WorksPhysics and Astronomy
Certified Separation Between Coherent and Adaptive Temporal Processing of Quantum-Error-Correction Readout
Abstract
Repeated quantum-error-correction readout is normally converted into a classical time series: each syndrome pulse is measured on arrival and only the resulting classical record is retained. Nonorthogonal readout signals raise the question of whether preserving quantum information between rounds can improve the final recovery decision. We formulate this comparison as a loss-weighted state-discrimination problem. For an idealized two-round benchmark with quadrature phase-shift-keying (QPSK) signals, a receiver with a seven-dimensional persistent register has certified recovery-decision error below , whereas every separate adaptive receiver has error above . The certified separation is at least percentage points. The loss counts incorrect final error-mask decisions in the benchmark. In the strict architecture, where the counted register is the entire persistent state, four dimensions are necessary and sufficient to outperform the adaptive class. Certified separations also occur at three and four rounds with register dimensions and , compared with and for full-prefix storage. We relate memory dimension to the structure of the temporal measurement, give a sufficient memory bound for collective measurements, and quantify robustness to register dephasing and accumulated channel error.
Introduction
Repeated quantum error correction (QEC) is inherently a temporal inference problem. A single syndrome observation can be ambiguous because data faults and readout noise both affect the observed record, so practical decoders combine information across multiple extraction rounds. Recent work has further shown that retaining analog or soft readout information can improve on hard syndrome decisions [1, 2, 3]. These results make clear that not only the syndrome record, but also how that record is represented and retained, can affect the final recovery decision.
In the usual measurement-based architecture, each syndrome extraction produces a physical readout signal that is measured as it arrives. The resulting classical record can be retained in considerable detail: a decoder may keep analog values, assign likelihoods, model correlations across rounds, and use the complete history when selecting a recovery. Such processing can preserve much more information than a hard syndrome bit. It nevertheless begins after the arriving quantum signal has been measured. By the time temporal decoding takes place, the information available to the decoder is classical.
At the other end of the design space, coherent and measurement-free QEC architectures avoid the conventional measurement-and-feedback loop by retaining quantum information within a coherent controller or by implementing correction through autonomous dynamics [4, 5, 6, 7]. These approaches show that syndrome information need not always be converted immediately into a classical record, but they also change how error correction itself is implemented.
Here we consider an intermediate possibility. Before measurement, the finite-strength signals associated with distinct syndromes need not be orthogonal, so there is an operational distinction between remembering the past after measurement and preserving part of the past before measurement. We leave the syndrome-extraction process and the emitted readout signals unchanged, but alter when those signals must become classical. Rather than measuring every signal separately as it arrives, a finite quantum memory may retain information from earlier signals and interact with later ones, allowing some or all measurement to be deferred until later signals have arrived.
We therefore ask whether, for the same sequential readout ensemble and recovery objective, preserving quantum information across rounds can outperform every receiver that measures each signal separately and retains only classical information between rounds.
Answering this question requires more than comparing against a particular decoder. Failure of one classical algorithm could reflect the choice of decoder rather than a limitation imposed by measurement itself. We therefore define a deliberately broad separate adaptive class: each arriving signal may undergo an arbitrary quantum instrument, the receiver may retain an unrestricted classical transcript, and all later instruments may depend on that transcript. The only excluded resource is a quantum state carrying information from an earlier signal into a later round.
Direct optimization over this adaptive class is difficult. Every such receiver induces measurement operators that are separable across the temporal signal factors and therefore positive under partial transpose (PPT). This containment gives a tractable semidefinite-programming relaxation: a lower bound obtained over the larger PPT class is automatically a lower bound on the error attainable by any receiver in the separate adaptive class. Consequently, a feasible coherent receiver whose error lies below this bound is separated from the entire measure-and-feed-forward class, rather than from a particular decoder found numerically.
For the two-round QPSK benchmark studied here, a streaming receiver achieves a recovery-decision error of approximately 0.2421, while the certified adaptive lower bound is approximately 0.2514. The resulting separation is about , or 0.93 percentage points; the outward-rounded values are stated in Theorem 2. This gap certifies a separation between the two receiver classes under the same signal ensemble and recovery objective.
We next determine the persistent-memory requirements of this separation. We distinguish the broad coherent class, which may retain classical side information, from a strict fixed-memory architecture in which the counted register is the only persistent state. For the latter model, the two-round benchmark has the exact threshold , while . For each explicit receiver, the certified advantage persists up to a rigorously bounded dephasing strength on the persistent register. More generally, the entanglement-breaking analysis shows that any receiver exhibiting such an advantage must have at least one persistent boundary whose complete classically flagged instrument is not measure-and-prepare. The temporal-separability argument extends to arbitrary depth, and certified separations are obtained with at and at . These finite-depth results establish persistence of the effect beyond two rounds; asymptotic memory scaling remains open.
We also derive a diamond-norm accumulation bound that gives sufficient implementation-error budgets for the explicit streaming receivers. Separately, a stored-pulse model translates phase retention into approximate storage-lifetime requirements and compares their scale with experimentally reported component coherence times. This comparison provides a component-level feasibility check. The architecture also changes where latency enters the recovery pipeline: information that would otherwise be converted into a classical record between rounds may instead be carried forward coherently, replacing some inter-round measurement and feed-forward requirements with coherent storage and control. This suggests a latency–coherence tradeoff whose quantitative characterization requires a hardware-specific timing model.
Related work
a. Soft syndrome decoding. Experimental readout produces analog voltages or photon counts before thresholding, and a hard decision discards information about the confidence of that assignment. Soft and analog decoders retain part of this information and can lower logical error relative to hard-decision decoding [1, 2, 3]. These results establish that the representation of readout information affects QEC performance even when the syndrome-extraction circuit remains unchanged. However, soft decoding operates on information that has already been converted into a classical measurement record. Our work asks whether the measurement itself imposes a fundamental limitation, even when all available classical information is retained. To isolate this distinction, our separate adaptive class permits arbitrary local measurements, continuous outcomes, an unrestricted classical transcript, and unrestricted classical post-processing. We establish a certified separation from this entire class, demonstrating that, for the benchmark considered, coherent processing achieves a lower recovery error than any separate adaptive receiver, even with unrestricted classical memory and post-processing.
b. Coherent and measurement-free QEC. Coherent feedback and measurement-free error correction avoid the standard mid-circuit measurement loop [4, 5, 6, 7, 8]. In these architectures, a coherent controller may participate directly in syndrome extraction, feedback, or autonomous correction. Recent theory provides scalable fault-tolerant and noise-biased measurement-free constructions [6, 8], while trapped-ion experiments have demonstrated measurement-free logical operations [7]. A separate spatio-temporal framework describes adaptive error correction using sequential interventions [9]. Our work addresses a different question. Rather than modifying syndrome extraction or implementing coherent correction, we hold the extraction circuit, emitted pulse ensemble, trajectory prior, and recovery loss fixed. Only the processing of the emitted temporal signals changes: each pulse is either measured on arrival, with its outcome retained classically, or processed through a persistent quantum memory. This isolates the value of quantum memory at the receiver level and establishes a certified advantage without attributing the improvement to changes in the underlying error-correction circuit.
c. Restricted quantum measurements. Minimum-error discrimination admits a primal-dual semidefinite programming (SDP) formulation [10, 11, 12]. For multipartite measurements, local operations and classical communication (LOCC), separable measurements, and positive-partial-transpose (PPT) measurements form a useful hierarchy [13, 14, 15, 16]. PPT constraints provide a tractable outer relaxation of separable measurements. We apply this hierarchy to a temporal discrimination problem in which the tensor factors represent readout signals arriving in successive rounds. Measuring each signal and retaining only its classical outcome produces adaptive-tree effects that are separable and therefore PPT. The resulting SDP consequently bounds the entire defined measure-and-feed-forward class, rather than a particular decoder or measurement policy. Our contribution is to combine this relaxation with a coherent receiver construction and rigorous verification of the optimization bounds. The resulting certificate establishes a strict separation between the receiver classes, rather than relying on numerical comparisons between individual measurement strategies.

