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

Holding a Qubit: The Retention Price of Quantum Memory

William Alexander Aldred

Published Nov 10, 2026• 0 citations
Publisher

Abstract

Classical retention costs almost nothing: a bit in DRAM costs 1.57e-8 J per bit-hour in self-refresh, a bit in flash at most 5.70e-14 J per bit-hour as an amortised rewrite bound, and a bit on tape on a shelf costs nothing. This paper prices the quantum analogue: holding one logical qubit, with no computation performed on it, below a cumulative failure probability of 1% for one hour, one day, and one year. Because bare qubits survive only microseconds (superconducting) to minutes (trapped ion), retention requires continuous surface-code error correction at a 100% refresh duty cycle, with code distance growing as the interval lengthens. Holding one logical qubit takes 15,741 W, 17,569 W, and 20,784 W continuously on the superconducting platform (physical error rate 1e-3, 1 microsecond cycles, distances 27/31/37, 10 kW fixed cryostat plant included) and 2,721 W, 3,057 W, and 3,681 W on the trapped-ion platform (1 ms cycles, distances 19/23/29). The energy per logical-qubit-hour exceeds the classical DRAM bit-hour by a factor of 6.23e14 to 4.76e15, the flash bit-hour by about 1e21, and is unbounded against tape on a shelf at zero holding cost. Quantum memory with present architectures is not a storage medium but a continuous power commitment. Model calculations are labelled SIMULATED, NOT MEASURED. The deposit includes the paper PDF and markdown, the computation script (compute.py) that produces every number in the paper, and its outputs (results.json, results.txt).

DOI
10.5281/zenodo.24110764

Related papers

The Energy Budget of Quantum Error Correction: Power and Energy per Logical Qubit for a Surface-Code Architecture

William Alexander Aldred

Model calculation (SIMULATED, NOT MEASURED) of the energy budget of quantum error correction for a rotated surface-code architecture. At a physical error rate of 0.001 per operation, reaching a logical error probability of 1e-12 per syndrome cycle requires code distance 27 and 1457 physical qubits per logical qubit. Combining the cryogenic load, evaluated with the refrigerator coefficient-of-performance penalty at 4 K and at 15 mK, with room-temperature control and decoding power gives 3.94 W per physical qubit, 5741 W per logical qubit, and 0.155 J per logical operation at a 1 microsecond syndrome cycle time. This is an overhead of about 7.7e9 relative to a practical classical operation at 20 pJ. Room-temperature control and decoding dominate the budget. All numbers are produced by the deposited computation script.

•Nov 10, 2026•0 citations

The Price of a Secret Bit: Energy per Delivered Key Bit in Quantum Key Distribution

William Alexander Aldred

Quantum key distribution is sold on security, never on energy. This paper prices a stated decoy-state BB84 link end to end: the secret-key rate is computed from the standard asymptotic decoy-state yield model (Lo, Ma and Chen 2005; Ma et al. 2005) at a 1 GHz clock, mean photon number 0.5, 0.2 dB/km fibre loss, and published detector classes, and divided into the stated total wall-plug power including detector cooling and classical post-processing. With InGaAs avalanche photodiodes (190 W total) the link delivers 7.68 Mbit/s at zero distance at 2.48e-05 J per secret bit and 57.37 kbit/s at 100 km at 3.31e-03 J per bit, with no key beyond about 134 km. With superconducting nanowire detectors (1690 W total, including a 1500 W closed-cycle cryocooler) it delivers 44.72 Mbit/s at zero distance at 3.78e-05 J per bit, 431.62 kbit/s at 100 km at 3.92e-03 J per bit, and 41.32 kbit/s at 150 km at 4.09e-02 J per bit, with no key beyond about 222 km. Energy per bit rises by a factor of 133.8 (avalanche) and 103.6 (nanowire) between 0 km and 100 km. A stated classical X25519 handshake costs 1.00e-03 J, or 3.91e-06 J per key bit, so the quantum link is 848 times (avalanche) and 1002 times (nanowire) as expensive per key bit at 100 km. Information-theoretic key security costs microjoules per bit at short range and millijoules per bit at 100 km, rising exponentially thereafter. Model calculations are labelled SIMULATED, NOT MEASURED. The deposit includes the paper PDF and text, the computation script (compute.py) that produces every number in the paper, and its outputs (results.json, results.txt).

