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

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

Robust qubit–memory transfer against quasi-static frequency noise costs time and exposure: exact bounds from Dubins curves

Carles Marín Muñoz

Published Nov 15, 2026• 1 citation
Publisher

Abstract

A quasi-static offset of the qubit or memory frequency, the slowest part of the noise in a superconducting quantum memory, is described by no Markovian rate. For the exchange of one excitation between a qubit and a memory mode at a coupling bounded by G, we characterise every complete transfer by the times A and B it spends in the qubit and in the memory and by its first-order leakage ζ. For a coupling of any phase and any detuning the duration obeys T ≥ T0 + arccos(|ζ|/z0)/G for |ζ| ≤ z0 and the exposures obey (|ζ| − z0)2 ≤ (A − A0)(B − A0), with T0 = π/(2G), A0 = π/(4G), z0 = 1/(2G); both are sharp, and for a coupling of fixed phase and no detuning they follow from known results on paths of bounded curvature. Cancelling the leakage with a pulse of fixed phase and no detuning costs exactly (γq + γm)/(2G)·arccos(|γq − γm|/(γq + γm)) of first-order population loss above that of the square pulse (a third of it in average infidelity), through the shortest Dubins U-turn; for any pulse, with relaxation only, cancelling it costs at least max(√(γqγm), (π/2) min(γq, γm))/G in the same units. At second order in the offsets every pulse has error [2E(ξq − ξm)2|ζ|2 + E(ξqA + ξmB)2]/6: with noise on the qubit no pulse goes below π2σ2/(96G2), against (8 + π2)σ2/(96G2) for the square pulse; with noise on both elements the square pulse is optimal for 0.137 ≤ σm2/σq2 ≤ 7.30 and explicit shaped pulses beat it outside (0.1264, 7.91), certified by interval arithmetic; for pulses of fixed phase and no detuning, a planar exposure bound derived from the reachable set of the Dubins car gives the exact threshold ratio 0.12645352248541…, the crossing of the bang–bang family. With memory loss the square pulse stops being optimal (for fixed phase and no detuning exactly at σ2 = 7.9078851606875… γmG), and for large σ2/(γmG) a pulse that waits in the memory attains the optimal error to leading order. Positively correlated offsets favour the square pulse, with an elementary sufficient condition that also covers relaxation of both elements, and the band of optimality of the square pulse can only widen as the correlation grows; for fixed phase and no detuning we certify the exact band with correlated offsets and the exact threshold with qubit relaxation at several parameter values, and prove, given the planar bound, that this threshold grows as 2√r + π/2 + 4/√3 + O(r-1/2) in the ratio r of qubit to memory relaxation. English and Spanish versions, LaTeX sources and the scripts behind the certificates are included. The robustness inequalities, the minimisation of the cost, the unitarity step of the second-order expansion and the algebra of the bounds are machine-checked in Isabelle/HOL (https://github.com/karlesmarin/formalizations); the companion site https://karlesmarin.github.io/blueprints/robust-offset/ says, for each statement, where the formal statement ends and the written proof begins. We worked with an AI assistant.

DOI
10.5281/zenodo.24250502

Related papers

Exact limits on writing a superconducting quantum memory: when can pulse shaping beat the resonant swap?

