Research
Closing the loop on memory: accurate, shallow and transferable decoding of quantum LDPC codes
Releases a qLDPC decoder that cuts gross-code logical failure from 2.51% to 0.73% and ships code, making circuit-level fault tolerance more credible.
- arXiv
- 2611.21086
- Published
- 2026-11-16
- Authors
- Zhou You, Ying Li, Xiao Yuan, Yuan Yao
AI summary
Overview
Research area: Quantum error correction — specifically decoding algorithms for quantum low-density parity-check (qLDPC) codes under circuit-level noise, combining belief propagation with online feedback control theory.
Technical level: Advanced. The paper assumes familiarity with belief propagation / min-sum message passing, Tanner graphs, detector error models, syndrome-extraction rounds, and basic linear-dynamical-systems language (eigenvalues, unit circle, loop gain).
Scope (one sentence): The paper introduces TEMPO-BP, a training-free decoder that continuously adjusts each variable's memory strength during decoding, and reports its accuracy, iteration depth, runtime, and cross-code transferability against BP, memory-BP variants, Relay-BP, and BP+OSD+CS-10 on the gross code and five other code families.
What This Paper Is About
qLDPC codes promise fault-tolerant quantum computing with less qubit overhead, but only if their syndrome data can be decoded reliably and fast enough. Standard belief propagation often fails to converge under circuit-level noise, and existing "memory-based" fixes hold each variable's memory strength fixed during a decoding run, which forces a trade-off between speed and convergence and requires many sequential retries to escape trapped states. The authors replace that fixed knob with a local feedback controller that adjusts each variable's memory strength on the fly, aiming for one decoder configuration that is accurate, shallow in iteration depth, and portable across code families without retuning.
Key Contributions
-
TEMPO-BP (TEmpered Memory via Predictive Online control): a decoder that exploits an exact local identity — once the next incoming check messages are known, a variable's next belief is an affine function of its own memory strength — to steer that belief toward a small opposite-sign target, projecting the result onto an admissible interval and taking a bounded step toward it. The required check messages are computed one step early and reused, so no extra check-node updates are added and message passing stays local and parallel.
-
A gated, state-dependent controller: updates act as a "brake" (lowering memory where a belief is being reinforced) or a "hold" (raising it where incoming messages would flip the sign), gated by the local fraction of unsatisfied syndrome checks and by variable confidence. The control interval is K = 4 iterations on the gross code and transfer codes, K = 8 on BB-72.
-
A loop-gain theory of memory: memory strength is interpreted as the gain of a feedback loop. On an exactly solvable zero-prior family satisfying a two-dimensional linear map with characteristic equation λ² − (b + γ)λ − γ = 0 (b = c − 1), plain BP sits at unit gain when b = 1, positive memory pushes the dominant root outside the unit circle (λ₊ ≈ 1 + 2γ), and −1 < γ < 0 places both roots inside. A second, temporal amplification channel arises from repeated syndrome-extraction rounds, with asymptotic growth factor m^{1/(T+1)} for m ≥ 2 every T + 1 iterations.
-
Parallel ensembles as a latency-control mechanism: independent TEMPO-BP replicas run concurrently with per-member iteration caps of ⌊18,080/S⌋ (S = 15 on the gross code, 1,205 iterations each; S = 10 on BB-72, 1,808 each), returning the valid correction most probable under the priors.
Main Findings
-
Single-run accuracy on the gross code: At a 0.4% physical error rate on the [[144,12,12]] bivariate bicycle (gross) code (100,000 shots), one TEMPO-BP run fails on 0.73% of shots (730 failures; 6.10 × 10⁻⁴ per round). Comparators were 2.51% for BP+OSD+CS-10 (2.12 × 10⁻³ per round), 4.82% for warm-started disordered memory DMem-warm (4.11 × 10⁻³), 7.49% for Mem-BP (6.46 × 10⁻³), 9.84% for BP (8.59 × 10⁻³), and 15.3% for un-warm-started DMem-BP. That is 3.4 times fewer failures than BP+OSD+CS-10 (730 versus 2,510).
