=1$ can be checked, but does not prove it in the text.

Editorial extensions

If this is right

For any transition kernel satisfying the uniform lower bound $\delta<1/9$, the classical ceiling $1-2\delta$ applies, so the quantum strategy's reward of $1$ beats it by a positive margin $2\delta$.
A fixed rate of two fresh EPR pairs per time step suffices; the advantage does not require a growing quantum memory or pre-shared entangled states of increasing size.
The classical upper bound holds even for relaxed strategies in which each agent sees the other's past observations, so the quantum advantage survives considerable information leakage between the agents.
One-shot quantum advantage does not imply dynamical quantum advantage: in the periodic-walk example of Section 4, classical strategies learn the other agent's current observation after two steps and attain reward $1$, despite the static advantage of $1$ versus at most $7/9$.
The construction works for an arbitrary Markov kernel subject only to the lower-bound condition, so the qualitative conclusion does not depend on a specially tailored transition law.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Although the paper stays with one concrete example, the mechanism indicates that the quantum advantage is driven by the local information structure (each agent sees only one component of the state) rather than by the specific transition kernel, since the kernel enters only through the uniform lower bound $\delta$.
A natural extension is to replace the ideal EPR pairs with depolarized or amplitude-damped entangled states and compute the long-term reward as a function of noise; the noise level at which the reward crosses $1-2\delta$ would quantify how much imperfection destroys the advantage.
The Section 4 example suggests a design principle: dynamical quantum advantage disappears when the state dynamics quickly reveal each agent's current observation to the other, so constructing POMDPs with persistent hidden state components may be the right route to robust dynamical advantages.
One could also ask whether a single shared EPR pair reused across time via quantum memory would preserve the advantage at a lower entanglement rate; the paper's strategy uses two fresh pairs per step, leaving the memory-versus-rate trade-off open.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this paper delivers what it promises. The new thing is not the Mermin-Peres square itself but the embedding: a two-agent decentralized POMDP with long-term average reward where entangled pairs supplied each step strictly beat any classical strategy with common randomness. That is a real extension of the known one-shot quantum advantage in team problems, and the Section 4 example showing one-shot advantage can vanish dynamically is a useful caution. The classical upper bound is proved by a clean coupling argument: conditional on the common past, every pair (i,j) has probability at least delta, so the per-stage reward is at most 1-2delta; the limsup then cannot exceed that. The quantum strategy achieves reward 1 pointwise, so the gap is explicit.

I checked the two load-bearing ingredients. The Mermin-Peres correlation is asserted "it can be checked" rather than derived in the text, but it is standard and correct: on the two EPR pairs, each cell's joint observable has eigenvalue +1, so Alice's and Bob's outcomes multiply to 1. The classical upper bound's proof does have a blemish: Corollary 2 in Appendix A states Bob's product constraint as +1 instead of -1, which makes the lemma false as written. On reading, though, this is clearly a typo—Corollary 1 and the main text use the correct -1, and the proof in Section 3.1 invokes Corollary 1 in the conditional step. I also noticed a small algebraic shorthand slip in Lemma 1's proof; the lemma itself is true.

Soft spots are modest. The model assumes two fresh, noiseless EPR pairs delivered every time step; the quantum advantage obviously degrades with noise, and the paper doesn't quantify how much. That is fine for an existence result. The paper restricts attention to a simple measurement-per-step strategy rather than the most general adaptive quantum controller; the authors acknowledge this in the closing remarks. For the point being made—there exists a dynamical quantum advantage—the restricted strategy suffices.

I agree with the reader's conditional verdict, but I'd go a bit further: the remaining issues are typographical. The paper is self-contained, honest about what is standard versus new, and the citation pattern looks proper; the prior work on quantum advantage in static team problems is cited, and the new contribution is clearly flagged. Who is this for? Control theorists who haven't followed the quantum information literature, and quantum information people curious about control. Both groups can read it without a lot of background. It is short, clear, and correct after minor fixes. I would send it out for review; a good referee will catch the Corollary 2 sign and ask for the Mermin-Peres check to be written out, but nothing deeper is needed.

Referee Report

0 major / 5 minor

Summary. The paper considers a decentralized POMDP with two agents, Alice and Bob, who cooperate to maximize long-term average reward. It specializes to X=Y={1,2,3}, with Alice's actions being sign vectors of product +1 and Bob's sign vectors of product -1, reward r(i,j,u,v)=u(j)v(i), and transition probabilities bounded below by delta < 1/9. The main result is that any classical strategy, even with common randomness, has limsup average reward at most 1-2*delta, while an entanglement-assisted strategy using two fresh EPR pairs per step and the Mermin-Peres-square measurements achieves average reward 1. The paper also constructs a second example, with deterministic periodic transitions, where a one-shot quantum advantage exists but the dynamic advantage disappears.

