Surface Code Performance Under Biased Noise

Hello! Welcome to the third lesson in our module on recent advances in quantum error correction.

In our previous lesson, we explored quantum LDPC codes as a way to improve upon the surface code's poor encoding rate, aiming for a significant reduction in qubit overhead. We saw this as one major axis of research: developing codes with better asymptotic properties.

Today, we will investigate a complementary and equally important research direction: tailoring codes to the specific, non-ideal noise found in real quantum hardware. Our focus will be to analyze the performance degradation of the standard surface code under a biased noise model where dephasing errors are more probable than bit-flip errors. By understanding why the standard code is suboptimal in this common scenario, we will lay the groundwork for understanding the next generation of noise-aware codes.

1. The Reality of Biased Noise

In theoretical analyses, we often assume a simple, symmetric noise model like the depolarizing channel, where Pauli $X$, $Y$, and $Z$ errors occur with equal probability. However, many physical qubit implementations exhibit a strong asymmetry or bias in their noise.

For many qubit platforms, the relaxation time $T_1$ (which governs energy decay, associated with $X$ and $Y$ bit-flip type errors) is much longer than the dephasing time $T_2$ (which governs the loss of phase coherence, associated with $Z$ phase-flip errors). This condition, $T_1\gg T_2$, leads to a biased noise model where $Z$ errors are far more frequent than $X$ or $Y$ errors.

To formalize this, please read the following brief introductions from two key papers on this topic. They define the biased noise model and list some of the qubit technologies where it is prevalent.

As these resources clarify, the biased noise channel is defined by the total error probability $p=p_x+p_y+p_z$ and the bias parameter $\eta =p_z/(p_x+p_y)$. In the limit $\eta \to \infty$, we have pure dephasing or pure Z noise. The depolarizing channel corresponds to $\eta =1/2$ (since $p_x=p_y=p_z$ gives $p_z/(p_z+p_z)=1/2$). Given that this bias is a natural feature of many promising qubit platforms, understanding its impact on our leading error correction code is critical.

2. The Standard Surface Code's "Blind Spot"

Let's briefly recall the structure of the standard (or CSS) surface code. Qubits reside on the edges of a 2D square lattice. The stabilizer generators are:

• Vertex stabilizers ($A_v$): Products of Pauli-$X$ operators on the four qubits surrounding a vertex.
• Plaquette stabilizers ($B_p$): Products of Pauli-$Z$ operators on the four qubits bordering a plaquette.

An error is detected when it anti-commutes with a stabilizer, flipping its eigenvalue from +1 to -1 and creating a syndrome excitation.

• A $Z$ error anti-commutes with $X$ operators $⟹$ Detected by vertex stabilizers.
• An $X$ error anti-commutes with $Z$ operators $⟹$ Detected by plaquette stabilizers.
• A $Y$ error anti-commutes with both $⟹$ Detected by both vertex and plaquette stabilizers.

Now, consider the case of highly Z-biased noise ($\eta \gg 1$). The vast majority of errors will be single-qubit $Z$ errors. What happens when such an error occurs?

It anti-commutes with the two $X$-type vertex stabilizers at its endpoints, creating a pair of syndrome defects. However, it commutes with all $Z$-type plaquette stabilizers. This leads to a crucial problem.

The plaquette stabilizers are blind to the dominant error type.

In a highly biased noise environment, half of your stabilizers provide no information about the errors that are actually happening. The decoder, which relies on the syndrome to infer the error configuration, is effectively working with half the information it was designed to have.

This insight is the core of the issue. The following reading explains this "blindness" and introduces the concept of tailoring the code to fix it—a topic we will explore in the next lesson.

3. Quantifying the Performance Degradation

One might intuitively think that more predictable noise (i.e., higher bias, lower entropy) should make error correction easier. However, because the standard surface code is mismatched to this noise structure, its performance actually degrades. The most direct measure of this is the error correction threshold, $p_c$, the maximum physical error rate below which the logical error rate can be made arbitrarily low.

• For the standard surface code with a standard minimum-weight perfect matching (MWPM) decoder under depolarizing noise ($\eta =1/2$), the threshold is approximately 15.5%. Other decoders can push this to ~18%.
• For the same code and decoder under pure Z noise ($\eta \to \infty$), the problem of correcting Z errors using only vertex stabilizers becomes equivalent to a different statistical mechanics model. The result, as cited in the Tuckett et al. paper (Ref. [7]), is that the threshold drops to 10.9%.

This is a profound and counter-intuitive result: as the noise becomes more structured and seemingly simpler to correct, the code's performance gets worse. The fault lies not in the noise, but in the code's inability to leverage that structure.

4. The Compounding Effect of Circuit-Level Noise

The situation is further complicated when we move from this simplified "code-capacity" model (where only data qubits are noisy) to a more realistic circuit-level noise model. In this model, the gates used to perform syndrome measurements are also faulty.

To measure the $X$-stabilizers needed to detect $Z$ errors, the syndrome extraction circuit typically involves CNOT gates and Hadamard gates. As your background in simulating physical systems might suggest, the Hamiltonians that generate these gates can themselves mix different error types.

A Z error (${\sigma }_z$) on a target qubit undergoing a CNOT interaction may not commute with the gate Hamiltonian, causing the evolution to transform the $Z$ error into a $Y$ or $X$ error. In essence, the gates required to detect the dominant error can inadvertently "un-bias" the noise, converting "easy" $Z$ errors into "harder" $X$ or $Y$ errors that are less frequent and thus harder for the decoder to diagnose.

The following reading explores this subtlety.

This analysis reveals that the performance of the standard surface code under biased noise is even worse than the code-capacity model suggests. The very act of running the error correction protocol can actively work against the potential advantages of the biased noise, further degrading performance.

Conclusion

In this lesson, we have dissected the reasons behind the standard surface code's poor performance in the presence of a common, realistic noise model.

Key Takeaways:

• Biased Noise: Many quantum systems exhibit biased noise ($T_1\gg T_2$), where dephasing ($Z$) errors are much more likely than bit-flip ($X,Y$) errors.
• A Fundamental Mismatch: The standard surface code's $Z$-type plaquette stabilizers are blind to $Z$ errors. This means that in a highly Z-biased setting, half of the syndrome measurements provide no information about the dominant error source.
• Performance Degradation: This mismatch leads to a counter-intuitive drop in the error correction threshold, from $\sim 15.5\%$ for depolarizing noise to $\sim 10.9\%$ for pure Z noise.
• Circuit-Level Effects: The problem is compounded at the circuit level, where non-bias-preserving gates like CNOT and Hadamard can convert the dominant $Z$ errors into other error types, reducing the effective bias and further hindering performance.

Preview of the next lesson:
The shortcomings of the standard surface code in this setting present a clear opportunity for improvement. If the problem is a mismatch between the stabilizers and the noise, the solution is to change the stabilizers. In our next lesson, we will describe the design of a tailored stabilizer code, such as the XZZX surface code, and explain how its structure provides enhanced protection against a specific biased noise model. We will see how a simple change to the stabilizers can dramatically improve the error correction threshold by ensuring all stabilizers are sensitive to the most likely errors.