Carles Marín Muñoz

We study quantum-memory writing implemented by the exchange of one excitation between a short-lived superconducting qubit and a long-lived memory mode, at a coupling rate bounded by G, and ask how small pulse shaping can make the error of that exchange. With relaxation at unequal rates and Markovian dephasing, no pulse transfers more population than the plain resonant swap, and at a prescribed duration none beats a resonant pulse combined with a wait in the less lossy element; the proof is a Lyapunov function, with no condition on G. Without dephasing the maximum has a closed form. For the average infidelity of the write operation this gives a lower bound which the resonant pulse misses by less than 1.5·10-6 for a recent phonon-memory processor and attains for one of its modes: the limits are 1.00–1.14 % per access for its three modes against 1.71 % measured on average. Along a fixed small-rate direction the resonant pulse is exactly optimal whenever four times the smaller coherence decay rate exceeds the total pure dephasing rate, and otherwise the undecided width is of second order. Without relaxation and under strong dephasing the optimum is the trivial value 5/12: for equal rates at and above a unique threshold φ*, with 2.0941221534 < φ*/G < 2.0941221542, and, for fixed-phase, zero-detuning controls with 0 ≤ g ≤ G and |φq − φm| ≥ 2√5 G, exactly when the product of the rates is at least 3G2; between 2G and 2√5 G an explicit product of the rates still suffices. A write–wait–read cycle with a reset of the qubit composes exactly, which gives its floor; without the reset, for cycles that close in the absence of noise and to first order in the rates, the optimum reduces to a scalar minimisation: it never completes the swap when the memory decays or dephases, and accessing the memory pays only when the memory loses less than the qubit, and then only after an explicit waiting time. Five recent devices are evaluated within the model. Slow noise is not Markovian, and a counterexample shows that shaping can then help; for a quasi-static offset we give two bounds, and its first- and second-order theory is the subject of a companion paper. English and Spanish versions, LaTeX sources and the scripts behind the numbers, interval enclosures and certificates are included. The bound without dephasing, the quasi-static square-pulse bound, the Lyapunov step of the population bound for continuous controls (given the derivative identities of the master equation) and the elementary steps of the other proofs are machine-checked in Isabelle/HOL (https://github.com/karlesmarin/formalizations); the companion site https://karlesmarin.github.io/blueprints/memory-floors/ says, for each statement, what is formalised, what is certified by computation and what is a written proof. We worked with an AI assistant.

•Nov 15, 2026•1 citation

Exact Robustness Frontiers and Fidelity Limits for Qubit Memories

Ying Xie

We determine exact resource limits for a qubit memory with retained classical records. For every qubit noise channel A, the largest output generalized entanglement-breaking robustness under an input budget R_G ≤ r is a semidefinite program in two 4 × 4 positive matrices, attained with at most five classical outcomes. For amplitude damping we solve the full budget curve, including two mixed-state regimes, and attain it with one classical bit. Its contraction coefficient is one, approached with vanishing input resource, although every fixed positive resource budget contracts strictly. A different argument controls memories prepared from a perfect qubit: for prescribed independent noise channels A_1, …, A_n, arbitrary adaptive instruments and finite classical records obey R_G(Γ_n) ≤ R(Γ_n) ≤ ∏_j ν(A_j), where ν(A) = max{0, 2λ_max J_A − 1} and J_A is normalized. The proof propagates a positive 2 × 2 operator over the branches. For amplitude-damping survivals with product b, both robustness optima equal b; the optimal average and worst-case pure-state fidelities are (2+b)/3, the optimal Bell fidelity is (1+b)/2, and the minimum half-diamond error is (1−b)/2. Finite protocols attain all four operational optima simultaneously. Thus classical control can improve fidelity while leaving optimal robustness unchanged. The model permits one persistent qubit and excludes quantum bypass, environment access, and postselection. A visual guide records the reasoning routes, their joint premises, and the distinction between prior tools and the results developed here.

•Nov 12, 2026•0 citations

A Closed-Form Excess-Population Latency Budget for Active Reset of Superconducting Qubits

Kuzey Kaya

Active reset returns a superconducting qubit to its ground state: read it out, apply a π-pulse if it reads excited. The classical decision between the two takes time, and the qubit relaxes meanwhile. Salathé and co-workers tied this cost to the ratio of feedback latency to lifetime and expected the error to scale as one minus its exponential, and Ristè and co-workers had already written the reset error to first order as readout errors plus equilibration terms linear in the latency. This work makes that accounting exact at a fixed evaluation time. A transmon with T₁ = 50 μs is prepared in its excited state and read out by an ideal terminal projection followed by a classical confusion channel with symmetric assignment error P_AE. Under that instrument the ensemble-averaged residual population after one feedback cycle is an exact weighted sum of four deterministic Lindblad evolutions, conditioned on the projected and the reported outcome. The sum is evaluated in closed form, together with the critical latency τ_crit at which the residual exceeds its zero-latency value by a chosen budget, defined throughout as an absolute increase in excitation probability of 0.01, that is one percentage point. To leading order τ_crit is the budget times T₁ divided by the pulse-firing probability p_pulse, times a prefactor of 1.019 for the parameters considered here. Discarding that prefactor while retaining the division by p_pulse costs under 2%; the further simplification to budget times T₁ carries 3.4% to 7.3% error across P_AE = 1% to 5%. Because P_AE enters only through p_pulse, the budget is weakly sensitive to assignment error over the tested range for an initially excited qubit, giving 518 to 539 ns; the readout duration enters separately, through the observation window and the post-readout population. A critical latency exists only for budgets below the maximum excess attainable within that window. The four-branch representation is exact for the stated measurement model; the finite-pulse map used to evaluate it is approximate, and is replaced here by its exact affine propagator. The closed form agrees with a master-equation solver to 3.3×10⁻⁷ and is cross-checked against stochastic trajectories.