•Nov 10, 2026•0 citations

Error Correction Is a Landauer Machine: The Thermodynamic Floor of Quantum Error-Correction Overhead

Rowan Brad Quni-Gudzinas

Quantum error correction is priced in combinatorial currency — rate, distance, threshold — while the physical cost of performing correction goes uncounted. Every cycle of active (measurement-based) error correction erases information: syndrome extraction, majority voting, and ancilla re-initialization each reset redundancy, and Landauer's principle prices each erasure at no less than kT·ln2 per bit. This paper derives the resulting thermodynamic floor for correction overhead — (n−k)/k · kT·ln2 per logical qubit per round for an [[n,k]] code — verifies its scaling across code families (linear for repetition, quadratic for surface, constant for constant-rate), and applies the JPCUB joules-per-correct-answer metric to rank correction architectures. A two-level nested (tree-structured) code is compared with flat Hamming codes of equal rate on a clustered NAND-like channel: hierarchical decoding cuts the per-round erasure count by a factor of 1.6 to 3, while the naive construction pays a residual-error penalty — a partial null that leaves the floor claim intact and defines the protocol for an industrial flash-memory test. Biological systems that run robust quantum processes without correction (photosynthetic energy transfer, radical-pair magnetoreception) show the floor is an architecture choice: structure prepays in capital energy what correction pays in operating erasures. Autonomous and dissipative correction is identified as the boundary of the claim. A closing section documents the 2026 quantum-algorithms-and-QEC curriculum snapshot: the energy bill the paper prices is absent from the field's own teaching materials as of 2026-08-26.

•Aug 26, 2026•0 citations

A Lower Bound on Energy Cost per Logical-Qubit Operation in Surface-Code Quantum Error Correction

Rowan Brad Quni-Gudzinas

Surface-code quantum error correction (QEC) is pivotal for fault-tolerant quantum computing, but its energy efficiency remains underexplored. This work derives a practical lower bound on the energy cost per logical-qubit operation, accounting for gate-fidelity decay's impact on code distance and physical-qubit overhead. By modeling the energy consumption of gate operations and error correction, we express the total cost as $ E = (E_g N_g + E_c N_c)/N_l $, where $ E_g $, $ E_c $, $ N_g $, $ N_c $, and $ N_l $ represent gate and physical-qubit energy costs, gate counts, physical-qubit counts, and logical qubit counts, respectively. Gate-fidelity decay increases the required code distance $ d $, scaling $ N_c $ and thus $ E $. We propose normalizing computational cost in joules per logical-qubit operation (JPCUB), isolating logical work from error-correction overhead. Quantitative predictions for $ E_g $, $ E_c $, and $ N_c $ remain [to verify], but the functional dependence on $ d $ and gate fidelities is established. This framework informs energy-efficient QEC design and benchmarks. ## 1. Introduction Surface codes enable fault-tolerant quantum computation by encoding logical qubits in a 2D lattice of physical qubits, correcting errors through syndrome measurements. However, the energy cost per logical operation—critical for scalability—remains poorly quantified. While theoretical foundations exist (Raussendorf & Terhal, 2006; Kitaev, 2003), prior work focuses on error-correction thresholds rather than energy efficiency. Here, we address this gap by deriving a lower bound on energy cost, considering how gate-fidelity decay necessitates larger code distances $ d $, increasing physical-qubit overhead $ N_c $. This analysis connects to the joules-per-compute (JPCUB) benchmark, advocating for energy normalization per logical qubit to disentangle computational work from error-correction costs. ## 2. Background Surface codes encode logical qubits using a 2D grid of

•Sep 3, 2026•0 citations

Single-Shot Error Correction at Optimal Spacetime Cost

Kishor Bharti, Tobias Haug, Derek Khu +1 more

Recent work established that, for independent erasures, storing $K$ logical qubits for $S$ time steps with error at most $\varepsilon$ requires a spacetime cost of $Ω(S(K+\log(S/\varepsilon)))$. A matching construction was also given, but assumes ideal error correction. Here we show that the same scaling can be achieved with an explicit noisy error-correction circuit and efficient decoding, counting all state-preparation, gate, measurement, and wait locations, provided that the hardware supports long-range qubit connectivity and fast, reliable classical processing. Thus, reliability adds only a logarithmic overhead shared by all stored qubits. The construction works for sufficiently weak but otherwise general circuit noise, including faults during syndrome extraction and correlated faults, and for independent erasures it matches the best known lower bound up to constant factors. We attain this scaling using quantum Tanner codes, with each memory step consisting of one round of syndrome extraction followed by a fixed number of parallel classical decoder steps that reduce the residual error enough to keep later faults correctable, regardless of memory size, storage time, or target accuracy. Together, this gives exponentially small failure probability per step while keeping the cost of each memory step linear in the code size. The same analysis also extends to certain constant-depth logical Clifford operations.

quant-pharXiv:2610.03684•Oct 5, 2026•0 citations

Idle Decoherence Within One Syndrome Cycle: Duty-Cycle Arithmetic for Data Qubits Waiting on Measurement

William Alexander Aldred

In every syndrome cycle a data qubit spends most of its time doing nothing, waiting while its neighbouring measure qubits are read out and reset, and nothing is not free: idle decoherence accrues at a rate set by the ratio of waiting time to coherence time. This paper prices the wait. For a superconducting cycle of 1.0 microsecond in which entangling gates occupy a quarter of the schedule, against an energy-relaxation time of 100 microseconds, the idle error per data qubit per cycle is about 7.50e-3, which across the 441 data qubits of a distance-twenty-one patch amounts to an expectation of 3.3 idle error events per cycle, comparable to the entire gate-error contribution the cycle was designed around. The wait is a first-class error source, and the paper budgets the three ways to pay it down: faster cycles, longer coherence, and schedules that keep data qubits usefully busy, including dynamical decoupling, whose own costs are charged.