-
Consistent lead across noise rates: On the gross code from 0.2% to 0.6%, TEMPO-BP reduces failures by factors of 3.5 to 46 relative to BP, 2.8 to 9.5 relative to DMem-warm, and 2.3 to 3.4 relative to BP+OSD+CS-10. On BB-72 at 0.4% (20,000 shots) it fails on 4.65% (929 failures; 7.90 × 10⁻³ per round) versus 5.51% for Mem-BP, 5.75% for DMem-warm, and 6.29% for BP+OSD+CS-10. Paired tests remain significant after Holm correction (P < 10⁻³⁰⁰ on the gross code; P ≤ 8.8 × 10⁻¹³ on BB-72).
-
Failure modes differ by code distance: The gross code's fixed-memory baselines fail mostly by not converging (9.82% unconverged for BP and 4.28% for DMem-warm at 0.4%); TEMPO-BP cuts non-convergence to 231 shots (0.23%) alongside 499 valid wrong-class corrections (0.50%). On distance-6 BB-72 convergence is rarely limiting — TEMPO-BP leaves only 38 of 20,000 shots unconverged (0.19%), and wrong-logical-class corrections dominate.
-
Fixed memory is fundamentally limited: In a 10,000-shot uniform scan at 0.4%, plain BP fails on 9.7%; positive memory speeds converged shots (median iterations 46 for BP to 14 at γ = 0.4) but failures dip only to 6.7% at γ = 0.05 before rising to 25% at γ = 0.5. Negative memory reverses the trade-off (45% failures at γ = −0.1; no convergence at γ = −0.3). Of 42 scanned random interval fields, the best confirmed rate was 6.6%, versus 0.71% for TEMPO-BP on the same shots — a 9.4-fold reduction (95% CI 7.6 to 12.0). The best fixed field fails about nine times more often than TEMPO-BP.
-
Lower iteration depth and runtime: Under a 10,000-iteration cap at 0.4%, TEMPO-BP averages 92 iterations with a 99th percentile of 1,278, against 533 mean iterations and a saturated 10,000-iteration tail for DMem-warm. On isolated single-core timings (5,000 shots), TEMPO-BP averages 19.2 ms per shot versus 62.1 ms for BP+OSD+CS-10, 86.2 ms for DMem-warm, 100.5 ms for Mem-BP, 150.4 ms for TEMPO-static, 175.4 ms for BP, and 269.5 ms for DMem-BP.
-
Ensembles match or beat Relay-BP-5 at all ten operating points: On 48,000 common shots at 0.4%, the ensemble overtakes Relay-BP-5 from S = 8 on the gross code (222 versus 227 failures) and from S = 2 on BB-72. At full width it achieves fewer failures at eight points and identical counts at the two lowest gross-code rates (13 in 500,000 at 0.2%; 36 in 100,000 at 0.3%). At 0.4%, rates fall from 0.47% to 0.35% on the gross code and from 4.39% to 3.72% on BB-72; at 0.5% and 0.6% on the gross code, failures drop 39% and 36%. Per-member caps are 15 times below Relay-BP-5's 18,080-iteration serial limit (1,205 versus 18,080).
-
Bounded decoding depth: Mean ensemble depth is below Relay-BP-5's mean iteration count at all ten points by factors of 1.5 to 10, and 99th-percentile depth is lower at nine of ten points (1,097 against 828 on BB-72 at 0.4%). At 0.6% on the gross code, Relay-BP-5 saturates its 18,080-iteration limit while ensemble depth never exceeds 1,205. Summed iterations across members are 0.87× Relay-BP-5 at gross-code 0.6% and 1.04 to 2.68 times as many on average across the remaining nine points.
-
Transfer without retuning: Frozen gross-code parameters applied to LP20 [[1122,148,≤20]], Kasai-96 [[1152,580,≤12]], Coset-128 [[128,16,12]], and Mitten-540 [[540,108,18]] gave fewer failures than Relay-BP-5 (in its published gross-code configuration) at all twelve tested points. At center points, single-run TEMPO-BP cut failures by 8.5× on LP20 (17 versus 145 in 2,000 shots), 3.4× on Kasai-96 (36 versus 123), 1.5× on Coset-128 (330 versus 485 in 10,000 shots), and 7.9× on Mitten-540 (8 versus 63 in 2,000). Speedups on one core were 118-fold on LP20 (0.39 s versus 46 s at 0.175%), 34-fold on Mitten-540, 12-fold on Kasai-96, and 4.4-fold on Coset-128.