•Oct 10, 2026•0 citations

Causal quantum-channel simulation: memory beyond entropy

Seth Douglas, Nidhal Mghirbi

How much memory does it take to apply the same quantum channel many times, when each output must be released before the next input arrives and every retained qubit counts? One-use invariants do not decide it. Two explicit channels on dimension 1664 have identical normalized Choi spectra (32 eigenvalues 1/32), the same maximal complementary entropy 5, and zero asymptotic purity cost. At the same optimal exchange rate, one is implemented exactly with five memory qubits for any number of uses, while the other needs Omega(n^(1/(2 log_2(107)+2))) memory qubits for n uses, at every fixed error at most 1/16 and whatever purity is available; memory O(n^(1/3) log^(4/3) n) suffices for it. The hard channel is built from Houghton's group H_3. Behind the example is a general theory: the achievable exchange and purity rates, the purity cost kappa, which vanishes exactly on the closure of finite tracial factorizations, and memory bounds from how well a channel's environment is approximated by finite ones, and, for channels built from finite presentations, from the approximation profile of the group. Version 1.1 adds where the expensive channels lie: on the relative boundary of the closure; any fixed depolarizing admixture brings them down to polylogarithmic memory at the rank rate; on a face containing the Houghton channel a single moment of a trace on the group decides expense; and one free fermion per site already forces polynomial memory. Version 1.1.1 adds a map of what the main separation depends on (published inputs, companion results and new arguments), sets finite-budget and asymptotic-rate statements side by side, cites the companion papers at their published versions, pins the tested software versions, and makes every check refuse to run under optimized Python. The mathematical results are unchanged. Version 1.1.2 corrects attributions and makes two headline statements precise. The separation of Theorem A is stated at the exchange budget 5n/2, and the channels that need polynomial memory at every exchange rate are those with purity limited to O(log n). Credits added or corrected: Bisio, D'Ariano, Perinotti and Chiribella, and Bisio, D'Ariano, Perinotti and Sedlak; Boes et al. for the square-root dephasing construction; Lie, Son, Boes, Ng and Wilming for the exact case of the tracialization lemma; Faist, Berta and Brandao for the relation to the thermodynamic capacity; Haagerup, Musat and Rordam for the moment criteria; Thoma, Johnson, Watrous, Horodecki-Oppenheim-Winter and Lie-Jeong; Cornulier for the Baumslag-Solitar bound. Known results are moved out of numbered statements. The mathematical results and the numbering are unchanged. All operational bounds use complete-experiment error against adaptive observers with quantum references. The general rate-region achievability result uses a cited closed-device theorem; the explicit separation does not. This preprint archive includes the PDF, TeX source, finite verification scripts, and scoped Lean arithmetic proofs. The artifacts do not formally verify the entire manuscript; their scope and external dependencies are documented in verification/README.md.

•Sep 29, 2026•1 citation

Quantum state transfer through a lossy bus: fundamental error limits for every control pulse, with application to IBM's meter-scale interconnects

