Foundations of quantum mechanics · Theory

Quantifying the irrealism required by quantum random access codes

Pedro R. Dieguez

arXiv:2611.26340v1 ↗PDF ↗Submitted 19 November 2026

Question
In a \(2_d\to1\) random access code Alice packs two \(d\)-valued symbols into one \(d\)-level system and Bob, asked for one of them, measures one of two decoding observables. Quantum codes beat the classical success probability \(P_C(d)=\tfrac12(1+1/d)\). A criterion of Bilobran and Angelo makes the EPR notion of an element of reality operational: an observable is real for a preparation if measuring it without revealing the outcome leaves the preparation unchanged, and its irrealism is the entropy such a measurement adds. How much irrealism must Bob's decoding observables carry for a code to reach a given success probability, with and without shared entanglement?
Result
For prepare-and-measure codes decoded in mutually unbiased bases, the least total irrealism of the two decoders is found exactly in every dimension, and it is attained only by Gibbs states of the decoding operator \(\Pi_i+\Pi'_j\). For qubits it is \(2[h_2(P)-h_2(\tfrac12+(2P-1)/\sqrt2)]\) bits, which is 0.421 bits at the classical value \(3/4\) and 1.202 bits at the optimum. Either decoder on its own can stay an element of reality exactly up to \(P_C(d)\). For entanglement-assisted qubit codes the frontier is \(2[1-h_2(\tfrac12+(2P-1)/\sqrt2)]\), for any quantum system on the sender's side. It exceeds the prepare-and-measure frontier by exactly \(2[1-h_2(P)]\), which is the realism-based nonlocality of the contexts that the optimal code uses. With rank-one projective sender measurements, a success probability above 0.8145 certifies a minimum realism-based nonlocality of the shared state.

The paper in detail

A random access code compresses more data than the channel can carry and lets the receiver choose afterwards which part to read. In the \(2_d\to1\) version, Alice receives \((i,j)\in\{0,\dots,d-1\}^2\) and sends a \(d\)-level state \(\rho_{ij}\). Bob reads \(i\) by measuring \(B\), with projectors \(\Pi_k\), or reads \(j\) by measuring \(B'\), with projectors \(\Pi'_k\). The success probability is the average of \(s_{ij}=\tfrac12\operatorname{tr}[\rho_{ij}(\Pi_i+\Pi'_j)]\). Sending one symbol classically gives \(P_C(d)=\tfrac12(1+1/d)\), and no classical \(d\)-level message does better. With mutually unbiased decoders the best quantum value is \(P_Q(d)=\tfrac12(1+1/\sqrt d)\), the largest eigenvalue of \(\Pi_i+\Pi'_j\) divided by two. In the entanglement-assisted version, Alice measures her half of a shared state and sends one bit. The success probability is then \(\tfrac12+\mathcal S/8\), with \(\mathcal S\) the CHSH value.

The irrealism of an observable \(B\) for a preparation \(\rho\) is \(\mathfrak I_B(\rho)=S(\Phi_B(\rho))-S(\rho)\), where \(\Phi_B\) is the unrevealed measurement of \(B\) (dephasing in its eigenbasis). It is the relative-entropy distance to the nearest preparation in which \(B\) has a definite value, and it vanishes exactly when \(B\) is such an element of reality. For a single system it is the relative entropy of coherence in the eigenbasis of \(B\). For a bipartite state it adds the mutual information that the unrevealed measurement destroys, so it captures in one quantity the two resources a code can use: coherence of a transmitted system, and correlations with a distant party. The paper's question is the frontier \(\mathcal C_{\min}(P)\): the least value of \(\mathfrak I_B+\mathfrak I_{B'}\), averaged over messages, among codes whose success probability reaches \(P\).