Figure 1. Receiver-class comparison. Both receivers receive the same pulse ensemble, prior, clock, and recovery objective. The separate adaptive receiver may preserve an arbitrarily rich classical record. The coherent receiver additionally preserves a finite-dimensional quantum system between rounds. In the strict memory- model, the labeled -dimensional register is the only persistent state; no additional persistent classical side register is available.
d. Quantum memories and entanglement-breaking channels. Entanglement-breaking (EB) channels are precisely measure-and-prepare channels [17]. Resource-theoretic approaches have used discrimination tasks to verify whether a quantum memory is not entanglement-breaking [18]. However, the temporal interpretation of this property requires care: EB channels can still support nonclassical multitime correlations under broader operational notions of memory [19]. Our result is therefore specific to the causal receiver architecture considered here. If the complete classically flagged instrument at each persistent boundary is measure-and-prepare, the resulting statistics can be reproduced by a separate adaptive receiver. Consequently, a receiver that achieves the certified separation must have at least one boundary that is not measure-and-prepare. This gives the separation a resource interpretation: the advantage requires genuinely quantum memory, not merely a different classical encoding of earlier measurement outcomes. The simulability statement does not extend to arbitrary multitime intervention models.
e. Collective and sequential quantum receivers. Collective optical receivers can outperform symbol-by-symbol detection on structured coherent-state ensembles [20, 21]. Such improvements arise because joint measurements need
not reduce each signal to a classical outcome before subsequent signals are processed. Previous work has also shown that small persistent quantum memories can reproduce collective discrimination performance in streaming settings [22, 23]. Our work builds on this connection but considers a recovery-weighted objective and compares finite-memory coherent receivers against the entire class of arbitrary local measurements with classical memory and feed-forward. A finite persistent register restricts the operator-space complexity of the induced temporal effects. In the strict architecture considered here, this appears as an operator-Schmidt-rank bound across each chronological cut, corresponding to the matrix-product-operator bond-dimension constraint associated with finite sequential memory [24, 25]. We use these structural constraints to establish lower bounds on the required memory dimension and combine them with certified receiver constructions. For the two-round benchmark, this determines the exact minimum memory dimension needed to outperform every separate adaptive receiver, strengthening the connection between sequential discrimination performance and the quantum resources required to achieve it.
f. Quantum reservoir computing. Quantum reservoir computing (QRC) uses a persistent dynamical system and a simple trained readout to process temporal data [26, 27]. Its use of persistent quantum dynamics makes it a natural candidate for processing temporally correlated readout signals without measuring each signal independently. Appendix J examines this connection for the benchmark by comparing a homogeneous reservoir with a clocked construction. The clocked construction achieves lower error in all ten paired best-of-three comparisons with the matched homogeneous reservoir. Because the clocked construction was informed by the optimized channel, we treat these results as exploratory rather than as evidence of a general architectural advantage. They illustrate how the receiver framework may inform practical sequential architectures, while the formal contribution of this paper remains the certified separation and the associated memory-resource bounds.
Physical background and notation
From physical errors to readout signals
A physical qubit is an individual two-level quantum system. A logical qubit stores quantum information in a subspace of several physical qubits, so that specified errors can be detected without measuring the unknown logical amplitudes. In a stabilizer code, commuting check operators distinguish error sectors. Their outcomes form the syndrome: information about an error, rather than a measurement of the stored logical state. An ancillary qubit is a helper system coupled to the data to extract a check outcome, then read out and reset.
For example, the three-qubit repetition encoding , protects against a single bit flip, rather than arbitrary quantum noise; phase errors require additional protection. Here , denote computational-basis states, exchanges them, and . The parity checks and give syndrome bits and for error mask , where is addition modulo two. This convention illustrates syndrome extraction; the phase lookup used in the certified benchmark is a separate part of its specification.
In dispersive microwave readout, an ancilla changes the response of a resonator and thereby the state of an outgoing pulse [28]. Finite pulses associated with different states can overlap, so a measurement need not reveal their labels without error. Repeated extraction rounds help distinguish ambiguous observations and track faults occurring between rounds. A receiver combines the readout signals to select a correction, either applied physically or recorded in a Pauli frame, a classical record of corrections used when interpreting later operations and measurements.
The benchmark below models these emitted signals directly. Each two-bit syndrome is assigned to one of four coherent pulse states. The ancilla circuit is physical motivation rather than a simulated component of the numerical model. The question is whether a receiver can use the same pulse sequence more effectively by retaining quantum information between arrivals. Figure 2 shows this connection.

Figure 2. Readout generation and signal encoding. Left: a conceptual parity-check extraction circuit couples data qubits to ancillary systems and generates readout signals. The receiver processes these signals and selects a recovery; the extraction circuit is illustrative rather than simulated in the benchmark. Right: quadrature phase-shift keying (QPSK) places four coherent-state amplitudes at phases , , , , with radius . The axes are real and imaginary field amplitude. Distinct points represent nonorthogonal quantum states. The phase index is assigned to a syndrome by the benchmark lookup, not by the drawing.
Mathematical description of signals and receivers
A Hilbert space describes a quantum system, and is the space of linear operators on it. The set consists of density operators: positive semidefinite operators of trace one. We write when for every vector , and when .
The trace is the sum of diagonal entries, is the Hermitian adjoint, is transpose in the stated basis, and is the identity operator. A ket denotes a state vector and its pure-state density operator. Tensor products, written , describe joint systems; here their factors label different temporal signals. A partial trace discards subsystem , while transposes only the factors indexed by .
A positive operator-valued measure (POVM) is a general quantum measurement: its positive effects sum to the identity, and is the probability of decision for input . A quantum instrument also specifies the state update associated with each outcome. A completely positive trace-preserving (CPTP) map is a quantum channel: it maps valid states to valid states, including when applied to part of a larger system. An entanglement-breaking (EB) channel measures its input and prepares a new state depending only on the outcome.
We use for the number of readout rounds and for the Hilbert-space dimension of the persistent receiver register. Thus describes the size of the quantum state available from one round to the next, rather than a physical-qubit count. It embeds in ideal qubits, where denotes the ceiling function. The single-signal dimension is when common to all rounds. The separation rate symbol denotes code distance, the minimum weight of a nontrivial logical Pauli operator in a stabilizer code.
a. Operator norms and measurement bounds. The receiver comparison uses matrix constraints and norms to bound measurement performance and sensitivity to perturbations. For operators, is the trace norm, the largest singular value, and the Frobenius norm. An effect is separable across temporal factors when it is a limit of sums of tensor products of positive effects. Such effects have positive partial transpose (PPT), the necessary condition used to bound adaptive receivers. A semidefinite program (SDP) optimizes a linear objective with positive-matrix constraints; a feasible dual witness supplies a bound on every primal solution.
Sequential recovery problem and receiver classes
Loss-weighted pulse ensemble
Let denote the number of readout rounds. A latent trajectory specifies the underlying fault history together with the syndrome labels it produces over those rounds. The trajectory is assigned prior probability and generates an arbitrary receiver-independent temporal state
| Symbol | Meaning |
| , | Error trajectory and its probability |
| Readout state emitted for that trajectory | |
| , | Recovery action and its measurement effect |
| Loss for that action on trajectory | |
| , | Cost and success operators weighted by the prior |
| Average loss of a specified receiver | |
| , | Number of rounds and persistent-register dimension |
| , | Signal dimension and code distance |
| , | Class-optimal risk and minimum strict-register dimension that outperforms the adaptive class |
Table 1. Main quantities and their physical meanings.
Thus may be mixed and may contain arbitrary classical or quantum temporal correlations. Temporal separability below is a property of the receiver protocol, not of this input state. The only source assumption is that the joint state is fixed independently of receiver actions. If an earlier receiver intervention causally changes later emissions, a process tensor or quantum-comb description is required instead [29, 30]. The present benchmark has no such feedback into the source: syndrome extraction and the emitted ensemble are held fixed.
Let label a recovery action and its recovery loss. Operationally, can represent a physical correction or a Pauli-frame update. In the binary-loss benchmark below, the loss is zero exactly when the selected recovery mask equals the final physical-error mask, and one otherwise. Code-level logical success can identify different physical errors as equivalent; the benchmark instead uses exact mask matching. Define
For an effective POVM , the average recovery risk is the linear functional
This expression averages the loss over the possible trajectories and the receiver’s decisions. For the benchmark below, it is the probability of an incorrect final-mask decision, called the recovery-decision error. We distinguish this risk of a specified receiver from the class optima,
Here and denote the receiver classes defined below; is the risk of a specified receiver, whereas starred quantities denote class optima. Each incorporates the trajectory prior, emitted temporal state, and recovery loss associated with recovery action . The resulting receiver optimization is a linear quantum decision problem. The objective rewards successful recovery rather than exact reconstruction of the syndrome history.
Two-round benchmark
The numerical instance specializes this construction to two sequential readout rounds. In each round the relevant syndrome has four possible values, represented by two syndrome bits. A fixed lookup maps these four labels one-to-one to a QPSK phase , producing the coherent state
| Increment | Probability | Syndrome | Pulse |
| No flip, 0 | |||
| First bit, 1 | |||
| Second bit, 2 | |||
| Third bit, 4 |
Table II. Per-round benchmark increments. The syndrome and pulse columns describe a single increment starting from zero error. At later rounds the lookup is applied to the cumulative mask . The symbolic functions and denote the benchmark’s syndrome and phase lookups.
