# PriorTilt: Decoding quantum LDPC codes with sparse perturbations of belief propagation priors

Quantum.world

> Open-source PriorTilt cuts logical error rates on the [[144,12,12]] Gross code by up to 4.3× over Relay-BP while decoding faster, giving qLDPC builders a usable decoder.

**arXiv**: 2611.13373

**Published**: 2026-11-09

**Authors**: Min Ye, Nicolas Delfosse

## AI summary

## Overview

**Research area:** Quantum error correction, specifically decoding of quantum low-density parity-check (LDPC) codes under circuit-level noise.

**Technical level:** Intermediate. The paper assumes familiarity with belief propagation, Tanner graphs, log-likelihood ratios, and the syndrome-decoding problem, but its central idea is conceptually simple.

**Scope:** The paper introduces PriorTilt, a belief-propagation variant that perturbs only a small, adaptively selected subset of prior log-likelihood ratios, and benchmarks it against Relay-BP and a beam search decoder on three quantum LDPC codes.

## What This Paper Is About

Belief propagation is fast and effective for classical LDPC codes, but in the quantum setting short cycles, stabilizer degeneracy, and symmetries cause it to oscillate or fail to find a syndrome-consistent solution. The authors ask whether BP can be made substantially more accurate by changing almost nothing about it: instead of altering the message-passing rules, they modify the prior log-likelihood ratios of a very small fraction of error nodes for short, fixed-duration intervals. The goal is a decoder that closes much of the accuracy gap to state-of-the-art quantum decoders while staying close enough to standard BP to remain cheap and hardware-friendly.

## Key Contributions

1. **A minimally invasive BP variant.** PriorTilt replaces the prior LLR with an effective prior that adds a per-node offset, leaving the min-sum message-passing equations, the Tanner graph, and the decoding loop otherwise unchanged. No normalization, damping, or message resets are used.
2. **A statistics-driven, sparse perturbation rule.** Error-prone nodes are identified by comparing a sum of recent posterior LLRs against a threshold scaled by the node's prior LLR; only those nodes can receive a nonzero offset, and the offset magnitude is drawn from clipped extrema of recent posterior LLRs and outgoing messages. Fewer than 0.5% of error nodes receive nonzero offsets per modified-BP round on average.
3. **A multi-phase parameter schedule.** The four offset parameters can vary across consecutive groups of rounds. The paper formalizes an m-phase mode for any 1 ≤ m ≤ R, reports that going from one to two phases reduced logical error rate and a third phase gave a further reduction, that a fourth phase gave no clear additional improvement, and uses a three-phase mode (14 tunable quantities) throughout.
4. **Circuit-level benchmarking against two leading decoders.** PriorTilt is compared with Relay-BP and a beam search decoder on BB144 [[144,12,12]], BB70 [[70,6,9]], and GB102 [[102,22,9]], including decoding-time measurements and an XYZ-decoding comparison on BB144. Source code is released for reproducibility.