For prepare-and-measure codes, two short arguments fix the qualitative picture. If \(B\) is real for every transmitted state, Alice could send its eigenvalue classically and Bob could simulate \(B'\), so the code cannot beat \(P_C(d)\). Conversely, with compatible decoders a code with no irrealism reaches \(P_C(d)\). The exact frontier for mutually unbiased decoders is reached by the Gibbs code \(\rho_{ij}(\theta)\propto e^{\theta(\Pi_i+\Pi'_j)}\). It is strictly convex, smooth through the classical value, and zero only at the guessing value \(1/d\). Every code also obeys the upper bound \(2[h_2(P)+(1-P)\log_2(d-1)]\). The two bounds meet at \(P_Q\), where the success probability alone fixes the irrealism, \(\mathcal C_Q(d)\simeq\log_2d+2\) bits. When the decoders are optimised too, the least irrealism is zero up to \(P_C(d)\) and positive beyond it. For qubits, slightly misaligned decoders are cheaper than mutually unbiased ones below \(P_*\simeq0.833\), but they cannot reach high success probabilities.

In the entanglement-assisted code, Bob's reduced state is maximally mixed in the optimal codes, so all of his decoders' irrealism sits in correlations with Alice. A tight trade-off between the two irrealisms and the CHSH value, and the closed-form frontier, hold for any system on Alice's side. Once Alice measures, Bob's conditional states form a prepare-and-measure code with the same score, and relative to them his decoders are more real. The gain is the realism-based nonlocality of the contexts formed by Alice's and Bob's observables, and for the optimal code it accounts exactly for the gap between the two frontiers. Open problems listed: decoders for \(d>2\), untrusted or noisy decoders, \(n\to1\) codes, constraints on the full statistics, and entanglement-assisted codes with qudit receivers or sender settings that depend on both bits.

Transferable methods

  1. An excess identity that turns a stationary point into the global minimum

    Minimising \(\mathfrak I_B+\mathfrak I_{B'}\) at a fixed decoding probability is a non-convex-looking problem: it subtracts twice a concave entropy from two dephased entropies. The paper's device is an identity. If the unrevealed measurements map a candidate state into the span of the identity and the constrained observable, the excess irrealism of any competitor is a difference of relative entropies, which data processing makes non-negative. The Gibbs state of the decoding operator has this closure property for mutually unbiased bases, so it is optimal, and the equality conditions make it unique.

    Write \[F(\sigma)=\mathfrak I_B(\sigma)+\mathfrak I_{B'}(\sigma)=S(\Phi_B\sigma)+S(\Phi_{B'}\sigma)-2S(\sigma).\] Two identities hold for any full-rank \(\rho\). The first is \(-S(\sigma)=D(\sigma\|\rho)+\operatorname{tr}\sigma\log\rho\). The second is \(S(\Phi\sigma)=-D(\Phi\sigma\|\Phi\rho)-\operatorname{tr}\sigma\log\Phi\rho\), which uses that \(\log\Phi\rho\) is invariant under \(\Phi\). Adding them gives (editorial derivation of Eq. (19)) \[F(\sigma)=\sum_{\Phi=\Phi_B,\Phi_{B'}}\bigl[D(\sigma\|\rho)-D(\Phi\sigma\|\Phi\rho)\bigr]+\operatorname{tr}\bigl[\sigma\,(2\log\rho-\log\Phi_B\rho-\log\Phi_{B'}\rho)\bigr].\] Suppose the operator in the last trace is \(\kappa I+\zeta M\), the closure property (18). Then the last term takes the same value for every \(\sigma\) with \(\operatorname{tr}\sigma M=\operatorname{tr}\rho M\). Setting \(\sigma=\rho\) shows that \(F(\sigma)-F(\rho)\) equals the bracket, which is non-negative because unrevealed measurements cannot increase distinguishability.

