⚛️ quantum physics

Fair Decoder Baselines and Rigorous Finite-Size Scaling for Bivariate Bicycle Codes on the Quantum Erasure Channel

This paper establishes a fair benchmark for bivariate bicycle (BB) codes on the quantum erasure channel by correcting decoder-baseline biases and applying rigorous finite-size scaling to demonstrate that BB codes achieve an asymptotic threshold of approximately 0.488 with significantly lower normalized overhead compared to surface codes.

Original authors: Tushar Pandey

Published 2026-03-20

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Tushar Pandey

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to send a secret message across a stormy sea. The "storm" is noise that can destroy your message, and the "boat" is your quantum computer. To keep your message safe, you don't just send one copy; you wrap it in a complex, magical net called a Quantum Error Correction Code.

This paper is about testing two different types of nets to see which one is better at saving your message when the storm hits.

Here is the breakdown of the research, explained simply:

1. The Storm: The "Erasure Channel"

In the real world, errors are sneaky. You often don't know where the mistake happened. But in this specific experiment, the researchers used a "fair" storm called the Erasure Channel.

The Analogy: Imagine sending a letter, but instead of the ink smudging randomly, the post office just tells you, "Hey, this specific page is missing." You know exactly where the hole is, you just need to figure out what was written on it.
Why it matters: Because you know where the holes are, you can use a smarter strategy to fill them in.

2. The Two Nets: Surface Codes vs. Bicycle Codes

The researchers compared two famous types of nets:

Surface Codes (The Local Net): Think of this like a standard fishing net where every knot is only tied to its immediate neighbors. It's very reliable and easy to build, but it's heavy. To protect a lot of information, you need a huge net.
Bivariate Bicycle (BB) Codes (The Long-Reach Net): This is a newer, more complex design. Imagine a net where some knots are tied to knots far away across the ocean. It's more efficient, but the instructions to tie it are harder to follow.

3. The Big Mistake in Previous Studies

The paper starts by pointing out a major flaw in how people used to compare these nets.

The Problem: Previous studies tried to fix the "Local Net" (Surface Code) using a very dumb strategy. They acted like they didn't know where the holes were in the letter, even though the post office had told them!
The Result: It's like trying to solve a puzzle while blindfolded. The "Local Net" performed terribly, looking like it was barely better than random guessing.
The Fix: The authors said, "That's unfair!" They decided to test both nets using the smart strategy (knowing exactly where the holes are). This is the "Fair Baseline."

4. The Results: Who Wins?

Once they compared the nets fairly (both knowing where the holes were), here is what they found:

The "Threshold" (How strong the storm can be): Both nets are incredibly strong. They can survive storms up to about 49% of the qubits being erased. The Bicycle net is slightly stronger, but the difference is tiny.
The "Overhead" (How much material you need): This is the real winner.
- To protect a small amount of data, the Surface Net needs a massive amount of material (qubits).
- The Bicycle Net needs 12 times less material to do the exact same job.
- Analogy: If the Surface Net is a giant, heavy wool blanket, the Bicycle Net is a high-tech, lightweight thermal sheet that keeps you just as warm but weighs a fraction of the price.

5. The "Magic" Math (Finite-Size Scaling)

The researchers didn't just test small nets; they tested nets of five different sizes, from small to huge.

The Analogy: Imagine testing a bridge with 10 cars, then 100, then 1,000. You want to know if it will hold up when 1 million cars drive over it.
The Finding: By using a special mathematical trick called "Finite-Size Scaling," they predicted that as the Bicycle nets get infinitely large, they will get even closer to the theoretical limit of perfection (97.6% of the best possible performance).

6. Why This Matters for the Future

The main takeaway isn't that one net is "stronger" at stopping storms; they are both very strong. The breakthrough is efficiency.

Current Reality: Building quantum computers is hard and expensive. Every extra "knot" (qubit) you need to build a net costs money and engineering time.
The Verdict: The Bicycle (BB) Codes are the clear winner for the future because they achieve the same safety with 12 times fewer resources.

Summary in One Sentence

The authors fixed a "cheating" comparison method used in the past and proved that a new type of quantum code (Bicycle Codes) is just as tough as the old standard but requires 12 times less hardware to build, making it a much more practical choice for building real quantum computers.

1. Problem Statement

The paper addresses two critical methodological flaws in the current literature comparing Quantum Low-Density Parity-Check (QLDPC) codes, specifically Bivariate Bicycle (BB) codes, against standard Surface (Toric) codes on the Quantum Erasure Channel:

Decoder-Baseline Unfairness: Many prior studies compare BB codes (decoded with advanced algorithms like BP-OSD) against Surface codes decoded with Uninformed Minimum-Weight Perfect Matching (MWPM). On an erasure channel, uninformed MWPM ignores the known locations of erased qubits. The author demonstrates that in this regime, uninformed MWPM performs no better than random guessing (Word Error Rate $\approx 0.75$ ), rendering comparisons against it invalid for assessing code performance.
Confusion of Pseudo-Thresholds with Asymptotic Thresholds: Previous works often cite "pseudo-thresholds" (the error rate where a finite-size code reaches a specific Word Error Rate, e.g., 0.10) as the true asymptotic threshold. This paper argues that without rigorous Finite-Size Scaling (FSS), these values significantly underestimate the true capacity of the code family.