The amplitude is dimensionless, with mean photon number . Four equally spaced phases encode the four syndrome labels. Distinct amplitudes in phase space need not be orthogonal states. Coherent pointer states arise naturally in dispersive microwave readout, where finite-amplitude responses associated with different system states are not perfectly distinguishable [28]. The QPSK encoding used here provides a symmetric four-state benchmark of this nonorthogonal-readout setting. At the four pulses are nonorthogonal, so no measurement can identify all four labels perfectly. We set in the trajectory prior . The final decision set contains eight recovery masks. The loss and trajectory construction are specified below and in Appendix § 1.15.
a. Error trajectories and the recovery criterion. The per-round error increment is , where each nonzero integer mask flips one of the three physical bits. Increments are independent, with probabilities , , , . The cumulative error is the bitwise exclusive OR , and specifies the full trajectory. The syndrome of the cumulative mask is , giving phase index . The eight actions are all three-bit recovery masks, and the loss is
where the indicator is one if its condition holds and zero otherwise. An error near 24.21% therefore means an incorrect final mask on that fraction of the ensemble.
Exact mask matching cancels all modeled bit flips. For an encoded state, a nonzero residual mask may instead leave a detectable error or induce a logical operation. Counting every such residual as a failure is stricter than identifying only logical equivalence classes. The reported numbers measure this specified recovery decision, rather than a circuit-level logical-failure rate.
For example, is a first-bit flip in round one and no new flip in round two, with probability . The cumulative mask is one in both rounds, so both emitted pulses are . In the illustrative check convention of Sec. III, the syndrome is 10. Choosing has zero loss; every other mask has unit loss. The receiver sees only the readout signals and must infer the recovery without knowing .
Unlike the general formulation in Equation (1), this specific QPSK benchmark uses conditionally independent emitted modes,
The product form is therefore a specialization of the numerical instance, not an assumption of the receiver-class theorem.
Although the pulse modes have an infinite-dimensional Fock space, the four states used in each round span a four-dimensional subspace . Let project onto that subspace and write . Every benchmark state satisfies . For any receiver effect , the compression , regarded as an operator on , therefore gives the same outcome probabilities and risk. The compressed effects remain a POVM on that space. For a separate adaptive receiver, compression acts on each local factor of every product effect, preserving separability and PPT. Its conditional local measurements remain complete on . The coherent update maps can likewise be restricted to these supported signal inputs without changing their statistics. Thus the finite-dimensional receiver results apply exactly to this ensemble, without a photon-number cutoff.
We represent this finite-dimensional signal space directly through the coherent-state Gram matrix
Here is the overlap of pulses and , and the asterisk denotes complex conjugation. A factorization supplies coordinate vectors for the states without truncating their photon-number expansion. The four states span dimensions; this is an effective representation of the chosen ensemble, not a truncation of the electromagnetic field’s full Hilbert space. The certified calculation reconstructs these coordinates and the cost operators directly from the analytic Gram data using Arb/Acb ball arithmetic. No photon-number cutoff enters the proof. Appendix § 1.15 describes the analytic reconstruction and an independent truncated-Fock convergence check. The parameter is a benchmark trajectory parameter rather than a calibrated device error rate. The operating point is chosen to provide a nontrivial temporal-discrimination regime, rather than to model a fault-tolerant device setting.
Receiver classes
The separate adaptive receiver carries an unrestricted classical transcript between rounds. The general coherent receiver can retain quantum memory alongside that transcript. The strict receiver permits only the counted -dimensional register. These are constraints on retained information, rather than on classical computational power.
Definition 1 (Separate adaptive receiver). At round , the receiver applies an arbitrary quantum instrument to the current pulse. Its choice may depend on the clock and the entire earlier classical transcript. It may use arbitrary local ancillas, randomness, continuous outcomes, and unrestricted classical computation. No quantum state carried between rounds may depend on an earlier pulse; all information retained from previous rounds is contained in the classical transcript.
Thus a separate adaptive receiver may remember arbitrarily much about the past, including a complete analog record, but every persistent record must be classical. For , coarse-graining the complete transcript into the final action gives an effect
where is the first-round effect and is the conditional second-round POVM. Each is therefore separable across the temporal cut.
Definition 2 (Coherent receiver). A coherent receiver is a causal receiver that may retain a finite-dimensional quantum system between rounds. It may also retain an unrestricted classical transcript and condition later operations on that transcript. The union over all finite quantum-memory dimensions is denoted .
Definition 3 (Strict memory- streaming receiver). A strict memory- receiver consists of completely positive trace-preserving maps
followed by a final memory POVM . The -dimensional register is the only persistent state between rounds: no additional persistent classical record, quantum pointer, retained signal, or entangled ancillary system is available outside it. Fresh within-round ancillas may be introduced and discarded. The resulting receiver class is denoted .
The memory-dimension results use . The receiver-class separation instead compares with the broader class , which permits additional classical side information. In , the final POVM acts on the same counted -dimensional register.
The unrestricted collective receiver, which measures directly, supplies a global comparison point. The streaming model asks how much of that performance can be retained without storing the entire pulse block.
Arbitrary-depth structure of separate adaptive receivers
Theorem 1 (Temporal separability of classical-memory receivers). Let be finite-dimensional temporal signal systems. Every receiver that measures each signal locally, retains only classical information between rounds, and classically maps its complete transcript to a final action induces effects of the form
where every factor is positive semidefinite. Thus is a positive conic integral of product effects, equivalently a limit of finite positive sums of product effects, and is fully separable across . Consequently,
The receiver may use arbitrary discrete or continuous outcomes, private randomness, unrestricted classical computation, and fresh or resettable local ancillas. Outcome spaces are taken to be standard Borel, with measurable conditional measurement kernels.
The proof is given in Appendix A.
Corollary 1 (Multipartite PPT outer bound). For any -round recovery problem, minimizing the recovery loss over a measurement class constrained to be PPT across any selected temporal cuts gives a lower bound on the error of every separate adaptive receiver. If the risk of a feasible coherent receiver is below a certified dual lower bound for that relaxation, then it outperforms the entire separate adaptive class.
The benchmark emits product states conditional on the error trajectory. The coherent advantage therefore does not require entanglement between the emitted pulses. The relevant resource is the receiver’s ability to retain pulse-dependent quantum information. A boundary that measures its entire input and prepares a new state can instead be simulated by retaining the measurement outcome classically, as the next proposition establishes.
Proposition 1 (Entanglement-breaking persistent boundaries). Consider a receiver-independent temporal input and allow unrestricted classical side information. After each intermediate round , let denote the complete pulse-dependent quantum system retained for round , and include any outcome produced at that boundary in the retained classical history. Suppose that, for every
incoming classical history, the complete boundary instrument is measure-and-prepare: each of its classical-outcome branches admits a measure-and-prepare decomposition for the state of . If no other pulse-dependent quantum system survives between rounds, the final action statistics can be reproduced by a separate adaptive receiver. Hence a certified advantage requires at least one intermediate boundary whose complete, classically flagged instrument is not entanglement breaking.
Remark 1. The entanglement-breaking condition applies to the complete quantum system crossing the persistent-memory boundary. Testing a reduced memory channel after fixing the incoming signal is insufficient: a SWAP-type update can make that reduced map entanglement breaking while transferring the signal’s quantum information intact into the retained register.
Proof. Let be a fixed, input-independent state. Suppose inductively that, conditional on the classical history before round , the retained memory is a known prepared state , uncorrelated with the unprocessed signals. Those signals may have an arbitrary joint conditional state. Include the boundary’s classical output in the next history. The branchwise measure-and-prepare assumption gives
where , and summing over and integrating over gives on . Define effects on the current signal by
They are positive and sum (with integration) to the identity on . A separate adaptive receiver can measure this conditional POVM, retain classically, and prepare locally. This gives the same joint outcome distribution and conditional state of all untouched future signals as the original instrument. It also restores the induction hypothesis for the next round.
At the last round, let be the final POVM on . With the entering memory prepared in , the conditional effects
form a POVM on the last signal and reproduce the final action statistics. The resulting protocol measures each signal on arrival and retains only a classical transcript. No independence assumption on the emitted temporal state is needed.
Non-entanglement-breaking memory is therefore necessary, but not sufficient, for an advantage in this receiver model. This conclusion does not extend to more general multitime intervention settings [19].
Two-round separation
Collective and PPT programs
To bound the entire adaptive class, we optimize over a larger measurement class with directly checkable constraints. A feasible coherent measurement gives an achievable error upper bound. A feasible dual witness for the larger class gives a lower bound that applies to every adaptive receiver. If the upper bound is smaller, no assumption of global optimality of the coherent search is needed.
The unrestricted collective receiver minimizes the error over all POVMs on the full two-round signal space and provides a reference for the streaming receivers. The collective optimum is the SDP
Its dual, with Hermitian witness , maximizes subject to for every [12 Sec. 3.1.2]. Primal feasibility gives an achievable error; dual feasibility gives a lower bound on the collective optimum.