    For mutually unbiased bases take \(M=\Pi_i+\Pi'_j\), with eigenvalues \(1\pm d^{-1/2}\) and a \((d-2)\)-dimensional kernel. Then \(e^{\theta M}\) is a quadratic polynomial in \(M\). Unbiasedness gives \(\langle b_k|M|b_k\rangle=\langle b_k|M^2|b_k\rangle=1/d\) for every \(k\ne i\), so every diagonal element of \(\rho_\theta=e^{\theta M}/Z_d\) other than the \(i\)-th is equal: \[\Phi_B(\rho_\theta)=P\,\Pi_i+\frac{1-P}{d-1}(I-\Pi_i).\] The same holds for \(B'\) with \(\Pi'_j\). Hence \(2\log\rho_\theta-\log\Phi_B\rho_\theta-\log\Phi_{B'}\rho_\theta\) is \([2\theta-\ln\tfrac{(d-1)P}{1-P}]M\) plus a multiple of \(I\), and closure holds. In the equality case \(\rho^{-1/2}\sigma\rho^{-1/2}\) must commute with both projector families, which forces \(\sigma=\rho\) because every overlap \(\langle b_k|b'_l\rangle\) is non-zero.

    The frontier follows from the entropies. The dephased states have entropy \(f_d(P)=h_2(P)+(1-P)\log_2(d-1)\), and \(S(\rho_\theta)=\log_2Z_d-2\theta P/\ln2\) because \(\langle M\rangle=2P\). Therefore \[\mathcal C^{(d)}_{\min}(P)=2\Bigl[f_d(P)-\log_2Z_d(\theta)+\frac{2\theta P}{\ln2}\Bigr],\qquad Z_d=d-2+2e^{\theta}\cosh\frac{\theta}{\sqrt d},\] with \(\theta\) fixed by the success probability. For qubits this reduces to \(2[h_2(P)-h_2(\tfrac12+(2P-1)/\sqrt2)]\). The editorial check above confirmed numerically that no qutrit state beats it.

    How to reuse it

    Any problem that minimises a combination of dephasing-induced entropy gains under linear constraints can be attacked the same way: guess a maximum-entropy candidate, check whether the dephasing maps send its logarithm into the constraint span, and read off optimality and uniqueness. Candidate uses include coherence costs of state discrimination with unbiased measurements, states that saturate entropic uncertainty relations with a fixed score, and certified randomness from prepare-and-measure tests in higher dimensions.

  2. Irrealism as an adversary's uncertainty, and the steering accounting

    Two identities connect the resource to operational quantities. Irrealism equals the conditional entropy of the outcome given a system that purifies the preparation, so a frontier is a randomness bound. In the entanglement-assisted code, Alice's measurement splits the irrealism of Bob's decoders into two parts. One is the irrealism they keep relative to her outcome, which forms a prepare-and-measure code. The other is the irrealism her outcome removes, which is the realism-based nonlocality of the context. Comparing the two frontiers is then bookkeeping rather than a new optimisation.

    Let \(E\) purify \(\rho_{AB}\). Measuring a rank-one \(B\) on the pure state of \(ABE\) leaves \(AE\) pure for each outcome \(k\). Hence \(H(BE)=H(B)+\sum_kp_kS(\rho_A^{(k)})=S(\Phi_B\rho_{AB})\), while \(H(E)=S(\rho_{AB})\), so \[\mathfrak I_B(\rho_{AB})=H(B|E).\] For prepare-and-measure codes, the purification also carries a copy of Alice's input, and the total irrealism is \(H(b|EX)+H(b'|EX)\). The frontier is therefore the least uncertainty, about Bob's outcomes, of an adversary who knows the input and holds the purification, given only the success probability.

    Let \(\mathcal M_z\) be Alice's measurement with the outcome recorded but not revealed. Because \(\mathcal M_z\) acts on Alice's side and commutes with \(\Phi_B\), data processing gives a non-negative remainder in \[\mathfrak I_B(\rho_{AB})=\sum_ap(a|z)\,\mathfrak I_B(\rho_B^{a|z})+\Delta_B^z,\qquad\Delta_B^z\ge0.\] For a rank-one projective measurement, \(\Delta_B^z=I(B{:}a|E)\), which is also the realism-based nonlocality \(\eta_{B|A_z}\). Averaging over \(z\) and applying the prepare-and-measure frontier to the steered states gives \(\mathcal C_{\rm EA}\ge\mathcal C^{(2)}_{\min}(P)+\bar\Delta\). For the optimal resource each of the four \(\Delta\) terms equals \(1-h_2(P)\), so the two frontiers differ by exactly \(2[1-h_2(P)]\).