Carles Marín Muñoz

How small can pulse shaping make the error of quantum state transfer between two qubits, such as superconducting qubits in different modules, through a lossy bus (a resonator, a waveguide or a cable mode)? We give bounds that hold for every control pulse, every duration and any number of bus modes, knowing only the maximal coupling strength G, the qubit loss rate γ and the bus loss rate κ ≥ γ. The population delivered in time T obeys P(T) ≤ e^{−γT}/[1 + (κ−γ)/(G²T)]: going slowly costs qubit loss, going fast costs bus loss, and no pulse escapes both. The end qubits cannot be bypassed either: in the frame where the qubits are lossless, every transfer keeps the excitation in them for an integrated time bounded below by a quantity that tends to π/(2G). As γ/(κ−γ) → 0 and √(γ(κ−γ))/G → 0, the error cannot fall below 2√(2γ(κ−γ))/G (1 − o(1)); this constant was known as the optimum of perturbative and quasi-adiabatic pulse optimisation; here it bounds every control, and an explicit pulse attains it with explicit error terms. Markovian dephasing enters through the same equations with rates 2/T₂; with a different coherence time for each element the bounds give the floor of a real device, and a reciprocity relation makes the direction of transfer irrelevant. For the four links of IBM's meter-scale relay experiment (arXiv:2610.29986), no pulse can bring the error below 6.03–10.05 × 10⁻⁴ (certified with interval arithmetic) if dephasing is Markovian at the measured T₂*, and not below 1.5–1.8 × 10⁻⁴ for any dephasing noise; the measured errors are 11–16 times larger. Doubling the coupling halves the floor; the dephasing of the worse qubit is the next lever; a 26-ns trapezoidal pulse at 18 MHz brings all four links below 10⁻³ within the model. The floor, the occupation lemma, the frontier lemma and the window comparisons are machine-checked in Isabelle/HOL, and every number comes with the script that produces it. Part II of a study whose Part I is «Certified Speed-Limit Diagnostics for Deployed Superconducting Quantum Processors». Companion site (blueprint, with the laboratory «Check your link»): https://karlesmarin.github.io/blueprints/relay-floors/. English and Spanish versions; sources and supplementary package (scripts, data, Isabelle theories, certificates).

•Oct 31, 2026•0 citations

Probing Qubit Memory Errors at the Part-per-Million Level

M. A. Sepiol, A. C. Hughes, J. E. Tarlton +7 more

Robust qubit memory is essential for quantum computing, both for near-term devices operating without error correction, and for the long-term goal of a fault-tolerant processor. We directly measure the memory error ${\ensuremath{\epsilon}}_{m}$ for a $^{43}{\mathrm{Ca}}^{+}$ trapped-ion qubit in the small-error regime and find ${\ensuremath{\epsilon}}_{m}<{10}^{\ensuremath{-}4}$ for storage times $t\ensuremath{\lesssim}50\text{ }\text{ }\mathrm{ms}$. This exceeds gate or measurement times by three orders of magnitude. Using randomized benchmarking, at $t=1\text{ }\text{ }\mathrm{ms}$ we measure ${\ensuremath{\epsilon}}_{m}=1.2(7)\ifmmode\times\else\texttimes\fi{}{10}^{\ensuremath{-}6}$, around ten times smaller than that extrapolated from the ${T}_{2}^{*}$ time, and limited by instability of the atomic clock reference used to benchmark the qubit.