Let denote partial transpose on the first temporal factor, . Because every separate adaptive effect is separable across the two temporal factors, it remains positive under . Dropping separability and imposing only this necessary PPT condition enlarges the feasible measurement class and gives
Here the dual uses a Hermitian operator and one positive multiplier for each action. A convenient dual is
This is the loss-minimization form of the standard PPT discrimination primal–dual pair [16 Sec. 3.1]; see also [12 Sec. 1.2.3] for SDP duality.
Proposition 2 (Adaptive containment). Every two-round separate adaptive receiver induces a POVM feasible for Equation (18). Consequently,
| Quantity | Recovery-decision error |
| Collective optimum | |
| Certified risk upper bound | |
| Certified PPT lower bound on adaptive error | |
| Achievable adaptive risk upper bound | |
| Certified gap lower bound |
Table III. Certified two-round benchmark. Inequality directions follow from outward-rounded receiver and dual verification.
Proof. Each coarse-grained effect is an integral of positive product operators, as in Equation (10). It is therefore separable. Partial transpose maps to ; integration, convex mixing, local ancillas, continuous records, and randomization preserve this property. The completeness relation is unchanged.
A coherent receiver whose risk lies below a certified lower bound on therefore outperforms the entire separate adaptive class.
Floating-point optimization is used to find candidate receivers and dual witnesses, but it is not used directly for the theorem. The numerical receiver and dual candidates are embedded as exact rational data and checked against cost operators reconstructed directly from the analytic benchmark using 256-bit Arb/Acb ball arithmetic. Outward rounding yields a risk upper bound for each explicit receiver and a dual objective lower bound for the PPT relaxation. The procedure is described in Appendix § 1.3.4.
Separation criterion and numerical values
Theorem 2 (Certified two-round separation). For the two-round cost operators , the certified streaming receiver and PPT-dual witness satisfy
Consequently,
The upper bound concerns one explicitly feasible receiver, whereas lower-bounds the error of every adaptive receiver. Their strict ordering certifies the class separation. They are bounds rather than claims that both class optima are known exactly.
The theorem follows from weak duality and feasibility of the explicit receiver. The lower bound applies to every separate adaptive receiver; tightness of the PPT relaxation is unnecessary. The same calculation encloses the collective optimum as
The receiver is therefore within of the collective optimum. The enclosure does not establish equality with the collective optimum.

Figure 3. Certified two-round comparison. The horizontal axis is recovery-decision error. The coherent-risk upper bound lies below the certified PPT lower bound on the separate-adaptive optimum. The explicit adaptive receiver is shown as an achievable reference.
Proposition 3 (Stability of a certified separation). Let be a perturbed decision problem on the same physical Hilbert space and with the same receiver classes, where . Suppose is a certified lower bound and a fixed feasible coherent receiver has risk . Define
where the common Hermitian shift is independent of . If , then the perturbed class optima obey
Proof. For any POVM , , and hence, for every Hermitian ,
POVM completeness makes the common shift contribute . Therefore . Evaluating the same explicit coherent receiver on gives , while . Subtraction and optimization over prove Equation (24).
For the certified two-round gap, Proposition 3 guarantees separation whenever
The bound is sufficient and need not be tight.
Certified parameter neighborhood {#sec:C}
Interval verification gives the same strict two-round threshold throughout
with a whole-box separation lower bound of 0.001627522285899. Thus throughout this inner rectangle in the strict model. The verification includes both exclusion of , with a uniform margin of at least 0.034644013557275, and a feasible receiver throughout the rectangle. A separate certificate covers
with gap at least 0.001932932526545 and uniform exclusion margin at least 0.023874515270248; there the established memory statement is . All values of are represented in a common analytic signal frame; details are given in Appendix § 1.15. These rectangles give local parameter-robustness certificates around the benchmark point.
Temporal quantum-memory complexity
Streaming optimization
A collective POVM does not specify the sequential channels needed for a streaming implementation. We therefore optimize the channel sequence in Equation (11) using Choi-matrix representations, enforcing complete positivity and trace preservation for each update and positivity and completeness for the final POVM. With all other blocks fixed, one update channel or the final measurement enters the objective linearly and can be optimized by an SDP. Each completed sweep produces a candidate receiver, which is subsequently checked and, when necessary, repaired as described in Appendix § 1. The joint problem is nevertheless nonconvex, so the reported values are the best feasible values found rather than certified global optima. Appendix § 1 states the Choi convention, block structure, and feasibility checks used in this calculation.
A Choi matrix represents a channel as a positive matrix subject to a linear trace-preservation condition. This is why each individual update can be optimized by an SDP when the others are fixed. Its convention is given in Appendix § 1.
For the strict class in Definition 3, define
The minimum of an empty set is . Thus is the smallest dimension whose optimal strict receiver outperforms the adaptive optimum; the certificates determine it by upper and lower bounds. It is not defined by the outcome of a particular local search. These definitions apply to the strict counted-register convention and do not include architectures with an additional persistent classical record.
Proposition 4 (Attainment and monotonicity). For finite signal and action spaces, the minimum in Eq. (29) is attained and
Theorem 3 (Constructive full-storage bound). For signals of common dimension and final actions, every -outcome collective POVM can be realized by a strict streaming receiver with
The operator-Schmidt rank is the smallest number of product operators needed to express the effect across the cut between the first signals and the rest.
Theorem 4 (Finite-memory operator-Schmidt-rank bound). Every terminal effect induced by a strict memory- receiver obeys
across every chronological prefix-suffix cut . For a fixed receiver, the action effects share the same channel-generated left slice space at each chronological cut.
The -dimensional operator space of a memory limits the structure that can link earlier and later signals in each decision effect. This rank bound describes the receiver’s measurement, not entanglement of the emitted input.
The proofs are given in Appendix B. The operator-Schmidt statement is the standard matrix-product-operator bound constraint associated with a finite operator space [24, 25]; its use here is to formulate memory requirements for the recovery-weighted decision problem.
Global lower bounds for fixed memory dimension
The minimality statements below require lower bounds valid for every strict memory- receiver, not merely the best value found by a local search. Two such bounds are used. Both are stated with the success operators
which satisfy for the binary loss.
The success operator weights each trajectory state by its reward for action . In the relaxation below, is an effective input effect, bounds its operator norm, and the sum of these bounds is limited by the terminal register dimension.
Lemma 1 (Terminal norm budget). Let be the effective POVM on the full temporal input induced by a strict memory- receiver. Then
Proof. Write , where is the composed channel from the temporal input to the terminal -dimensional register and is the terminal POVM element. The adjoint is positive and unital, so . Since , , and .
Lemma 1 gives the convex outer relaxation
subject to
Every strict memory- receiver is feasible for Equation 36, so its optimum upper-bounds the achievable success probability. A convenient dual is
subject to
The dual follows by assigning a Hermitian multiplier to POVM completeness, positive multipliers to , and a nonnegative multiplier to the terminal norm budget, using standard SDP duality [12 Sec. 1.2.3]. Any certified feasible dual value therefore gives the global bound
For the two-round benchmark, interval verification of feasible duals gives
In the dual, is Hermitian, are positive multipliers, and is a nonnegative scalar. A feasible value upper-bounds success, so lower-bounds the risk. A second, purely combinatorial bound is useful at greater depth.
Lemma 2 (Terminal-dimension prior bound). For binary loss , let , and order these weights as , padding with zeros if . Then
The proof is given in Appendix B.
Memory-dimension frontier
Figure 4 shows the best values obtained in the fixed-memory search, while Table IV gives the separate certified bounds used to establish the strict memory threshold. The plotted and values are feasible search points; their exclusion instead follows from Equation (46). The and entries are explicit receivers whose risks are certified by outward rounding.

Figure 4. Two-round memory-performance search. The horizontal axis is memory dimension , and the vertical axis is recovery-decision error. Markers show the best values found at each tested dimension. The exact strict-model threshold follows from the global exclusion certificates in Table IV, not from the plotted points. The plotted collective line is the numerical estimate used in the search; Eq. (22) gives the certified enclosure.
| Object | Certified relation | Role |
| global value | ||
| global lower bound | ||
| global lower bound | ||
| adaptive strategy | ||
| explicit strict receiver | ||
| PPT lower bound |
Table IV. Certified bounds used for the two-round strict-memory threshold. The adaptive entry is an achievable risk upper bound.
For , global lower bounds exceed the error of an available adaptive strategy, excluding an advantage at these dimensions. A certified receiver instead has error below the lower bound for every adaptive receiver. Together these comparisons establish necessity and sufficiency. Four dimensions could be represented by two ideal qubits; the threshold counts the entire persistent and terminal register in the strict model.
Theorem 5 (Two-round strict-memory threshold). For the certified two-round benchmark and the strict terminal-
register model,
Proof. The first three rows of Table IV give for . Since the displayed adaptive strategy is feasible, ; therefore none of these dimensions can beat the adaptive optimum. Conversely, the certified receiver obeys
and , so suffices. This proves , and the same exclusion gives . Recovery action 7 is never correct for the two-round benchmark, so its success operator satisfies . Any weight assigned to outcome 7 can be reassigned to another outcome without decreasing the success probability, since all success operators are positive semidefinite. Hence an optimal collective POVM may be chosen with at most seven nonzero decision outcomes. Applying Theorem 3 with and gives .