•Nov 15, 2026•0 citations
Classical retention costs almost nothing: a bit in DRAM costs 1.57e-8 J per bit-hour in self-refresh, a bit in flash at most 5.70e-14 J per bit-hour as an amortised rewrite bound, and a bit on tape on a shelf costs nothing. This paper prices the quantum analogue: holding one logical qubit, with no computation performed on it, below a cumulative failure probability of 1% for one hour, one day, and one year. Because bare qubits survive only microseconds (superconducting) to minutes (trapped ion), retention requires continuous surface-code error correction at a 100% refresh duty cycle, with code distance growing as the interval lengthens. Holding one logical qubit takes 15,741 W, 17,569 W, and 20,784 W continuously on the superconducting platform (physical error rate 1e-3, 1 microsecond cycles, distances 27/31/37, 10 kW fixed cryostat plant included) and 2,721 W, 3,057 W, and 3,681 W on the trapped-ion platform (1 ms cycles, distances 19/23/29). The energy per logical-qubit-hour exceeds the classical DRAM bit-hour by a factor of 6.23e14 to 4.76e15, the flash bit-hour by about 1e21, and is unbounded against tape on a shelf at zero holding cost. Quantum memory with present architectures is not a storage medium but a continuous power commitment. Model calculations are labelled SIMULATED, NOT MEASURED. The deposit includes the paper PDF and markdown, the computation script (compute.py) that produces every number in the paper, and its outputs (results.json, results.txt).
Model calculation (SIMULATED, NOT MEASURED) of the energy budget of quantum error correction for a rotated surface-code architecture. At a physical error rate of 0.001 per operation, reaching a logical error probability of 1e-12 per syndrome cycle requires code distance 27 and 1457 physical qubits per logical qubit. Combining the cryogenic load, evaluated with the refrigerator coefficient-of-performance penalty at 4 K and at 15 mK, with room-temperature control and decoding power gives 3.94 W per physical qubit, 5741 W per logical qubit, and 0.155 J per logical operation at a 1 microsecond syndrome cycle time. This is an overhead of about 7.7e9 relative to a practical classical operation at 20 pJ. Room-temperature control and decoding dominate the budget. All numbers are produced by the deposited computation script.
Quantum key distribution is sold on security, never on energy. This paper prices a stated decoy-state BB84 link end to end: the secret-key rate is computed from the standard asymptotic decoy-state yield model (Lo, Ma and Chen 2005; Ma et al. 2005) at a 1 GHz clock, mean photon number 0.5, 0.2 dB/km fibre loss, and published detector classes, and divided into the stated total wall-plug power including detector cooling and classical post-processing. With InGaAs avalanche photodiodes (190 W total) the link delivers 7.68 Mbit/s at zero distance at 2.48e-05 J per secret bit and 57.37 kbit/s at 100 km at 3.31e-03 J per bit, with no key beyond about 134 km. With superconducting nanowire detectors (1690 W total, including a 1500 W closed-cycle cryocooler) it delivers 44.72 Mbit/s at zero distance at 3.78e-05 J per bit, 431.62 kbit/s at 100 km at 3.92e-03 J per bit, and 41.32 kbit/s at 150 km at 4.09e-02 J per bit, with no key beyond about 222 km. Energy per bit rises by a factor of 133.8 (avalanche) and 103.6 (nanowire) between 0 km and 100 km. A stated classical X25519 handshake costs 1.00e-03 J, or 3.91e-06 J per key bit, so the quantum link is 848 times (avalanche) and 1002 times (nanowire) as expensive per key bit at 100 km. Information-theoretic key security costs microjoules per bit at short range and millijoules per bit at 100 km, rising exponentially thereafter. Model calculations are labelled SIMULATED, NOT MEASURED. The deposit includes the paper PDF and text, the computation script (compute.py) that produces every number in the paper, and its outputs (results.json, results.txt).
Quantum error correction is priced in combinatorial currency — rate, distance, threshold — while the physical cost of performing correction goes uncounted. Every cycle of active (measurement-based) error correction erases information: syndrome extraction, majority voting, and ancilla re-initialization each reset redundancy, and Landauer's principle prices each erasure at no less than kT·ln2 per bit. This paper derives the resulting thermodynamic floor for correction overhead — (n−k)/k · kT·ln2 per logical qubit per round for an [[n,k]] code — verifies its scaling across code families (linear for repetition, quadratic for surface, constant for constant-rate), and applies the JPCUB joules-per-correct-answer metric to rank correction architectures. A two-level nested (tree-structured) code is compared with flat Hamming codes of equal rate on a clustered NAND-like channel: hierarchical decoding cuts the per-round erasure count by a factor of 1.6 to 3, while the naive construction pays a residual-error penalty — a partial null that leaves the floor claim intact and defines the protocol for an industrial flash-memory test. Biological systems that run robust quantum processes without correction (photosynthetic energy transfer, radical-pair magnetoreception) show the floor is an architecture choice: structure prepays in capital energy what correction pays in operating erasures. Autonomous and dissipative correction is identified as the boundary of the claim. A closing section documents the 2026 quantum-algorithms-and-QEC curriculum snapshot: the energy bill the paper prices is absent from the field's own teaching materials as of 2026-08-26.