arXiv:1905.06878•Physical Review Letters•Sep 13, 2019•31 citations
A quasi-static offset of the qubit or memory frequency, the slowest part of the noise in a superconducting quantum memory, is described by no Markovian rate. For the exchange of one excitation between a qubit and a memory mode at a coupling bounded by G, we characterise every complete transfer by the times A and B it spends in the qubit and in the memory and by its first-order leakage ζ. For a coupling of any phase and any detuning the duration obeys T ≥ T0 + arccos(|ζ|/z0)/G for |ζ| ≤ z0 and the exposures obey (|ζ| − z0)2 ≤ (A − A0)(B − A0), with T0 = π/(2G), A0 = π/(4G), z0 = 1/(2G); both are sharp, and for a coupling of fixed phase and no detuning they follow from known results on paths of bounded curvature. Cancelling the leakage with a pulse of fixed phase and no detuning costs exactly (γq + γm)/(2G)·arccos(|γq − γm|/(γq + γm)) of first-order population loss above that of the square pulse (a third of it in average infidelity), through the shortest Dubins U-turn; for any pulse, with relaxation only, cancelling it costs at least max(√(γqγm), (π/2) min(γq, γm))/G in the same units. At second order in the offsets every pulse has error [2E(ξq − ξm)2|ζ|2 + E(ξqA + ξmB)2]/6: with noise on the qubit no pulse goes below π2σ2/(96G2), against (8 + π2)σ2/(96G2) for the square pulse; with noise on both elements the square pulse is optimal for 0.137 ≤ σm2/σq2 ≤ 7.30 and explicit shaped pulses beat it outside (0.1264, 7.91), certified by interval arithmetic; for pulses of fixed phase and no detuning, a planar exposure bound derived from the reachable set of the Dubins car gives the exact threshold ratio 0.12645352248541…, the crossing of the bang–bang family. With memory loss the square pulse stops being optimal (for fixed phase and no detuning exactly at σ2 = 7.9078851606875… γmG), and for large σ2/(γmG) a pulse that waits in the memory attains the optimal error to leading order. Positively correlated offsets favour the square pulse, with an elementary sufficient condition that also covers relaxation of both elements, and the band of optimality of the square pulse can only widen as the correlation grows; for fixed phase and no detuning we certify the exact band with correlated offsets and the exact threshold with qubit relaxation at several parameter values, and prove, given the planar bound, that this threshold grows as 2√r + π/2 + 4/√3 + O(r-1/2) in the ratio r of qubit to memory relaxation. English and Spanish versions, LaTeX sources and the scripts behind the certificates are included. The robustness inequalities, the minimisation of the cost, the unitarity step of the second-order expansion and the algebra of the bounds are machine-checked in Isabelle/HOL (https://github.com/karlesmarin/formalizations); the companion site https://karlesmarin.github.io/blueprints/robust-offset/ says, for each statement, where the formal statement ends and the written proof begins. We worked with an AI assistant.
We study quantum-memory writing implemented by the exchange of one excitation between a short-lived superconducting qubit and a long-lived memory mode, at a coupling rate bounded by G, and ask how small pulse shaping can make the error of that exchange. With relaxation at unequal rates and Markovian dephasing, no pulse transfers more population than the plain resonant swap, and at a prescribed duration none beats a resonant pulse combined with a wait in the less lossy element; the proof is a Lyapunov function, with no condition on G. Without dephasing the maximum has a closed form. For the average infidelity of the write operation this gives a lower bound which the resonant pulse misses by less than 1.5·10-6 for a recent phonon-memory processor and attains for one of its modes: the limits are 1.00–1.14 % per access for its three modes against 1.71 % measured on average. Along a fixed small-rate direction the resonant pulse is exactly optimal whenever four times the smaller coherence decay rate exceeds the total pure dephasing rate, and otherwise the undecided width is of second order. Without relaxation and under strong dephasing the optimum is the trivial value 5/12: for equal rates at and above a unique threshold φ*, with 2.0941221534 < φ*/G < 2.0941221542, and, for fixed-phase, zero-detuning controls with 0 ≤ g ≤ G and |φq − φm| ≥ 2√5 G, exactly when the product of the rates is at least 3G2; between 2G and 2√5 G an explicit product of the rates still suffices. A write–wait–read cycle with a reset of the qubit composes exactly, which gives its floor; without the reset, for cycles that close in the absence of noise and to first order in the rates, the optimum reduces to a scalar minimisation: it never completes the swap when the memory decays or dephases, and accessing the memory pays only when the memory loses less than the qubit, and then only after an explicit waiting time. Five recent devices are evaluated within the model. Slow noise is not Markovian, and a counterexample shows that shaping can then help; for a quasi-static offset we give two bounds, and its first- and second-order theory is the subject of a companion paper. English and Spanish versions, LaTeX sources and the scripts behind the numbers, interval enclosures and certificates are included. The bound without dephasing, the quasi-static square-pulse bound, the Lyapunov step of the population bound for continuous controls (given the derivative identities of the master equation) and the elementary steps of the other proofs are machine-checked in Isabelle/HOL (https://github.com/karlesmarin/formalizations); the companion site https://karlesmarin.github.io/blueprints/memory-floors/ says, for each statement, what is formalised, what is certified by computation and what is a written proof. We worked with an AI assistant.