The threshold refers to the strict counted-register convention: the terminal register is included in , and no additional persistent classical history is available outside that register. By Proposition 1, the certified receiver must contain a non-entanglement-breaking persistent boundary. This is a statement about the channel, not about entanglement of the benchmark state.
Multi-round finite-memory results
For , temporal separability again permits PPT constraints across chosen temporal bipartitions. The certificate imposes all three one-versus-two cuts simultaneously (only and are chronological),
and gives . Only the balanced cut is imposed in the calculation, which gives .
The full-storage construction of Theorem 3 gives at and at for and . Certified separation is achieved with and , respectively. Table V collects the outward-safe bounds.
| Depth | Memory | Certified risk upper bound | PPT lower bound | Gap lower bound |
| 0.294297344448216 | 0.299647463773192 | 0.005350119324975 | ||
| 0.284502699489212 | 0.299647463773192 | 0.015144764283980 | ||
| 0.319370057047514 | 0.319542490402679 | 0.000172433355164 | ||
| 0.317529609881821 | 0.319542490402679 | 0.002012880520857 |
Table V. Certified streaming constructions at greater depth. The comparator uses all cuts in Eq. (50); the comparator uses only .
The and entries are explicit strict streaming realizations. The and entries are lower-memory streaming receivers obtained directly from the restricted prefix-memory construction described in Appendix § 1.16.
The lower endpoints at and follow from Lemma 2. Certified feasible product receivers give the adaptive-class upper bounds
The dimension-prior floor excludes against these feasible comparators, giving at both depths. The corresponding dimension-prior lower bounds lie above the certified feasible full-storage risk upper bounds at and at , establishing .
The reduction from 64 to 36 is in Hilbert-space dimension. Both need six qubits in a binary embedding: . A smaller dimension alone does not imply fewer physical qubits or a simpler control circuit.
At and , full-prefix storage is therefore not necessary for a certified separation. The tested depths do not determine asymptotic memory scaling. The certified margins are also not monotonic in : the stronger three-cut relaxation produces a larger full-storage margin than the two-round calculation, while the present bound uses only one balanced cut. The margin therefore depends on both the recovery problem and the tightness of the chosen outer relaxation. Appendix § 1.16 describes the deeper streaming constructions and certificates; the observation-only control remains separate in Appendix H.
Robustness and implementation error
Dephasing of the counted persistent register
We first apply noise to the register of an explicit streaming receiver while holding its channels and terminal measurement fixed. This tests whether its advantage survives imperfections. Section IX instead acts on the original pulse registers and reoptimizes a collective measurement.
For the strict streaming receivers, whole-register dephasing is applied after each intermediate update. The update channels and terminal POVM are otherwise kept fixed. We use for dephasing strength and for coherence retention:
This differs from reoptimizing a terminal measurement for each noise strength. At there is a single intermediate memory boundary, so the certified curve is a test of inter-round memory alone and the risk is affine in .
Interval verification brackets the crossing points near
The available bracket information is listed in Appendix C; exact two-round endpoints remain to be transcribed. Thus the fixed and receivers retain a certified advantage up to approximately 1.69% and 5.19% dephasing strength, respectively. These dephasing strengths are defined in the numerical memory basis and are not calibrated device error rates. Analogous fixed-receiver certificates are obtained for the deeper and constructions; their exact crossing brackets are listed in Appendix C.
Table VI. Current strict-memory bounds. Only the two-round separation threshold is exact. The prefix column is , before accounting for the terminal action register in Eq. (33).
Channel-distance error budget
Theorem 6 (Implementation-error accumulation). Let the ideal strict receiver be a composition of CPTP factors, including a terminal measurement channel, and let factor be replaced by an implementation at unhalved diamond distance . The diamond norm maximizes the output trace norm over joint inputs with an arbitrary reference system and input trace norm one. No factor of one half is included in . For any loss in ,
The proof is a telescoping hybrid argument followed by contractivity of trace distance and is given in Appendix C. Combining Eq. (53) with the certified W2 margins gives the sufficient budgets
These are dimensionless sufficient channel-error budgets, not calibrated hardware tolerances.
Stored-pulse hardware model
We also consider a stored-pulse model in which individual pulse registers undergo phase decay before an optimized collective readout. This model acts on the stored pulse registers rather than on the learned -dimensional streaming memory. At each trial value of , the final joint POVM is reoptimized while the stored pulses undergo pulse-dependent phase retention.
Here is the coherence-decay time, the energy-relaxation time, and the pure-dephasing time. The interval separates pulses, and is the final processing delay. For pulse , the modeled storage time and retention are
The earliest pulse has the longest modeled storage time,
Amplitude loss must be simulated separately, for example with
Although
relates the decay constants in a Markovian model, it does not replace separate amplitude-damping and pure-dephasing channels.
For the slow schedule, the timing anchors are
The timing parameters are modeling inputs drawn from the stated schedule and a literature anchor [31]; they are not measurements of an integrated implementation. The composition of is given in Appendix § 1.21. The corresponding longest storage times are 8.46, 14.28, and 20.10 for . At a numerical crossing we report the retention of the earliest pulse,
The resulting crossings, shown in Figure 6 of Appendix § 1.21, are , 767.970 , and 2656.778 for .

Figure 6. Stored-pulse collective-readout model. Solid curves show the slow-schedule calculation in which the final joint POVM is reoptimized at each . Each panel has coherence time on the horizontal axis and recovery-decision error on the vertical axis; panels correspond to . Horizontal dashed lines show the comparator values used in the stored-pulse root calculation; the root uses the single-cut comparator employed in that stored-pulse calculation rather than the three-cut bound in Eq. (50). The marked crossings give , , and for . Vertical lines indicate separately reported component-scale coherence values: approximately for a tantalum transmon [34], for an integrated cavity memory [33], and for a high- cavity qubit [35]. These coherence values were reported in separate experiments and do not represent an integrated realization of the receiver considered here.
These crossings are numerical outputs of the stored-pulse model and are not part of the interval-certified receiver bounds. The crossing lies within reported component-scale coherence times, whereas the and calculations require progressively longer storage. The comparison is made at the component level; reported component values and the timing assumptions behind them are collected in Appendix § 1.21.
Discussion
The certified coherent-risk upper bound lies below the certified PPT lower bound on the entire separate-adaptive class. That class already permits continuous outcomes, unrestricted classical memory, and unrestricted classical computation. The operational distinction is therefore whether pulse-dependent quantum information may persist between rounds.
Proposition 3 shows that the positive gap persists under sufficiently small perturbations of the cost operators. Proposition 1 shows that a receiver whose complete classically flagged boundary instruments are all measure-and-prepare can be simulated by a separate adaptive receiver. The certified separation therefore implies that at least one persistent boundary of that receiver is not measure-and-prepare. This implication is specific to the receiver architecture defined here; entanglement-breaking dynamics can behave differently under more general multitime interventions [19].
For the strict two-round architecture, the minimum terminal memory dimension required to outperform the separate adaptive class is exactly four. This threshold includes the terminal register and assumes that no additional persistent classical history is available outside the counted register. The fixed-receiver dephasing study acts on this counted register, while the diamond-norm theorem bounds the effect of imperfect channel implementations without choosing a dephasing basis.
At greater depth the exact threshold is not known, but certified separation is already obtained with at and at , compared with and for the full-prefix construction. At , the simultaneous three-cut relaxation gives a full-storage certified margin of 0.015144764283980, while the calculation presently uses only one balanced cut. The tested depths therefore do not establish a monotonic trend in the true coherent-adaptive gap or an asymptotic memory scaling law.
The architecture also changes where latency enters the recovery pipeline. A separate adaptive receiver converts each arriving signal into a classical record before any pulse-dependent information from that signal can participate in later-round processing, whereas a coherent receiver may carry such information forward in a quantum memory. This can replace some inter-round measurement and feed-forward requirements with coherent updates and storage. The present results do not establish lower end-to-end latency: coherent capture, memory operations, storage, and final readout introduce their own timing requirements, and, in the separate stored-pulse model considered here, the required coherence time increases substantially with depth. A quantitative comparison of these resources would require a hardware-specific timing model and is left for future work.
Several limitations remain. The pulse ensemble, extraction circuit, QPSK encoding, prior, and recovery loss are fixed. PPT is an outer relaxation and may become loose at larger temporal depth; Appendix G gives an analytic two-round task with a strict coherent advantage whose optimal measurement is nevertheless separable. The optimized channels have not yet been compiled to a low-depth device Hamiltonian. The QRC calculation is exploratory, and the stored-pulse hardware calculation is distinct from the learned streaming-memory architecture. The most immediate extensions are a stronger multi-cut relaxation, device-level compilation of the coherent updates, and tests with independently calibrated amplitude and phase noise. Appendix § 1.22 states a conditional flagged-sector inheritance criterion, but its hypotheses have not been verified for a natural surface-code family.