Surface-code quantum error correction (QEC) is pivotal for fault-tolerant quantum computing, but its energy efficiency remains underexplored. This work derives a practical lower bound on the energy cost per logical-qubit operation, accounting for gate-fidelity decay's impact on code distance and physical-qubit overhead. By modeling the energy consumption of gate operations and error correction, we express the total cost as $ E = (E_g N_g + E_c N_c)/N_l $, where $ E_g $, $ E_c $, $ N_g $, $ N_c $, and $ N_l $ represent gate and physical-qubit energy costs, gate counts, physical-qubit counts, and logical qubit counts, respectively. Gate-fidelity decay increases the required code distance $ d $, scaling $ N_c $ and thus $ E $. We propose normalizing computational cost in joules per logical-qubit operation (JPCUB), isolating logical work from error-correction overhead. Quantitative predictions for $ E_g $, $ E_c $, and $ N_c $ remain [to verify], but the functional dependence on $ d $ and gate fidelities is established. This framework informs energy-efficient QEC design and benchmarks. ## 1. Introduction Surface codes enable fault-tolerant quantum computation by encoding logical qubits in a 2D lattice of physical qubits, correcting errors through syndrome measurements. However, the energy cost per logical operation—critical for scalability—remains poorly quantified. While theoretical foundations exist (Raussendorf & Terhal, 2006; Kitaev, 2003), prior work focuses on error-correction thresholds rather than energy efficiency. Here, we address this gap by deriving a lower bound on energy cost, considering how gate-fidelity decay necessitates larger code distances $ d $, increasing physical-qubit overhead $ N_c $. This analysis connects to the joules-per-compute (JPCUB) benchmark, advocating for energy normalization per logical qubit to disentangle computational work from error-correction costs. ## 2. Background Surface codes encode logical qubits using a 2D grid of
Recent work established that, for independent erasures, storing KKK logical qubits for SSS time steps with error at most ε\varepsilonε requires a spacetime cost of Ω(S(K+log⁡(S/ε)))Ω(S(K+\log(S/\varepsilon)))Ω(S(K+log(S/ε))). A matching construction was also given, but assumes ideal error correction. Here we show that the same scaling can be achieved with an explicit noisy error-correction circuit and efficient decoding, counting all state-preparation, gate, measurement, and wait locations, provided that the hardware supports long-range qubit connectivity and fast, reliable classical processing. Thus, reliability adds only a logarithmic overhead shared by all stored qubits. The construction works for sufficiently weak but otherwise general circuit noise, including faults during syndrome extraction and correlated faults, and for independent erasures it matches the best known lower bound up to constant factors. We attain this scaling using quantum Tanner codes, with each memory step consisting of one round of syndrome extraction followed by a fixed number of parallel classical decoder steps that reduce the residual error enough to keep later faults correctable, regardless of memory size, storage time, or target accuracy. Together, this gives exponentially small failure probability per step while keeping the cost of each memory step linear in the code size. The same analysis also extends to certain constant-depth logical Clifford operations.
In every syndrome cycle a data qubit spends most of its time doing nothing, waiting while its neighbouring measure qubits are read out and reset, and nothing is not free: idle decoherence accrues at a rate set by the ratio of waiting time to coherence time. This paper prices the wait. For a superconducting cycle of 1.0 microsecond in which entangling gates occupy a quarter of the schedule, against an energy-relaxation time of 100 microseconds, the idle error per data qubit per cycle is about 7.50e-3, which across the 441 data qubits of a distance-twenty-one patch amounts to an expectation of 3.3 idle error events per cycle, comparable to the entire gate-error contribution the cycle was designed around. The wait is a first-class error source, and the paper budgets the three ways to pay it down: faster cycles, longer coherence, and schedules that keep data qubits usefully busy, including dynamical decoupling, whose own costs are charged.