## Main Findings

- **Lower logical error rate at every tested point.** For standard XZ decoding, PriorTilt achieves the lowest LER of the three decoders at every tested physical error rate on all three codes.
- **Size of the accuracy gain on BB144.** Reductions range from 1.7× to 4.3× relative to Relay-BP and from 1.9× to 3.1× relative to beam search. At p = 10⁻³ specifically, the reductions reach 4.3× versus Relay-BP and 3.1× versus beam search.
- **Gains on the other two codes.** For BB70, reductions are 1.6×–3.5× versus Relay-BP and 1.9×–2.7× versus beam search. For GB102 the gains are more modest but consistent: 1.2×–1.6× versus Relay-BP and 1.4×–2.0× versus beam search.
- **Sparsity of the intervention.** The mean fraction of error nodes satisfying the error-prone condition, which upper-bounds the fraction receiving nonzero offsets, is 0.19% (BB144, p = 0.001), 0.33% (BB144, p = 0.002), 0.16% (BB70, p = 0.001), 0.31% (BB70, p = 0.002), 0.33% (GB102, p = 0.001), and 0.26% (GB102, p = 0.002). All are below 0.5%.
- **Comparison against standard BP.** On BB144 at p = 0.001 with the same maximum budget of 18060 iterations, standard min-sum BP yields an LER of 1.27 × 10⁻⁵ while PriorTilt achieves 2.08 × 10⁻⁸, a reduction of approximately 610×.
- **Mean decoding times.** For Z-memory experiments at p = 0.001 over the same 100000 shots per code, PriorTilt's mean times are 0.907 ms (BB144), 0.281 ms (BB70), and 0.803 ms (GB102), versus 1.038 ms, 0.416 ms, and 1.224 ms for Relay-BP, and 2.377 ms, 0.829 ms, and 2.219 ms for beam search. PriorTilt therefore has lower mean time than Relay-BP on all three codes and mean speedups of roughly 2.6×–3.0× over beam search.
- **Tail latency.** PriorTilt's 99th-percentile times are lower than beam search on all three codes by roughly 1.7×–2.3×, and lower than Relay-BP on BB70 and GB102, but higher than Relay-BP on BB144 (4.52 ms versus 3.33 ms).
- **XYZ decoding on BB144 at p = 0.003.** PriorTilt reaches an LER per cycle of 1.04 × 10⁻⁵ versus 1.10 × 10⁻⁵ for Relay-BP-5, essentially the same accuracy, while using 43.5% less decoding time on average, a 1.77× speedup (601 ms versus 1064 ms). Both decoders do far better under XYZ than under XZ decoding, where the LER per cycle is 5.74 × 10⁻⁵ for PriorTilt and 1.09 × 10⁻⁴ for Relay-BP-1.
- **Comparison with BP-SF via reported figures.** The paper notes that, per Fig. 7 of the BP-SF reference, BP-SF reduces LER by roughly 40–50× relative to standard BP at nearly equal maximum aggregate iteration budgets of 10100 and 10000, whereas PriorTilt reduces LER by roughly 610× relative to standard BP at equal maximum budgets of 18060. The authors state that BP-SF's independent trials can run concurrently, while PriorTilt's successive rounds depend on the preceding round's statistics, though updates within an iteration remain parallelizable.
- **Consistent parameter pattern.** Across optimized configurations, the final phase has the largest c₂ and smaller π and c₁ than the initial phase, meaning later rounds more often select the nonpositive-offset branch (favoring the error hypothesis) with a larger magnitude, while applying nonnegative offsets less often and less strongly.

## Methodology in Plain English

The authors start from a standard min-sum BP decoder that runs for up to 60 ordinary iterations. If BP has not converged, decoding continues in rounds of six iterations each, up to a maximum of 3000 rounds, for a total budget of 18060 iterations.

At the boundary of each round, the decoder looks back at the previous six iterations for every error node and computes three simple statistics: the sum of that node's posterior LLRs, and the most negative and most positive values seen among its posterior LLRs and outgoing messages (clipped at zero). A node counts as "error-prone" if its posterior sum falls below a threshold equal to θ times its prior LLR — that is, if BP either clearly leans toward the error hypothesis or does not clearly lean either way.

Only error-prone nodes get a perturbation. For each of them, a coin flip weighted by π chooses between a nonnegative offset scaled by the node's most negative recent value and a nonpositive offset scaled by its most positive recent value, with coefficients c₁ and c₂. That offset is added to the node's prior LLR and held fixed for all six iterations of the round, then recomputed, replacing rather than accumulating on the previous value. Messages are never reset, so the perturbation acts as a short nudge on a continuously running BP trajectory.

The four parameters (θ, π, c₁, c₂) are allowed to change across three groups of rounds, giving 14 tunable quantities including the two internal boundaries. These are optimized separately for each code and each physical error rate using coordinate descent on a training pool containing only shots that fail to converge in the initial 60 ordinary iterations — since converged shots are unaffected by the parameters. Separate pools are built for X- and Z-memory circuits, and their combined failure count is minimized.

Benchmarks use d syndrome extraction cycles for distance d: 12 cycles for BB144, nine for BB70 and GB102. BB144 uses the circuits and uniform circuit-level noise model from an earlier reference; BB70 and GB102 use circuits with a moving-qubit circuit-level noise model, ignoring qubit loss and leakage. Monte Carlo samples are generated with Stim. The reported LER is the sum of X-memory and Z-memory LERs at the same physical error rate. PriorTilt runs 18060 iterations maximum versus 18080 for Relay-BP-1 (an 80-iteration initial leg plus 300 legs of 60), and the beam search decoder uses the beam32_340iters configuration. Timings are single-thread on one core of a 2023 Apple M3 Pro laptop over 100000 shots per code.