Conclusion
The point at which syndrome information becomes classical is itself a receiver-design decision. For the fixed recovery problem studied here, an explicit coherent streaming receiver has certified recovery risk below every separate adaptive receiver. In the strict counted-register model, the exact two-round separation threshold is , and the induced temporal effects have operator-Schmidt rank at most across every chronological cut. At greater depth, certified and receivers show that full-prefix storage is not required for separation at and .
Dephasing of the actual persistent register gives a direct robustness test for the explicit receivers, while the channel-distance bound gives a general sufficient implementation-error condition. Both results are defined within the receiver model used here. The distinction from the separate adaptive class is the ability to retain pulse-dependent quantum information across rounds. Exact and thresholds, a stronger multi-cut bound, and device-level compilation of the certified updates remain open.
Proof of temporal separability
Proof of Theorem 1. Condition on the receiver’s private randomness. Fresh local ancillas can be absorbed into the effective POVM on the current signal. At each node of the adaptive tree, the conditional POVM is a measurable operator-valued kernel. In finite dimension, it admits a positive density with respect to its scalar trace measure, which may depend on the preceding history. The integrals below are understood iteratively with these conditional measures. For a complete transcript , the associated leaf effect therefore factors as
and each factor is positive semidefinite. Let be the measurable set of transcripts that the classical decision rule maps to action . Coarse-graining those leaves and averaging over private randomness gives
which is exactly Equation (12). In finite dimension, this is a positive conic integral of product effects, equivalently a limit of finite positive sums of product effects, and therefore belongs to the closed fully separable cone.
Summing Equation (62) over the final actions, the sets partition the transcript space, and recursive use of the local completeness relations at each node gives . Finally, partial transpose on any subset of temporal factors maps each product integrand to another positive product operator. Positivity is preserved by integration, so for every temporal subset , completing the proof.
Memory-complexity proofs
For fixed , the Choi matrices of the strict update channels and the final POVM belong to closed and bounded subsets of finite-dimensional vector spaces. Their product is compact, and the risk is continuous; this proves attainment in Proposition 4. Embedding as a subspace of , extending each channel on the unused level, and ensuring that the added level is never reached from the embedded input subspace reproduces every memory- receiver at dimension . The final POVM is extended arbitrarily on the orthogonal one-dimensional complement, for example by assigning that complement to one fixed outcome. Because the added level is unreachable, this does not change the induced statistics and proves monotonicity.
For Theorem 3, store the first signals coherently in a prefix register. Immediately before the last signal arrives, its dimension is . Apply the desired collective POVM when the last signal arrives, write its outcome into one of orthogonal terminal states, and measure that state. A common register of dimension accommodates both stages, proving the claimed sufficiency.
For the archived , and , realizations, the induced measurement was additionally checked on a complete matrix-unit basis. Consequently, those explicit channel sequences reproduce their target collective measurements on arbitrary signal inputs and under extension by an external reference system.
To prove Theorem 4, cut the channel sequence after round . Let be the composed prefix channel from
to the -dimensional memory. Pull the terminal effect for action back through the remaining channels, obtaining an operator on . Choose an operator basis of and expand
Pulling this operator back through the prefix channel gives
The rank is therefore at most , and the left factors depend only on the prefix channels, not on .
Proof of the terminal-dimension prior bound
Let be the effective input effect associated with action , and define
Because , one has . Moreover, Lemma 1 gives
Since and ,
Hence every strict memory- receiver satisfies
Therefore
which proves Lemma 2.
Global bounds used for the two-round threshold
The certified global lower bounds for in Equation (46) are obtained from interval-verified feasible duals of Equation 37. Their comparison with the explicit adaptive upper bound and the certified receiver is the argument given in the proof of Theorem 5. The minimality result uses only the certified global lower bounds.
Robustness proofs and controls
For Theorem 6, replace the ideal CPTP factors one at a time. The telescoping hybrid expansion and diamond-norm contractivity give an unhalved trace-norm change at most in the terminal classical state. The expectation of any loss in changes by at most one half of that distance and never by more than one, proving Eq. (53).
Certified crossing brackets
Table VII gives the available dephasing-strength metadata, in the convention of Eq. (51). The trace-preservation completion used in the certificate is included when the noisy curves are evaluated.
| Receiver | Exact bracket | Decimal |
| , | – | |
| , | – | |
| , | ||
| , |
Table VII. Certified dephasing-strength brackets for the fixed streaming receivers. Exact endpoints for the two-round receivers are pending transcription from the verification archive.
For these particular receivers, terminal-hold dephasing applied after the last update leaves the decision statistics unchanged: the two-round terminal POVM effects are diagonal in the dephasing basis, and the deeper measure-and-encode constructions produce diagonal terminal states. This is a property of these constructions, not a general statement about processing noise.
Pulse-register control
As a separate control, the incoming pulse registers were dephased before a collective terminal measurement. At the coherent endpoint the optimized error is approximately 0.24210, while complete pulse-register dephasing gives approximately 0.29165. The terminal POVM is reoptimized in this control. These values do not describe noise on the learned persistent memory and are not used for the certified thresholds in Eq. (52).

Figure 5. Pulse-register dephasing control. The horizontal axis is retained pulse coherence, and the vertical axis is recovery-decision error. The points show the reoptimized end-point errors and the dashed line the adaptive comparison value used in this control. This auxiliary calculation acts on stored input pulses before collective readout, not on the counted streaming register.
Certificate validation
The certified bounds are evaluated using 256-bit Arb/Acb ball arithmetic. The QPSK Gram matrix, state coordinates, trajectory probabilities, and loss-weighted cost operators are reconstructed from the analytic benchmark rather than read as rounded cost matrices. Binary64 receiver and dual candidates are embedded as exact rational numbers before verification, so their input conversion adds no untracked rounding error.
For each candidate receiver, each Choi matrix and terminal effect is replaced by its Hermitian part in ball arithmetic. After positivity of the terminal POVM effects is verified, those effects are contracted by the exact rational factor to create completeness slack. The positive remainder is assigned to a fixed outcome, restoring exact completeness. Channel trace-preservation remainders are completed by the completely positive map , with interval verification that . The completed maps are then checked for complete positivity and trace preservation. The risk is evaluated from the analytic cost operators with outward rounding, yielding an upper bound on the explicit receiver’s error. Search optimality is not required.
For the PPT dual, the exact-rational witnesses and are inserted into
Interval factorizations verify every cone and slack, and the objective is rounded downward. Weak duality then gives a lower bound on the analytic benchmark. For a congruence with invertible , suppose a directed Gershgorin calculation gives . A lower bound on the smallest eigenvalue of is
The norm factors are replaced by certified upper bounds, with outward rounding in the appropriate direction. The accompanying archive records the exact rational witnesses, interval outputs, arithmetic precision, solver metadata, and cryptographic hashes used to reproduce these directed bounds.
Benchmark construction and convergence
The benchmark generator enumerates a finite latent-trajectory set . The per-round physical-error increments are
corresponding to no error or a single flip on one of the three physical bits. With two rounds, at most two distinct bits can be flipped, so the reachable final masks are 0, 1, 2, 3, 4, 5, 6; mask 7 is unreachable. The eight recovery actions are the masks 0, …, 7, and the loss is zero exactly when the recovery mask equals the final physical-error mask. Each record contains a fault trajectory, its probability , the emitted syndrome labels, and the residual physical-error mask associated with every recovery action. If is the two-bit syndrome in round , the fixed syndrome-to-QPSK lookup assigns a phase index , giving
Temporal factors are ordered chronologically, with round 1 as the leftmost tensor factor; the archive uses the same ordering. The certified calculation constructs the finite coherent-state span directly from the analytic Gram entries in Eq. (9). Arb/Acb ball arithmetic is used for the certified cost operators and objective contractions; tensor products of the resulting coordinates represent complete trajectories without a Fock cutoff. In the parameter-neighborhood verification, different values of are compared in a common analytic four-dimensional frame using explicit isometries, rather than by subtracting unrelated Gram factorizations. For final physical-error mask , the residual mask is . The benchmark loss is
where the Kronecker delta is one when its indices agree and zero otherwise. This is Eq. (7), a comparison of complete masks rather than logical equivalence classes. Substitution of the enumerated , lookup , and loss table into Equation (2) determines every without receiver-dependent choices.
As an independent floating-point check, let denote the numerical Gram factor and the analytic Gram matrix. The check gives
The maximum eigenvalue discrepancy was . The two-norm condition numbers are approximately for the Gram factor and for the Gram matrix. These are independent floating-point consistency diagnostics. The directed theorem bounds instead use the analytic Gram construction and interval procedure of Appendix D.