    The certificate combines the entanglement-assisted frontier from below with the upper bound \(2h_2(P)\) on any steered qubit code. Each \(\Delta\) is at most \(\eta_{\max}\), so \[\eta_{\max}(\rho_{AB})\ \ge\ \varphi\bigl(\sqrt2(2P-1)\bigr)-h_2(P),\qquad\varphi(t)=1-h_2\bigl(\tfrac{1+t}2\bigr).\] The right-hand side is positive for \(P>0.8145\) and reaches 0.399 bits at the Tsirelson point; the editorial check reproduced both numbers.

    How to reuse it

    Whenever a quantity defined by dephasing can be rewritten as a conditional entropy against a purifying adversary, a resource bound becomes a randomness bound. Splitting a bipartite cost into "what remains relative to the other party's outcome" plus "what that outcome removes" is the same accounting that underlies one-sided device-independent key distribution and remote state preparation. It can be reused to compare entanglement-assisted and prepare-and-measure versions of other communication tasks.

Source: Sec. II.2 (irrealism, Eqs. (1)–(8)); Sec. III.2.1 (excess identity, Eqs. (18)–(23), purification identity, Eq. (29)); Sec. III.3 (entanglement-assisted frontier and accounting, Eqs. (35)–(41)). Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. The paper turns an interpretive notion, the absence of definite values, into an exact resource count for a standard communication task. Irrealism equals the conditional entropy of the decoding outcome given a system that purifies the preparation. The frontier is therefore also an exact, dimension-general lower bound on the randomness that a code's score alone certifies against such an adversary, when Bob's decoders are trusted and mutually unbiased. The proof device is an identity that turns a stationary point into the unique global minimum, and it applies to any weighted sum of relative entropies of coherence. It is therefore reusable in other coherence-minimisation problems with mutually unbiased bases.

Limitations. The decoders are trusted and rank-one projective. The main frontier assumes they are exactly mutually unbiased, and for \(d>2\) the role of the decoders is open. For qubits, that mutually unbiased decoders are optimal above \(P_*\simeq0.833\) is shown only locally, with numerics suggesting it holds globally. Whether the upper bound obeyed by every code is attained for \(d>2\) below the optimum is open; numerics support it for Fourier pairs with \(d\le7\). The entanglement-assisted results assume a qubit receiver, two-outcome sender measurements, and a sender setting that depends only on \(x_0\oplus x_1\). The randomness reading assumes collective attacks on independent rounds. The nonlocality certificate is not tight. This is a single-author theoretical paper, and it contains no AI-use statement.

Description based on the full arXiv HTML of v1, read 2026-11-22: the whole main text (irrealism and the information–reality exclusion, the prepare-and-measure frontier with Eqs. (14)–(30), the dependence on the decoders, the entanglement-assisted frontier with Eqs. (31)–(41), and the conclusion); the proofs in Appendices A–F were not reproduced. Editorial checks: Eq. (23) gives 0.421 and 0.761 bits at the classical value for \(d=2,3\) (exclusion bounds 0.377 and 0.667) and 1.202 and 1.911 bits at the optimum (0.798 and 1.259), with increments 0.78, 1.15 and 1.37 bits for \(d=2,3,4\), and it agrees with the qubit closed form (26). Constrained minimisation over all qutrit states, for the Fourier pair of bases at \(P=0.62\), 0.667, 0.72 and 0.76, reproduces the Gibbs value to five decimals. At \(P=0.8\) the single-observable minimum is 0.094 bits (exclusion 0.029), with 0.917 bits on the other observable, while the Gibbs code has 0.335 bits on each. For qubit decoders at angle \(\gamma\), per-message minimisation gives 0.134 bits at \(P=0.76\) (0.463 for mutually unbiased decoders), and the optimal angle reaches \(90^\circ\) between \(P=0.832\) and 0.834. The entanglement-assisted values 0.798, 0.8145 and 0.399, the chord crossing at 2.695 and the tangent point \(\mathcal S_t=2.776\) (\(P=0.847\)) also reproduce.