## Why This Matters

**Impact on research.** The paper shows that a decoder can gain more than two orders of magnitude in accuracy over standard BP without abandoning the message-passing structure — no Gaussian elimination, no OSD post-processing, no ordered statistics, no restarts. That reframes the design space for quantum LDPC decoding: the choice is not simply "BP plus an expensive classical post-processor" versus "plain BP." It also provides a direct, like-for-like comparison against Relay-BP at nearly matched iteration budgets, and a concrete sparsity statistic (under 0.5% of nodes) that quantifies how little of the prior actually needs to change.

**Real-world applications.**

- Fault-tolerant quantum computing with quantum LDPC codes, where decoder latency and accuracy both directly affect the physical-qubit overhead needed to reach a target logical error rate.
- Real-time decoding in control electronics for trapped-ion and superconducting architectures, where the mean and 99th-percentile decoding times reported here are the relevant engineering constraints.
- Memory experiments and hardware demonstrations that need a decoder accurate enough to resolve logical error rates below what standard BP can reach.
- Classical-inspired decoder hardware (FPGA/ASIC) benefiting from a decoder that preserves the standard BP message-passing schedule.

**Industry relevance.** The authors are at IonQ Inc., and the paper is framed around practical decoder latency as well as accuracy. The emphasis on preserving standard BP message passing, keeping per-round perturbations sparse, and reducing both mean and tail decoding times points toward deployable hardware. The reported 1.77× speedup over Relay-BP-5 at essentially identical LER under XYZ decoding, and the 2.6×–3.0× mean speedup over beam search, are the kinds of numbers that matter when decoding must keep pace with syndrome extraction.

## Future Directions

- **Hardware realization.** The authors suggest that PriorTilt's proximity to standard BP may facilitate efficient hardware implementations, but no FPGA or ASIC implementation is reported here; realizing one is the obvious next step.
- **Exploiting parallelism.** Successive PriorTilt rounds depend on statistics from the preceding round, whereas BP-SF's independent trials can run concurrently (as demonstrated in a GPU implementation). Closing that parallelism gap, or restructuring rounds to be less sequentially dependent, is an open engineering question.
- **Richer parameter schedules and other codes.** The paper reports that a fourth phase gave no clear additional improvement, and that intermediate-phase settings and θ behave in a code-dependent way. Whether the m-phase construction helps on other code families, or on noise models beyond those tested, is untested — the simulations ignore qubit loss and leakage, and beam search has been applied to simulations beyond Pauli stochastic noise.
- **Reducing tuning cost.** The 14 parameters are optimized separately for each code and each physical error rate. Whether parameters transfer across error rates, across codes, or can be chosen by a cheaper procedure is not resolved, though the paper notes one Relay-BP parameter set per code sufficed for all rates while PriorTilt's did not.

## Target Audience

Researchers and engineers working on quantum error correction and decoder design for fault-tolerant quantum computing, particularly those building or benchmarking BP-based decoders for quantum LDPC codes. It will also interest readers tracking the practical comparison between BP variants (Relay-BP, BP-SF, beam search) and post-processing decoders such as BP-OSD, and hardware engineers assessing whether a decoder's structure is compatible with real-time implementation. Readers without background in LDPC decoding or the syndrome-decoding formalism will need to consult the cited background literature first.

## Authors’ abstract

Although belief propagation (BP) is a widely adopted decoder for classical low-density parity-check (LDPC) codes, its accuracy is limited in the quantum setting. We introduce \PriorTilt{}, a variant of BP obtained by selectively perturbing a small fraction of its prior log-likelihood ratios using offsets determined by recent posterior and outgoing-message statistics, while preserving the same message-passing rules. We compare \PriorTilt{} with Relay-BP and beam search…

[The full abstract on arXiv](https://arxiv.org/abs/2611.13373)

[Read the original paper](https://arxiv.org/abs/2611.13373)

---

Canonical: https://quantum.world/paper/priortilt-decoding-quantum-ldpc-codes-with-sparse-perturbations-of-2611.13373
