⚡ electrical engineering

Auxiliary Nodes for BP Decoding of Quantum LDPC Codes

This paper proposes a general framework for enhancing belief propagation decoding of CSS quantum LDPC codes by introducing auxiliary variable and check nodes into the decoding graph, a method that unifies existing techniques like 4-cycle removal and subcode ensemble decoding while demonstrating significant reductions in logical error rates under circuit-level noise.

Original authors: Daniel Tandler, Paul Bezner, Stephan ten Brink

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Daniel Tandler, Paul Bezner, Stephan ten Brink

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 solve a giant, three-dimensional maze where the walls are made of invisible quantum blocks. Your goal is to find a hidden path (the correct error correction) without bumping into the walls. In the world of quantum computers, this maze is called a Quantum Low-Density Parity-Check (QLDPC) code.

To solve this maze, scientists usually use a strategy called Belief Propagation (BP). Think of BP as a swarm of tiny, curious ants marching through the maze. They pass notes to each other saying, "I think the path is here!" or "No, it's over there!" Over time, they hope to agree on the right path.

But here's the problem: sometimes the maze has short loops (like a 4-cycle, which is a tiny square loop). When the ants hit these loops, they get confused. They start passing the same wrong note back and forth, reinforcing a bad idea until they all get stuck in a "trapping set." It's like a group of friends all agreeing on a wrong direction because they keep talking to each other in a circle.

The New Idea: Adding "Helper" Nodes

The authors of this paper, Daniel Tandler and his team, propose a clever way to fix the maze without changing the actual quantum hardware. They suggest adding Auxiliary Nodes—extra "helper" spots in the maze that don't exist in the original design but are added just for the ants to use while they solve the puzzle.

They introduce two types of helpers:

Auxiliary Check Nodes (ACNs): These are like new signposts added to the maze.
Auxiliary Variable Nodes (AVNs): These are like extra empty rooms added to the map.

The magic is that these helpers are temporary. Once the ants solve the puzzle using the helpers, the team can mathematically "erase" the helpers and translate the solution back to the original maze. It's like giving a student a cheat sheet to study a difficult math problem, then taking the cheat sheet away before the final exam to see if they really learned it.

Two Ways to Use the Helpers

The paper shows that these helpers can be used in two distinct ways, which the authors prove are actually two sides of the same coin:

1. Breaking the Loops (4-Cycle Removal)
Sometimes, the maze has those tiny, confusing square loops. The team uses the helpers to "break" these loops. They add a helper node that forces the ants to take a slightly different route, effectively cutting the loop open.

The Catch: The paper finds that this doesn't always make the ants faster. In their simulations (computer tests), the success of this method depends heavily on how many times the ants are allowed to pass notes (the number of iterations) and how loudly they shout their messages (a scaling factor called $\alpha$ ).
The Result: For some settings, breaking the loops helps a lot. But if the ants don't get enough time to think (low iteration count), adding these helpers can actually make things worse because the maze gets bigger and more confusing before it gets simpler.

2. Splitting the Confusion (Subcode Ensemble)
Quantum mazes have a unique problem called degeneracy. This means there might be two or more different paths that look exactly the same to the ants (they produce the same "syndrome" or clue). The ants get stuck because they can't tell which path is the real one.

The Fix: The team uses the helpers to "split" the maze. They create two versions of the puzzle: one where they assume the extra helper is "on" and one where it's "off." This forces the ants to pick a specific path in each version, breaking the symmetry.
The Ensemble: Instead of just running one swarm of ants, they run a whole team (an ensemble) of swarms, each trying a different combination of helper settings. If one swarm finds a valid path, they pick the best one.

What the Numbers Say

The team tested these ideas on a specific quantum code called the [[72, 12, 6]] bivariate bicycle (BB) code. They simulated errors at a rate of $p = 0.001$ and ran 6 measurement rounds ( $r=6$ ).

The Loop Breaker: When they removed the 4-cycles, the logical error rate (how often the maze solver fails) dropped, but only if they let the ants run for enough steps (iterations). If they stopped too early, the extra helpers just made the graph bigger without helping.
The Team Approach: The most exciting result came from the ensemble decoder. By using the helpers generated during the loop-breaking process as "splitting" tools, they created a team of decoders.
- For the smaller code, an adaptive ensemble (where the team decides which helpers to use based on the current mess) with 24 members performed almost as well as a much more complex, slower method called BP+OSD-0.
- For a larger code ([[90, 8, 10]]), even a team of 128 members didn't quite catch up to the best possible performance, suggesting that for bigger mazes, they might need even smarter tricks (like windowed decoding) to help the information spread faster.

What They Don't Claim

It's important to note what this paper doesn't say:

They do not claim this is a magic bullet that solves all quantum errors.
They do not say that removing loops is always better; in fact, they show it can be worse if the decoder doesn't run long enough.
They do not claim that the "adaptive" method is perfect; they suggest that their current way of picking helpers might not be the best possible way, and a smarter selection strategy could improve results further.

The Bottom Line

The paper proposes a general framework where you can temporarily add "helper" nodes to a quantum decoding graph to fix confusing loops and break symmetries. In simulations, this approach allows a team of simple decoders to work together and significantly reduce errors compared to a single decoder. However, the success depends on tuning the process carefully, and for larger codes, there is still room for improvement. It's a promising new tool in the toolbox, but the job isn't finished yet.