Significance. The main contribution is a clean separation between classical common-randomness strategies and entanglement-assisted strategies in a dynamic setting, which is new relative to the existing one-shot team-theory results. The classical upper bound is elementary and self-contained, and the quantum lower bound is imported from the standard Mermin-Peres square; no ad hoc assumptions or fitted parameters appear beyond the single threshold delta. The Section 4 example is a useful caution against inferring dynamic advantage from one-shot advantage. If the presentation issues below are addressed, the paper will be a valuable proof-of-concept for the control community.

minor comments (5)

[Appendix A, Corollary 2] The assumption on v is written as product_{k=1}^{3} v^{(k)}(Y,Z) = 1, but for the lemma to be true, and for consistency with Corollary 1 and the action set V, this must be -1. As printed, taking u = v = (1,1,1) gives P(u^{(Y)}v^{(X)} = -1) = 0, contradicting the conclusion. Section 4 explicitly cites Corollary 2 for the 7/9 one-shot upper bound, so this sign must be corrected and the citation rechecked.
[Section 3.1, eq. (8)] The text says the bound is an immediate consequence of Corollary 2, but the displayed step (c) invokes Corollary 1. Please align the cross-reference. Also, the claim is stated for every n >= 0, but for n = 0 the conditioning variable Z_0 contains no past observations; the proof as written needs a separate (or omitted) treatment of n = 0, which is harmless because that term is O(1/N) in the limsup.
[Section 3.2 / Appendix C.2] The key identity U_n^{(Y_n)} V_n^{(X_n)} = 1 is stated as 'it can be checked.' Since this identity is the entire mechanism behind the quantum lower bound, and the appendix promises rigorous background, please include the explicit calculation, for example by showing that each cell of the Mermin-Peres square has eigenvalue +1 on the state |Phi+> tensor |Phi+> with the stated assignments of row and column measurements.
[Appendix A, Lemma 1 and Corollary 1] The proofs contain an algebraic shorthand: the product over all i,j of u^{(j)}(i)v^{(i)}(j) is not simply (product_l u(l))(product_k v(k)); the correct computation involves the third power of each product. The statements are correct, but the written derivation could confuse readers and should be rewritten.
[Throughout] There are several typos: in eq. (2), 'Y0:n.W0:n' should be 'Y0:n, W0:n'; in Section 4 the definition of U has a missing brace and an extra parenthesis in 'u=(u1), u(2), u(3))'; 'indepedent' should be 'independent'; and the dedication line has a stray space in 'Pravin P . V araiya'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the classical bound is proved from model primitives and the quantum advantage is imported from the externally established Mermin-Peres square.

full rationale

The paper's central derivation is self-contained against external benchmarks and does not reduce any prediction to fitted inputs. The classical upper bound in Section 3.1 is proved from the model primitives: using the relaxation to strategies with access to the other agent's past observations and the conditional transition lower bound q(...) > delta, the argument conditions on Z_n = (X_{0:n-1}, Y_{0:n-1}, W_{0:n}) and applies Corollary 1 to conclude P(r_n = -1 | Z_n) >= delta, hence E[r_n | Z_n] <= 1 - 2*delta; the limsup bound follows. The Mermin-Peres quantum strategy is not derived from the paper's own conclusions; it is imported as a standard, externally established correlation (references [2], [11], [13], [15]), and Appendix C.2 states the required property. There are no fitted parameters, no statistically forced predictions, and no load-bearing self-citations. The only textual issues are a sign typo in Corollary 2 (Bob's product constraint written as +1 instead of -1) and an algebraic shorthand in Lemma 1's proof, but neither is load-bearing because the main proof invokes Corollary 1 directly and the corrected statement is immediate. Therefore no circular step can be exhibited.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

This is an existence proof built on standard quantum information rather than on new postulates. The only hand-chosen parameter is the transition lower bound delta, which is not fitted to data. The key external input is the Mermin-Peres correlation, imported from the quantum information literature. No new entities are introduced.

free parameters (1)

delta = delta in (0, 1/9), chosen by hand
Lower bound on every one-step transition probability in the example POMDP (eqn. (4)); the classical upper bound in eqn. (6) is 1 - 2*delta. Since the result holds for every delta in this range, it is a design parameter rather than a fitted constant.