We determine exact resource limits for a qubit memory with retained classical records. For every qubit noise channel A, the largest output generalized entanglement-breaking robustness under an input budget R_G ≤ r is a semidefinite program in two 4 × 4 positive matrices, attained with at most five classical outcomes. For amplitude damping we solve the full budget curve, including two mixed-state regimes, and attain it with one classical bit. Its contraction coefficient is one, approached with vanishing input resource, although every fixed positive resource budget contracts strictly. A different argument controls memories prepared from a perfect qubit: for prescribed independent noise channels A_1, …, A_n, arbitrary adaptive instruments and finite classical records obey R_G(Γ_n) ≤ R(Γ_n) ≤ ∏_j ν(A_j), where ν(A) = max{0, 2λ_max J_A − 1} and J_A is normalized. The proof propagates a positive 2 × 2 operator over the branches. For amplitude-damping survivals with product b, both robustness optima equal b; the optimal average and worst-case pure-state fidelities are (2+b)/3, the optimal Bell fidelity is (1+b)/2, and the minimum half-diamond error is (1−b)/2. Finite protocols attain all four operational optima simultaneously. Thus classical control can improve fidelity while leaving optimal robustness unchanged. The model permits one persistent qubit and excludes quantum bypass, environment access, and postselection. A visual guide records the reasoning routes, their joint premises, and the distinction between prior tools and the results developed here.
Active reset returns a superconducting qubit to its ground state: read it out, apply a π-pulse if it reads excited. The classical decision between the two takes time, and the qubit relaxes meanwhile. Salathé and co-workers tied this cost to the ratio of feedback latency to lifetime and expected the error to scale as one minus its exponential, and Ristè and co-workers had already written the reset error to first order as readout errors plus equilibration terms linear in the latency. This work makes that accounting exact at a fixed evaluation time. A transmon with T₁ = 50 μs is prepared in its excited state and read out by an ideal terminal projection followed by a classical confusion channel with symmetric assignment error P_AE. Under that instrument the ensemble-averaged residual population after one feedback cycle is an exact weighted sum of four deterministic Lindblad evolutions, conditioned on the projected and the reported outcome. The sum is evaluated in closed form, together with the critical latency τ_crit at which the residual exceeds its zero-latency value by a chosen budget, defined throughout as an absolute increase in excitation probability of 0.01, that is one percentage point. To leading order τ_crit is the budget times T₁ divided by the pulse-firing probability p_pulse, times a prefactor of 1.019 for the parameters considered here. Discarding that prefactor while retaining the division by p_pulse costs under 2%; the further simplification to budget times T₁ carries 3.4% to 7.3% error across P_AE = 1% to 5%. Because P_AE enters only through p_pulse, the budget is weakly sensitive to assignment error over the tested range for an initially excited qubit, giving 518 to 539 ns; the readout duration enters separately, through the observation window and the post-readout population. A critical latency exists only for budgets below the maximum excess attainable within that window. The four-branch representation is exact for the stated measurement model; the finite-pulse map used to evaluate it is approximate, and is replaced here by its exact affine propagator. The closed form agrees with a master-equation solver to 3.3×10⁻⁷ and is cross-checked against stochastic trajectories.
How much memory does it take to apply the same quantum channel many times, when each output must be released before the next input arrives and every retained qubit counts? One-use invariants do not decide it. Two explicit channels on dimension 1664 have identical normalized Choi spectra (32 eigenvalues 1/32), the same maximal complementary entropy 5, and zero asymptotic purity cost. At the same optimal exchange rate, one is implemented exactly with five memory qubits for any number of uses, while the other needs Omega(n^(1/(2 log_2(107)+2))) memory qubits for n uses, at every fixed error at most 1/16 and whatever purity is available; memory O(n^(1/3) log^(4/3) n) suffices for it. The hard channel is built from Houghton's group H_3. Behind the example is a general theory: the achievable exchange and purity rates, the purity cost kappa, which vanishes exactly on the closure of finite tracial factorizations, and memory bounds from how well a channel's environment is approximated by finite