As an independent convergence check, the same states were also reconstructed in a normalized truncated Fock basis,
For each tested , the cost operators and both optimization objectives were rebuilt. Relative to the reference, the first tested cutoff satisfying a convergence criterion is . At that cutoff, the absolute differences in coherent error, repaired PPT lower bound, and gap are , , and , respectively. This Fock sweep is a numerical cross-check rather than an interval bound; no cutoff enters the certified theorem.
Multi-round PPT relaxation and deeper constructions
By Theorem 1, every final adaptive effect is fully separable across temporal factors. A tractable outer relaxation imposes
together with . Adding temporal partial-transpose constraints tightens the relaxation. Omitting a cut does not invalidate containment, but it may weaken the resulting adaptive bound. For , all eight analytic residual matrices and all twenty-four PPT multipliers were verified at 256 bits; the scalar-repair objective debit was exactly 0.000672565248.
The , and , entries are explicit full-storage streaming realizations constructed using Theorem 3. Their Choi matrices and terminal POVMs are materialized exactly in sparse rational and Gaussian-rational form; the matrices are represented sparsely.
The remaining entries are lower-memory streaming receivers obtained from a restricted prefix-memory construction. Selected prefix computational-basis vectors are preserved coherently, the discarded population is reset to , and the selection keeps the vectors of largest average diagonal mass with deterministic tie-breaking. Only the terminal discrimination POVM is then numerically optimized and certified. Because the construction is restricted, a failure at some does not exclude that dimension globally. The tested dimensions were at and at . For at this measure-and-encode family supports at most terminal labels, so the negative results there are constructional rather than global; this is why the statement remains . For , the relaxation uses the cuts , , and ; for , it uses . In particular, the certificate does not establish an advantage for a constrained receiver.
| Receiver upper bound | PPT lower bound | Certified gap | ||
| 2 | 7 | 0.242099929928125 | 0.251431736761034 | 0.009331806832908 |
| 3 | 8 | 0.294297344448216 | 0.299647463773192 | 0.005350119324975 |
| 3 | 10 | 0.288350172949501 | 0.299647463773192 | 0.011297290823690 |
| 3 | 12 | 0.286175261363248 | 0.299647463773192 | 0.013472202409943 |
| 3 | 16 | 0.284502699489212 | 0.299647463773192 | 0.015144764283980 |
| 4 | 36 | 0.319370057047514 | 0.319542490402679 | 0.000172433355164 |
| 4 | 48 | 0.317851443330621 | 0.319542490402679 | 0.001691047072057 |
| 4 | 64 | 0.317529609881821 | 0.319542490402679 | 0.002012880520857 |
Table IX. Certified depth bounds and explicit streaming receivers. All inequalities are directed; cut metadata are stated in the text.
A separable-optimum example
Separability across the temporal cut does not by itself preclude a coherent advantage. Let and be independent and uniform. The first signal is
where , and label their measurement bases. The second signal is an orthogonal basis flag . Thus
and the final action is a guess with . A two-dimensional coherent memory can retain the first qubit until the basis flag arrives and then measure in the indicated basis, giving zero error. A receiver that must measure the first signal before learning the basis has optimal success probability [32], and hence error . Therefore
The optimal collective measurement is nevertheless separable across the temporal cut,
The example shows that a PPT or separability relaxation cannot detect every temporal-order advantage, and it is not used in any certificate above.
Observation-only depth control
As a separate numerical control, additional observations were introduced without additional data-error opportunities. This differs from both the certified two-round benchmark and the slow timing schedule used for Figure 6. Within that control model, the coherent errors decrease from approximately 8.761% to 4.093% to 1.814% for two, three, and four observations. The corresponding dephasing-threshold parameters are , , and . Here denotes the coherence-retention parameter used in this control calculation. These values pertain only to the observation-only control and are not used in the stored-pulse hardware calculation.
Numerical optimization of streaming updates
Consider a channel , where and are the input and output dimensions. Let be an orthonormal input basis. Its Choi matrix is defined by
Complete positivity and trace preservation are equivalent to
The action of the channel is recovered from
After composing the update maps with the loss-weighted input ensemble, the average risk is multilinear in the channel Choi matrices and linear in the final POVM. Holding every block but one fixed therefore leaves a linear objective over the positive semidefinite and trace-preserving constraints in Equation (81); optimizing that block is an SDP. The final POVM block is likewise an SDP with positivity and completeness constraints.
The streaming search uses block-coordinate optimization. It begins from feasible channel and POVM blocks, replaces one block at a time by the solution of its convex subproblem, and repeats channel and measurement updates until the prescribed stopping criterion is reached. An exactly solved block update cannot increase the objective, but finite solver tolerances can obscure changes at the last reported digits. Figure 4 shows the search values; Table IV gives the certified bounds used for minimality.
Feasibility can be checked independently of the optimization expression. The candidate receivers are evaluated by reconstructing the induced channel action with Equation (82) and recomputing the recovery risk from the resulting output states. The feasibility check records the smallest Choi and POVM eigenvalues, the trace-preservation residuals, and the Frobenius-norm POVM completeness residual . These diagnostics are used to screen the candidate receivers before interval verification.
For the and receivers, the feasibility check found trace-preservation residuals below and Choi minimum eigenvalues no more than below zero. The final POVMs were positive semidefinite and complete to floating-point precision. Before interval verification, the repaired floating-point candidates had risks 0.242099914992664 for and 0.249065671918806 for . The certified theorem instead uses the outward-safe upper bounds in Table VIII.
| Quantity | Certified bound |
| receiver risk | |
| Separate-adaptive risk | |
| separation | |
| receiver risk | |
| separation | |
| Collective optimum | |
| Adaptive strategy risk |
Table VIII. Outward-safe two-round bounds for the analytic benchmark.
Block-coordinate optimization therefore supplies candidate receivers but does not establish their global optimality. Feasibility and outward-safe risk bounds for the receivers used in the certified results are obtained independently by the procedure of Appendix § 1.3.4. The reproducibility archive records the initialization rules, block order, stopping tolerance, completed starts, solver versions, and feasibility residuals for the reported memory sweep.
Clocked quantum-reservoir pilot
We compare homogeneous and clocked reservoir constructions on the pilot benchmark with reservoir memory dimension , where is the dimension of the persistent reservoir register carried between rounds. The clocked and parameter-matched homogeneous families were compared over thirty seeds, grouped into ten matched best-of-three trials. Clocking gave lower error in 29 of 30 raw seed comparisons and in 10 of 10 grouped trials, with a mean difference of 0.00200414101285111 in absolute error, that is 0.200414101285111 percentage points. Because the comparison was not designed as a prespecified inferential study, we report these results
| Quantity | Reported result |
| Raw seed comparisons | 29/30 clocked wins |
| Grouped best-of-three trials | 10/10 clocked wins |
| Mean error reduction | 0.200414101285111 percentage points |
Table X. Descriptive clocked-reservoir comparison. No inferential interval is reported.
descriptively. Within this guided comparison, the clocked family gives lower error than the matched homogeneous family. Because the clocked maps were informed by the optimized coherent receiver, we treat this comparison as exploratory. The per-seed results and the grouping rule are included in the archive; the original training source and checkpoints are not.
Hardware-noise sensitivity
The modeled round duration is
combining a separately simulated pulse duration with assumed memory-interaction and inter-round gap times. All three components are modeling assumptions rather than measurements of the receiver studied here. The three crossings in Figure 6 were located by Brent root finding in , with an unreconstructed joint POVM reoptimized at each trial point.
For context, superconducting microwave-cavity memories have demonstrated long-lived storage [33], tantalum transmons have exceeded under dynamical decoupling [34], and a single-photon cavity qubit reached [35]. Separate experiments have demonstrated superconducting-qubit readout [36] and a measurement-to-conditional-pulse sequence [31]. These records were not achieved together as a coherent sequential receiver, and the result characterizes a fast measurement step rather than coherent capture, storage, and reinteraction. Because maximum latency depends on whether pulse integration, control, and memory interaction overlap, we report lifetime values rather than a single architecture-independent latency ceiling.
| Depth | Longest storage time | ||
| 0.967978 | |||
| 0.981577 | |||
| 0.992463 |
Table XI. Stored-pulse slow-schedule crossing values. The values are model-dependent and do not represent integrated-receiver performance.
Equation (55) models phase retention, whereas amplitude damping is controlled separately by Equation (57). Under a Markovian decomposition, Equation (58) separates the pure-dephasing time from the amplitude-relaxation time . This calculation includes phase retention only; it is not a complete device-noise model. A multilevel memory also requires a specified encoding and decay law for its off-diagonal matrix elements; the cited qubit coherence times alone do not determine those rates. A device-level assessment would vary and independently and propagate timing and control uncertainties.
Flagged direct-sum inheritance criterion
A certified base-sector separation can be inherited by a larger block-diagonal decision problem under the following conditions.