Title: Auxiliary Nodes for BP Decoding of Quantum LDPC Codes
Authors: Daniel Tandler, Paul Bezner, and Stephan ten Brink (University of Stuttgart)

Problem Statement

Belief propagation (BP) decoding is a low-complexity method for Quantum Low-Density Parity-Check (QLDPC) codes, particularly Calderbank–Shor–Steane (CSS) codes. However, its performance is highly sensitive to the structure of the underlying Tanner graph. Two primary factors degrade BP performance in this context:

Short Cycles: Specifically 4-cycles, which cause messages in the BP algorithm to become strongly correlated, oscillate, and reinforce incorrect beliefs.
Degeneracy: A phenomenon unique to quantum syndrome-based decoding where multiple distinct error hypotheses yield the same syndrome. This ambiguity can lead to convergence failures (quantum-type trapping sets), especially in codes with low-weight stabilizers.

Existing solutions often involve post-processing BP results with more complex decoders (increasing latency) or modifying the decoding trajectory. Other approaches attempt to structurally modify the decoding graph, such as removing 4-cycles or splitting degeneracy sets, but these are often treated as separate techniques.

Methodology

The authors propose a general framework for augmenting the decoding graph of CSS codes using Auxiliary Variable Nodes (AVNs) and Auxiliary Check Nodes (ACNs). This framework is designed to be compatible with standard stabilizer measurement circuits without requiring additional physical measurements.

The core mathematical construction involves:

AVN Insertion: Appending an all-zero column to the parity-check matrix $H$ .
ACN Insertion: Appending a new row $c_{aug}^T = (c'^T, 1)$ to the matrix.
Syndrome Transformation: Mapping the original syndrome $s$ to an augmented syndrome $s_{aug}$ via a linear map $A$ .

This transforms the original system $He = s$ into an augmented system $H_{aug}e_{aug} = s_{aug}$ . The authors prove that the projection of the solution space of the augmented system back to the original variables is a bijection, ensuring the decoding problem remains equivalent.

The framework unifies two distinct interpretations:

Structural Modification (4-Cycle Removal): By choosing specific $c'$ vectors, ACNs can be used to eliminate 4-cycles in the graph (similar to the GARI method and optimized row-addition strategies).
Degeneracy Splitting (Affine Subcode Ensemble): Fixing an AVN to a specific value imposes an affine constraint ( $c'^T e = \text{const}$ ). If $c'$ is linearly independent of the existing rows (and logical operators), this splits the degeneracy set, effectively creating a subcode ensemble.

Key Contributions

Unified Framework: The paper formalizes a general algebraic framework that connects 4-cycle removal and affine subcode ensemble decoding (ASCED) through the use of AVNs and ACNs.
Dual Interpretation: It demonstrates that ACNs can serve to structurally modify the graph (removing cycles), while fixing AVNs imposes constraints that split degeneracy sets.
Graph-Derived Ensemble Decoder: The authors propose a new decoder that reuses the auxiliary nodes generated during the 4-cycle removal process as candidate constraints for a subcode ensemble. They implement two selection strategies:
- Static: Iteratively selects candidates that minimize the logical error rate or maximize coverage of failed error events.
- Adaptive: Runs a standard BP pass; if it fails to converge, it ranks AVNs by their Log-Likelihood Ratio (LLR) magnitudes and selects the least reliable ones to form the ensemble.
Back-Transformation Algorithm: An algorithm (Alg. 1) is provided to construct the mapping matrix $A$ required to compute the augmented syndrome from the measured syndrome, ensuring compatibility with standard hardware.

Results

The authors evaluated their methods on the [[72, 12, 6]] and [[90, 8, 10]] bivariate bicycle (BB) codes under circuit-level noise ( $p=0.001$ ).

4-Cycle Removal: The performance gains from removing 4-cycles depend heavily on the number of BP iterations ( $I$ ) and the check-node message scaling factor ( $\alpha$ ). While removing cycles generally improves performance, the augmented graph requires sufficient iterations for information to propagate. For small iteration counts, adding AVNs can initially worsen performance.
Ensemble Decoding: Both static and adaptive ensemble decoders significantly reduce the per-round logical error rate ( $p_L$ $p_{L}$ ) compared to standard BP on both the original and 4-cycle-free graphs.
- For the [[72, 12, 6]] code, an adaptive ensemble with 24 members nearly matches the performance of a BP decoder combined with Ordered Statistics Decoding (OSD-0), while maintaining latency comparable to plain BP due to parallel execution.
- For the larger [[90, 8, 10]] code, the ensemble decoder improves performance but still shows a gap compared to BP+OSD-0, even with 128 ensemble members.
Trade-offs: The paper notes that 4-cycle removal can increase the number of edges in the decoding graph (introducing 5 edges while removing 4), creating a trade-off between decoding complexity and performance.

Significance and Claims

The paper claims to provide a common algebraic interpretation for previously distinct techniques (short-cycle removal and subcode ensemble decoding). By formalizing the use of AVNs and ACNs, the authors offer a flexible degree of freedom in decoding graph design.

The authors modestly state that while their proposed ensemble decoders substantially reduce logical error rates compared to standard BP, further gains for larger codes may require complementary techniques, such as layered or windowed decoding, to reduce the effective problem size. They also note that the current LLR-based selection strategy for the adaptive ensemble might not always identify the optimal splitter candidates, suggesting room for refinement. The work establishes that these graph augmentations can be integrated into the standard stabilizer measurement framework without requiring new physical measurements.