ones, and, for channels built from finite presentations, from the approximation profile of the group. Version 1.1 adds where the expensive channels lie: on the relative boundary of the closure; any fixed depolarizing admixture brings them down to polylogarithmic memory at the rank rate; on a face containing the Houghton channel a single moment of a trace on the group decides expense; and one free fermion per site already forces polynomial memory. Version 1.1.1 adds a map of what the main separation depends on (published inputs, companion results and new arguments), sets finite-budget and asymptotic-rate statements side by side, cites the companion papers at their published versions, pins the tested software versions, and makes every check refuse to run under optimized Python. The mathematical results are unchanged. Version 1.1.2 corrects attributions and makes two headline statements precise. The separation of Theorem A is stated at the exchange budget 5n/2, and the channels that need polynomial memory at every exchange rate are those with purity limited to O(log n). Credits added or corrected: Bisio, D'Ariano, Perinotti and Chiribella, and Bisio, D'Ariano, Perinotti and Sedlak; Boes et al. for the square-root dephasing construction; Lie, Son, Boes, Ng and Wilming for the exact case of the tracialization lemma; Faist, Berta and Brandao for the relation to the thermodynamic capacity; Haagerup, Musat and Rordam for the moment criteria; Thoma, Johnson, Watrous, Horodecki-Oppenheim-Winter and Lie-Jeong; Cornulier for the Baumslag-Solitar bound. Known results are moved out of numbered statements. The mathematical results and the numbering are unchanged. All operational bounds use complete-experiment error against adaptive observers with quantum references. The general rate-region achievability result uses a cited closed-device theorem; the explicit separation does not. This preprint archive includes the PDF, TeX source, finite verification scripts, and scoped Lean arithmetic proofs. The artifacts do not formally verify the entire manuscript; their scope and external dependencies are documented in verification/README.md.
How small can pulse shaping make the error of quantum state transfer between two qubits, such as superconducting qubits in different modules, through a lossy bus (a resonator, a waveguide or a cable mode)? We give bounds that hold for every control pulse, every duration and any number of bus modes, knowing only the maximal coupling strength G, the qubit loss rate γ and the bus loss rate κ ≥ γ. The population delivered in time T obeys P(T) ≤ e^{−γT}/[1 + (κ−γ)/(G²T)]: going slowly costs qubit loss, going fast costs bus loss, and no pulse escapes both. The end qubits cannot be bypassed either: in the frame where the qubits are lossless, every transfer keeps the excitation in them for an integrated time bounded below by a quantity that tends to π/(2G). As γ/(κ−γ) → 0 and √(γ(κ−γ))/G → 0, the error cannot fall below 2√(2γ(κ−γ))/G (1 − o(1)); this constant was known as the optimum of perturbative and quasi-adiabatic pulse optimisation; here it bounds every control, and an explicit pulse attains it with explicit error terms. Markovian dephasing enters through the same equations with rates 2/T₂; with a different coherence time for each element the bounds give the floor of a real device, and a reciprocity relation makes the direction of transfer irrelevant. For the four links of IBM's meter-scale relay experiment (arXiv:2610.29986), no pulse can bring the error below 6.03–10.05 × 10⁻⁴ (certified with interval arithmetic) if dephasing is Markovian at the measured T₂*, and not below 1.5–1.8 × 10⁻⁴ for any dephasing noise; the measured errors are 11–16 times larger. Doubling the coupling halves the floor; the dephasing of the worse qubit is the next lever; a 26-ns trapezoidal pulse at 18 MHz brings all four links below 10⁻³ within the model. The floor, the occupation lemma, the frontier lemma and the window comparisons are machine-checked in Isabelle/HOL, and every number comes with the script that produces it. Part II of a study whose Part I is «Certified Speed-Limit Diagnostics for Deployed Superconducting Quantum Processors». Companion site (blueprint, with the laboratory «Check your link»): https://karlesmarin.github.io/blueprints/relay-floors/. English and Spanish versions; sources and supplementary package (scripts, data, Isabelle theories, certificates).
Robust qubit memory is essential for quantum computing, both for near-term devices operating without error correction, and for the long-term goal of a fault-tolerant processor. We directly measure the memory error ${\ensuremath{\epsilon}}_{m}$ for a 43Ca+^{43}{\mathrm{Ca}}^{+}43Ca+ trapped-ion qubit in the small-error regime and find ${\ensuremath{\epsilon}}_{m}<{10}^{\ensuremath{-}4}$ for storage times $t\ensuremath{\lesssim}50\text{ }\text{ }\mathrm{ms}$. This exceeds gate or measurement times by three orders of magnitude. Using randomized benchmarking, at t=1  mst=1\text{ }\text{ }\mathrm{ms}t=1  ms we measure ${\ensuremath{\epsilon}}_{m}=1.2(7)\ifmmode\times\else\texttimes\fi{}{10}^{\ensuremath{-}6}$, around ten times smaller than that extrapolated from the T2∗{T}_{2}^{*}T2∗​ time, and limited by instability of the atomic clock reference used to benchmark the qubit.