Proposition 5 (Locally flagged inheritance). Let an orthogonal classical flag be accessible to both receiver classes before any operation whose choice depends on that flag. Assume the enlarged receiver classes admit the causal embedding of the base-sector strategies, and that restriction of an enlarged adaptive strategy to the base sector is an allowed base adaptive strategy. Let its register belong wholly to one local temporal tensor factor, and require that no partial transpose used by the certificate splits from that factor. In the fixed flag basis define the pinching map
Suppose the enlarged cost operators are block diagonal,
where is independent of the action. If the base sector has an adaptive lower bound , a fixed receiver with explicit risk , and certified gap , then the enlarged problem has the explicit-versus-certified separation
If the conditional base-sector costs are perturbed by , the right-hand side becomes , with from Equation (23).
Proof. Pinching a feasible POVM in the orthogonal flag basis preserves positivity, completeness, and the block-diagonal objective. Since the flag belongs to one temporal factor and its projectors are diagonal in the transpose basis, for every certified cut ,
Thus pinching also preserves the PPT relaxation and permits sectorwise analysis. The lower bound for the adaptive class uses a separate assumption: restricting an enlarged adaptive protocol to the base sector must give an allowed base-sector adaptive protocol. Its conditional risk is therefore at least . Conversely, the causal-embedding assumption lets the coherent receiver use on the base sector. In the other sector, the action-independent cost gives , irrespective of the receiver. This common contribution cancels, leaving a gap of at least . The perturbed statement follows by applying Proposition 3 to the base-sector costs.
The same reasoning applies when the complementary sector is not action-independent only if a nonnegative explicit-versus-certified contribution is established there separately. The inheritance therefore requires the flag, cost decomposition, and relaxation to respect the same direct-sum structure. For a family indexed by code distance , a uniform inherited floor would require both and , giving
Neither condition has been established for a natural surface-code family; in particular, the probability of a chosen sector may vanish with distance. Thus the proposition is conditional and does not establish scaling for a natural surface-code family.
Acknowledgments
This work was authored in part by the National Laboratory of the Rockies (NLR) for the U.S. Department of Energy (DOE), operated under Contract No. DE-AC36-08GO28308. Funding for the NLR authors provided by the Modeling, Simulation, Optimization Capability for Grid Modernization, Energy Systems Integration, Mobility Electrification (MSOC 4 GEM) at the National Laboratory of the Rockies. The views expressed in the article do not necessarily represent the views of the DOE or the U.S. Government. The U.S. Government retains and the publisher, by accepting the article for publication, acknowledges that the U.S. Government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this work, or allow others to do so, for U.S. Government purposes.
Grammarly was used for grammar checking during the preparation of this manuscript. The authors reviewed and edited the suggested changes and take full responsibility for the final content.
References
References
- [1]C. A. Pattison, M. E. Beverland, M. P. da Silva, and N. Delfosse, Improved quantum error correction using soft information (2021), arXiv:2107.13589 [quant-ph].arxiv.org/abs/2107.13589
- [2]H. Ali, J. Marques, O. Crawford, J. Majaniemi, M. Serra-Peralta, D. Byfield, B. M. Varbanov, B. M. Terhal, L. Di-Carlo, and E. T. Campbell, Physical Review Applied 22, 044031 (2024).
- [3]M. D. Hanisch, B. Hétènyi, and J. R. Wootton, APS Open Science 1, 000019 (2026).
- [4]J. Kerckhoff, H. I. Nurdin, D. S. Pavlichin, and H. Mabuchi, Physical Review Letters 105, 045002 (2010), arXiv:0907.0236.
- [5]D. Crow, R. J. Joynt, and M. Saffman, Physical Review Letters 117, 130503 (2016).
- [6]F. Butt, D. F. Locher, K. Brechtelsbauer, H. P. Büchler, and M. Müller, Science Advances 11, eadv2590 (2025), arXiv:2410.13568.
- [7]F. Butt, I. Pogorelov, R. Freund, A. Steiner, M. Meyer, T. Monz, and M. Müller, Nature Communications 17, 995 (2026).
- [8]K. Brechtelsbauer, F. Butt, D. F. Locher, S. Higuera Quintero, S. Weber, M. Müller, and H. P. Büchler, PRX Quantum 6, 040349 (2025), arXiv:2505.15669.
- [9]A. Tanggara, M. Gu, and K. Bharti, Strategic code: A unified spatio-temporal framework for quantum error-correction (2024), arXiv:2405.17567 [quant-ph].arxiv.org/abs/2405.17567
- [10]C. W. Helstrom, Quantum Detection and Estimation Theory (Academic Press, New York, 1976).
- [11]H. P. Yuen, R. S. Kennedy, and M. Lax, IEEE Transactions on Information Theory 21, 125 (1975).
- [12]J. Watrous, The Theory of Quantum Information (Cambridge University Press, 2018).DOI
- [13]W. Matthews, S. Wehner, and A. Winter, Communications in Mathematical Physics 291, 813 (2009), arXiv:0810.2327.
- [14]A. Peres, Physical Review Letters 77, 1413 (1996).
- [15]M. Horodecki, P. Horodecki, and R. Horodecki, Physics Letters A 223, 1 (1996).
- [16]A. Cosentino, Physical Review A 87, 012321 (2013), arXiv:1205.1031.
- [17]M. Horodecki, P. W. Shor, and M. B. Ruskai, Reviews in Mathematical Physics 15, 629 (2003), arXiv:quant-ph/0302031.
- [18]D. Rosset, F. Buscemi, and Y.-C. Liang, Physical Review X 8, 023033 (2018).
- [19]L. B. Vieira, H.-Y. Ku, and C. Budroni, Physical Review Research 7, 043281 (2025).
- [20]S. Guha, Physical Review Letters 106, 240502 (2011).
- [21]S. M. Barnett and S. Croke, Advances in Optics and Photonics 1, 238 (2009).
- [22]R. Blume-Kohout, S. Croke, and M. Zwolak, Ideal state discrimination with an O(1)-qubit quantum computer (2012), arXiv:1201.6625 [quant-ph].
- [23]A. Klimek, M. Jachura, W. Wasilewski, and K. Banaszek, Journal of Modern Optics 63, 2074 (2016).
- [24]F. Verstraete, J. J. García-Ripoll, and J. I. Cirac, Physical Review Letters 93, 207204 (2004), arXiv:cond-mat/0406426.
- [25]U. Schollwöck, Annals of Physics 326, 96 (2011), arXiv:1008.3477.
- [26]K. Fujii and K. Nakajima, Physical Review Applied 8, 024030 (2017).
- [27]G. Tanaka, T. Yamane, J. B. Héroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Neural Networks 115, 100 (2019).
- [28]A. Blais, A. L. Grimsmo, S. M. Girvin, and A. Wallraff, Reviews of Modern Physics 93, 025005 (2021).
- [29]F. A. Pollock, C. Rodríguez-Rosario, T. Frauenheim, M. Paternostro, and K. Modi, Physical Review A 97, 012127 (2018).
- [30]G. White, Nature Reviews Physics 4, 287 (2022).
- [31]D. Ristè, C. C. Bultink, K. W. Lehnert, and L. DiCarlo, Physical Review Letters 109, 240502 (2012).
- [32]M. A. Ballester, S. Wehner, and A. Winter, IEEE Transactions on Information Theory 54, 4183 (2008), arXiv:quant-ph/0608014.
- [33]M. Reagor, W. Pfaff, C. Axline, R. W. Heeres, N. Ofek, K. Sliwa, E. Holland, C. Wang, J. Blumoff, K. Chou, M. J. Hatridge, L. Frunzio, M. H. Devoret, L. Jiang, and R. J. Schoelkopf, Physical Review B 94, 014506 (2016), arXiv:1508.05882.
- [34]A. P. M. Place, L. V. H. Rodgers, P. Mundada, B. M. Smitham, M. Fitzpatrick, Z. Leng, A. Premkumar, J. Bryon, A. Vrajitoarea, S. Sussman, G. Cheng, T. Madhavan, H. K. Babla, B. Jäck, A. Gyenis, N. Yao, R. J. Cava, N. P. de Leon, and A. A. Houck, Nature Communications 12, 1779 (2021), arXiv:2003.00024.
- [35]O. Milul, B. Guttel, U. Goldblatt, S. Hazanov, L. M. Joshi, D. Chausovsky, N. Kahn, E. Çiftüyürek, F. Lafont, and S. Rosenblum, PRX Quantum 4, 030336 (2023), arXiv:2302.06442.
- [36]E. Jeffrey, D. Sank, J. Y. Mutus, T. C. White, J. Kelly, R. Barends, Y. Chen, Z. Chen, B. Chiaro, A. Dunsworth, A. Megrant, P. J. J. O’Malley, C. Neill, P. Roushan, A. Vainsencher, J. Wenner, A. N. Cleland, and J. M. Martinis, Physical Review Letters 112, 190504 (2014), arXiv:1401.0257.