assumptions (6)

domain assumption Standard quantum mechanics: states are density matrices, measurements are POVMs/PVMs, and the Born rule gives outcome probabilities.
Introduced in Appendix B; needed for the entangled resource and the Mermin-Peres measurements.
domain assumption The two EPR pairs, each in state (|00>+|11>)/sqrt(2), can be supplied to Alice and Bob at each time step and are independent across time.
Hardware or resource assumption; the quantum strategy in Section 3.2 requires fresh entangled pairs each round.
domain assumption Mermin-Peres square properties: row and column entries commute, row products equal +I, column products equal -I, and the intersection correlation a_{ij} b_{ij} = 1 holds on the two-EPR-pair state.
External result from quantum information, cited to [11,13,15], used in Section 3.2 and Appendix C.2; asserted but not fully derived in the text.
domain assumption The transition kernel satisfies q(.,.|x,y,u,v) > delta for all arguments, with delta < 1/9.
Modeling assumption of the example (eqn. (4)); it ensures the conditional law of (Xn,Yn) given history has each pair with probability at least delta.
domain assumption The common randomness sequence (W_n) is independent and causally available to both agents.
Modeling assumption in Section 2; needed so the classical upper bound covers the strongest classical resource.
standard math Private randomization can be subsumed by common randomness: a strategy with private seeds is a special case of a common-randomness strategy in which agents ignore irrelevant coordinates.
Used in Section 2 to state that proving the upper bound under common randomness covers all classical strategies.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Quantum advantage in decentralized control of POMDPs: A control-theoretic view of the Mermin-Peres square." pith.science (2026). https://pith.science/paper/RZZV65FC

@misc{pith2026250116690,
  author       = {Pith},
  title        = {Pith review of: Quantum advantage in decentralized control of POMDPs: A control-theoretic view of the Mermin-Peres square},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RZZV65FC}},
  note         = {Machine review of arXiv:2501.16690}
}

read the original abstract

Consider a decentralized partially-observed Markov decision problem (POMDP) with multiple cooperative agents aiming to maximize a long-term-average reward criterion. We observe that the availability, at a fixed rate, of entangled states of a product quantum system between the agents, where each agent has access to one of the component systems, can result in strictly improved performance even compared to the scenario where common randomness is provided to the agents, i.e. there is a quantum advantage in decentralized control. This observation comes from a simple reinterpretation of the conclusions of the well-known Mermin-Peres square, which underpins the Mermin-Peres game. While quantum advantage has been demonstrated earlier in one-shot team problems of this kind, it is notable that there are examples where there is a quantum advantage for the one-shot criterion but it disappears in the dynamical scenario. The presence of a quantum advantage in dynamical scenarios is thus seen to be a novel finding relative to the current state of knowledge about the achievable performance in decentralized control problems. This paper is dedicated to the memory of Pravin P. Varaiya.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages

[1]

Common randomness and dis- tributed control: A counterexample

Venkat Anantharam and Vivek Borkar. “Common randomness and dis- tributed control: A counterexample”, Systems and Control Letters , V ol. 56, 2007, pp. 568-572

work page 2007
[2]

Quantum mysteries revisited again

P. K. Aravind. “Quantum mysteries revisited again”, American Journal of Physics, V ol. 72, No. 10, 2004, pp. 1303-1307. 22

work page 2004
[3]

From Bell Inequalities to Tsirelson's Theorem: A Survey

David Avis, Sonoko Moriyama, and Masaki Owari. “From Bell inequalities to Tsirelson’s theorem: A survey”, arXiv:0812.4887 [quant-ph], 2008

work page Pith review arXiv 2008
[4]

The theory of teams: a selective annotated bibliography

Tamer Bas ¸ar and Rajesh Bansal. “The theory of teams: a selective annotated bibliography”, Lecture Notes in Control and Information Sciences, V ol. 119, Springer, 1989

work page 1989
[5]

The Quantum Advantage in Decentralized Control

Shashank A. Deshpande and Ankur A. Kulkarni. “The quantum advantage in decentralized control”, arXiv:2207.12075 [eess.SY]

work page Pith review arXiv
[6]

Beyond common random- ness: Quantum resources in decentralized control

Shashank A. Deshpande and Ankur A. Kulkarni. “Beyond common random- ness: Quantum resources in decentralized control”, 62nd IEEE Conference on Decision and Control, 2023, pp. 5906-5911

work page 2023
[7]

Deshpande and Ankur A

Shashank A. Deshpande and Ankur A. Kulkarni. “The quantum advantage in binary teams and the coordination dilemma: Part I, arXiv:2307.01762 [eess.SY]

arXiv
[8]

Deshpande and Ankur A

Shashank A. Deshpande and Ankur A. Kulkarni. “The quantum advantage in binary teams and the coordination dilemma: Part II, arXiv:2307.01766 [eess.SY]

work page arXiv

Show all 19 references

[9]

Geometry of the set of quantum correlations

Koon Tong Goh, Jedrzej Kaniewski, Elie Wolfe, Tam ´as V ´ertesi, Xingyao Wu, Yu Cai, Yeong-Cherng Liang, and Valerio Scarani. “Geometry of the set of quantum correlations”, arXiv:1710.05892 [quant-ph]

arXiv
[10]