-
Theory matches observed dynamics: The constructed b = 1 invariant state was verified in integer arithmetic on the full gross-code detector error model (936 detectors, 8,784 fault mechanisms, 30,672 edges). Across roughly 200 unconverged trajectories per strength over nine strengths in [0.0125, 0.4], the growing fraction rose from 34% to 96%, and over half of growing trajectories matched ln λ₊(γ) to four significant digits. LP20 and Mitten-540 mostly grow faster instead (1.22 versus 1.18 at γ = 0.1), consistent with the temporal channel.
-
Ablations support state-dependent control: On 10,000 gross-code diagnostic shots at 0.4%, TEMPO-BP fails 75 times; the always-brake ablation fails 2,552 (worse than 994 for plain BP, 1,227 for controller off, and 1,738 for blind decay), forbidding negative memory leaves 349, and dropping unsatisfied-check gating leaves 110. Although 99.0% of the 2 × 10⁹ controller actions request strengths below γ_min, the 0.8% above γ_max and 0.2% inside Γ supply 13% of total memory-strength change.
-
Hard-syndrome reproducibility: Decoding 400 hard syndromes (more than 100 initial iterations) with 128 random initial memory fields each gave convergence scales reproducible across disjoint seed halves (rank correlation 0.986), spanning a factor of about 630 between the 1st and 99th percentiles. Replicas cut failures on 100,000 gross-code shots from 1,360 (one) to 927 (two) and 419 (fifteen); parameter diversity beat identical-parameter replicas (376 versus 419; exact McNemar P = 0.009).
Methodology in Plain English
The decoder works on the detector error model that a noisy quantum circuit produces: a sparse binary matrix whose rows are parity checks (detectors) and whose columns are independent fault mechanisms with known prior probabilities. Standard min-sum belief propagation passes messages between checks and variables until the belief signs reproduce the measured syndrome.
Memory BP changes the update rule so that each variable mixes its prior with its own previous belief, weighted by a memory strength γ. That memory helps when it is right and hurts when it is wrong — and one global value cannot be right for every variable and every syndrome.
The key observation is algebraic: once the next round of check messages is known, a variable's next belief is exactly a straight line in its own memory strength — the intercept is what memoryless BP would produce, the slope is how far the variable has drifted from its prior. So the decoder can ask, per variable and per moment, "which memory strength would land my next belief on a small target of the opposite sign?" It solves for that value, clips it into a permitted range that spans negative and positive strengths, and moves partway toward it, taking bigger steps for confident variables and for variables touching unsatisfied checks. The check messages needed for this prediction are computed one iteration early and reused, so the trick costs no extra work and preserves the local, parallel structure that hardware likes.
For latency, the authors exploit the fact that independent replicas fail on different hard syndromes. They run many copies side by side with a per-copy iteration cap set so the total budget stays within the serial limit of the competing Relay-BP decoder, then pick the most probable valid correction under the priors.
To explain why this works, they isolate exactly solvable states where min-sum collapses to a small linear map, derive growth rates as a function of memory strength, and separately prove that repeated syndrome-extraction rounds create a second, temporal amplification channel along closed chains of repeated faults. They then check the predicted growth rates against tracked unconverged decoding trajectories.
Why This Matters
Impact on research. The paper reframes memory in iterative decoding as a per-variable control parameter rather than a hyperparameter, and gives an exactly solvable analysis of when memory helps and when it amplifies. It also offers a concrete competitor to both ordered-statistics post-processing (whose Gaussian elimination scales cubically on every unconverged shot)
Authors’ abstract
Quantum low-density parity-check (qLDPC) codes have attracted interest for fault-tolerant quantum computing because their high-rate constructions offer a route to reduced qubit overhead. Realizing this potential requires reliable decoding. Belief propagation (BP) forms the basis of many qLDPC decoders, but can fail to converge under circuit-level noise. Memory-based variants can improve convergence, yet selecting memory strengths that work reliably across different syndromes…
The full abstract on arXiv