2. Methodology

The study employs a rigorous Monte Carlo simulation framework with the following components:

Code Family: The authors focus on a specific family of BB codes defined by generator polynomials $A = x^3 + y + y^2$ and $B = y^3 + x + x^2$ over $\mathbb{Z}_L \times \mathbb{Z}_M$ . They simulate five code sizes ranging from $N=144$ to $N=1296$ physical qubits.
Channel Model: The Quantum Erasure Channel, where each qubit is erased with probability $p$ . Erased qubits are assigned independent Pauli $X$ and $Z$ errors with probability 0.5.
Decoders:
- BB Codes: Decoded using BP-OSD (Belief Propagation with Ordered Statistics Decoding, order 10) with erasure-informed priors.
- Surface Code Baselines: Two baselines are run for fairness:
  1. Uninformed MWPM: Converts erasure rate to depolarizing rate ( $q=p/2$ ) and runs standard MWPM.
  2. Erasure-Aware MWPM: Assigns weight 0 to erased qubits and high weight to non-erased ones, forcing corrections through known erasure locations. This is the fair baseline.
Statistical Rigor:
- Sample Size: 200,000 shots per data point.
- Uncertainty Quantification: 5,000-iteration parametric bootstrap to generate 95% confidence intervals (typically $\lesssim 0.001$ ).
- Reproducibility: All runs use recorded seeds and specific package versions (numpy, scipy, bposd, pymatching).
Finite-Size Scaling (FSS): The authors fit the Word Error Rate (WER) data to the scaling ansatz $WER(p, N) \approx f((p - p^*_\infty)N^{1/\nu})$ to extrapolate the asymptotic threshold ( $p^*_\infty$ ) and the critical exponent ( $\nu$ ).

3. Key Contributions

Demonstration of Baseline Failure: The paper proves that using uninformed MWPM on the erasure channel yields performance equivalent to random guessing ( $WER \approx 0.75$ ). Consequently, any study comparing QLDPC codes to this baseline is effectively comparing decoders, not code families.
Rigorous FSS Extrapolation: Instead of relying on finite-size crossing points, the authors use FSS to determine the true asymptotic threshold, providing a statistically robust estimate with error bars.
Fair Comparison Framework: By introducing the "Erasure-Aware MWPM" baseline, the paper establishes a fair metric for comparing non-local QLDPC codes against local Surface codes.
Reproducibility Pipeline: The work provides a fully reproducible simulation pipeline, addressing a common lack of transparency in quantum simulation literature.

4. Key Results

A. Asymptotic Threshold

BB Codes: The extrapolated asymptotic threshold for the BB family under BP-OSD is $p^*_\infty \approx 0.488$ (with a 95% CI of $\pm 0.001$ ).
Efficiency: This value is 97.6% of the zero-rate Shannon limit ( $p=0.5$ ) for the erasure channel, achieved without Maximum Likelihood (ML) decoding.
Critical Exponent: The scaling analysis yields a critical exponent $\nu \approx 1.18$ , consistent with a second-order phase transition, though the authors note the universality class may differ from the standard Random-Bond Ising Model due to the non-local nature of BB codes and the specific decoder used.

B. Pseudo-Thresholds

The pseudo-thresholds (WER = 0.10 crossing) increase with code size:

$N=144$ : $p^* \approx 0.370$
$N=1296$ : $p^* \approx 0.471$
The gap between the largest finite size and the asymptotic limit is small ( $\sim 0.017$ ), but significant enough to warrant FSS.

C. Comparison with Surface Codes

Threshold: At comparable sizes, BB codes show a modest threshold advantage over erasure-aware Surface codes (e.g., $p^* \approx 0.471$ for BB vs. $0.460$ for Surface at $N \approx 2592$ ).
Overhead (The Major Finding): The primary advantage of BB codes is qubit efficiency.
- The BB code $[[1296, 12]]$ achieves a threshold of 0.471 with 108 physical qubits per logical qubit ( $N/K = 108$ ).
- The Surface code achieving a similar threshold requires 1296 physical qubits per logical qubit (due to the constraint of $K=2$ for local planar codes).
- Result: BB codes offer a 12 $\times$ reduction in normalized overhead compared to Surface codes at comparable thresholds.

D. Rate-Threshold Trade-off

Larger BB codes trade encoding rate for higher thresholds. The $[[144, 12]]$ code has a high rate ( $K/N \approx 0.083$ ) but lower threshold, while the $[[1296, 12]]$ code has a lower rate ( $K/N \approx 0.009$ ) but approaches the asymptotic threshold.

5. Significance and Conclusion

This paper fundamentally shifts the benchmarking standard for QLDPC codes on erasure channels. It argues that qubit efficiency (overhead) is the more practical metric for BB codes than raw threshold gains, as they offer a massive reduction in physical qubit requirements compared to Surface codes.

The study concludes that:

Uninformed MWPM is an invalid baseline for erasure channels; future comparisons must use erasure-aware decoders.
BB codes are highly efficient: They approach the theoretical capacity limit with near-optimal decoders (BP-OSD) and require significantly fewer physical qubits per logical qubit than Surface codes.
Methodological Rigor: The combination of massive shot counts, bootstrap confidence intervals, and FSS provides a new standard for reliability in quantum error correction simulations.

The work suggests that while BB codes are non-local (posing architectural challenges), their code-capacity efficiency makes them a superior candidate for fault-tolerant quantum computation in settings where qubit count is the primary constraint.