Zero-sum games involving teams against teams: Existence of equilibria, and comparison and regularity in in- formation

Ian Hogeboom-Burr and Serdar Y ¨uksel. “Zero-sum games involving teams against teams: Existence of equilibria, and comparison and regularity in in- formation”, Systems and Control Letters, V ol. 172, 2023, No. 105454

2023
[11]

Alexander S. Holevo. Quantum Systems, Channels, Information: A Mathe- matical Introduction, De Gruyter, 2012

2012
[12]

Common information belief based dynamic programs for stochastic zero-sum games with compet- ing teams

Dhruva Kartik, Ashutosh Nayyar, and Urbashi Mitra. “Common information belief based dynamic programs for stochastic zero-sum games with compet- ing teams”, American Control Conference, 2022, pp. 605-612

2022
[13]

Simple unified form for the major no-hidden-variables theorems

N. David Mermin. “Simple unified form for the major no-hidden-variables theorems”, Physical Review Letters, V ol. 65, No. 27, 1990, pp. 3373-3376. 23

1990
[14]

Quantum advantage in Bayesian games

Igal Milchtaich. “Quantum advantage in Bayesian games”, Working Paper No. 2023-05, Department of Economics, Bar-Ilan University, https://hdl.handle.net/10419/279453

2023
[15]

Incompatible results of quantum measurements

Asher Peres. “Incompatible results of quantum measurements”, Physics Let- ters A, V ol. 151, Nos. 3 and 4, 1990, pp. 107-108

1990
[16]

Geometry of information structures, strategic measures and associated stochastic control topologies

Naci Saldi and Serdar Y¨uksel. “Geometry of information structures, strategic measures and associated stochastic control topologies”, Probability Surveys, V ol. 19, 2022, pp. 450-532

2022
[17]

Nash equilibria for exchange- able team-against-team games, their mean-field limit, and the role of com- mon randomness

Sina Sanjari, Naci Saldi, and Serdar Y ¨uksel. “Nash equilibria for exchange- able team-against-team games, their mean-field limit, and the role of com- mon randomness”, SIAM Journal on Control and Optimization, V ol. 62, No. 3, pp. 1437-1464
[18]

The Theory of Quantum Information , Cambridge University Press, 2018

John Watrous. The Theory of Quantum Information , Cambridge University Press, 2018

2018
[19]

Serdar Y ¨uksel and Tamer Bas ¸ar.Stochastic Teams, Games, and Control un- der Information Constraints, Birkh¨auser, 2024. 24

2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Common randomness and dis- tributed control: A counterexample

Venkat Anantharam and Vivek Borkar. “Common randomness and dis- tributed control: A counterexample”, Systems and Control Letters , V ol. 56, 2007, pp. 568-572

work page 2007

[2] [2]

Quantum mysteries revisited again

P. K. Aravind. “Quantum mysteries revisited again”, American Journal of Physics, V ol. 72, No. 10, 2004, pp. 1303-1307. 22

work page 2004

[3] [3]

From Bell Inequalities to Tsirelson's Theorem: A Survey

David Avis, Sonoko Moriyama, and Masaki Owari. “From Bell inequalities to Tsirelson’s theorem: A survey”, arXiv:0812.4887 [quant-ph], 2008

work page Pith review arXiv 2008

[4] [4]

The theory of teams: a selective annotated bibliography

Tamer Bas ¸ar and Rajesh Bansal. “The theory of teams: a selective annotated bibliography”, Lecture Notes in Control and Information Sciences, V ol. 119, Springer, 1989

work page 1989

[5] [5]

The Quantum Advantage in Decentralized Control

Shashank A. Deshpande and Ankur A. Kulkarni. “The quantum advantage in decentralized control”, arXiv:2207.12075 [eess.SY]

work page Pith review arXiv

[6] [6]

Beyond common random- ness: Quantum resources in decentralized control

Shashank A. Deshpande and Ankur A. Kulkarni. “Beyond common random- ness: Quantum resources in decentralized control”, 62nd IEEE Conference on Decision and Control, 2023, pp. 5906-5911

work page 2023

[7] [7]

Deshpande and Ankur A

Shashank A. Deshpande and Ankur A. Kulkarni. “The quantum advantage in binary teams and the coordination dilemma: Part I, arXiv:2307.01762 [eess.SY]

arXiv

[8] [8]

Deshpande and Ankur A

Shashank A. Deshpande and Ankur A. Kulkarni. “The quantum advantage in binary teams and the coordination dilemma: Part II, arXiv:2307.01766 [eess.SY]

work page arXiv