Noise Bias and Asymmetric Error Channels

In standard quantum error correction theory, errors are often modeled as depolarizing noise where $X$, $Y$, and $Z$ Pauli errors occur with equal probability. However, many physical qubit encodings have strongly asymmetric noise channels — one type of Pauli error is exponentially suppressed relative to others. Exploiting this noise bias enables dramatically more efficient error correction.

This note is about the directional axis of hidden, in-codespace errors. It is not a synonym for erasure dominance: erasure-error-vs-pauli-error concerns whether the decoder is told where a fault happened. Keep Pauli bias and erasure fraction as separate coordinates even when one device exploits both.

The Concept

A noise channel acting on a qubit has bias $\eta$ defined as:

$\eta =\frac{p_Z}{p_X+p_Y}$

where $p_X$, $p_Y$, $p_Z$ are the probabilities of the respective Pauli errors. For depolarizing noise, $\eta =1/2$. A biased-noise qubit has $\eta \gg 1$ (phase-flip dominated) or $\eta \ll 1$ (bit-flip dominated).

The general single-qubit Pauli channel is:

$\mathcal{E}(\rho )=(1-p_X-p_Y-p_Z)\rho +p_{X}X\rho X+p_{Y}Y\rho Y+p_{Z}Z\rho Z$

For a qubit with noise bias $\eta$ and total error rate $p$:

$p_Z=\frac{\eta }{1+\eta }p,p_X=p_Y=\frac{1}{2(1+\eta )}p,$

assuming $p_X=p_Y$ and the convention $\eta =p_Z/(p_X+p_Y)$ used above. Some papers instead define bias as $p_Z/p_X$; a numerical bias is meaningless unless the convention is stated.

Routing boundary

If the question is…	Start with…
Which hidden Pauli direction dominates after projection into the codespace?	This note
Does a detector, leakage check, or loss signature reveal the fault location?	erasure-error-vs-pauli-error
How should a decoder use both flagged events and biased unflagged residuals?	Read both and preserve the two-component channel model

Physical Realizations

Cat qubits (phase-flip biased)

The Kerr-cat qubit encodes $|0_L⟩$ and $|1_L⟩$ in coherent states $|\alpha ⟩$ and $|-\alpha ⟩$ stabilized by a two-photon drive. A bit-flip ($X$ error) requires the oscillator state to tunnel between the two wells in phase space, which is exponentially suppressed:

$p_X\propto e^{-2|\alpha {|}^2}$

Phase-flip errors ($Z$) arise from single-photon loss and grow linearly with $|\alpha {|}^2$. The resulting noise bias is:

${\eta }_{\text{cat}}\propto |\alpha {|}^2{e}^{2|\alpha {|}^2}$

reaching $\eta >10^4$ for $|\alpha {|}^2\gtrsim 4$ in experiments (Lescanne et al. 2020).

Erasure qubits (detected errors)

Erasure conversion is adjacent to, but distinct from, Pauli bias. The dominant process leaves the computational space and produces a detectable flag, while the residual in-codespace channel may still be unbiased or biased. Use erasure-error-vs-pauli-error for the decoding advantage from known fault locations; return here only to characterize the $X/Y/Z$ structure of the unflagged residual.

Superconducting 0-$\pi$ qubit

The 0-$\pi$ qubit is designed so that its two logical states are connected by a matrix element exponentially small in a circuit parameter, yielding exponential suppression of relaxation ($T_1$) errors while dephasing remains.

Exploiting Bias in QEC

For a qubit with noise bias $\eta \gg 1$:

1. Repetition code suffices for the dominant error: A simple $d$-qubit repetition code corrects $\lfloor (d-1)/2\rfloor$ phase-flip errors. Since bit-flips are exponentially rare, this provides exponential suppression of both error types.

2. Tailored surface codes: Rectangular surface codes with asymmetric dimensions ($d_Z\gg d_X$) match the noise asymmetry, reducing qubit overhead compared to square codes.

3. XZZX surface code: Ataides et al. (2021) showed that the XZZX variant of the surface code naturally exploits $Z$-biased noise, achieving higher thresholds ($\sim 50\%$ for pure $Z$ noise) than the standard CSS surface code.

4. Concatenation with bias: The cat qubit + repetition code concatenation (Guillaud & Mirrahimi 2019) can achieve a logical error rate:

$p_L\propto {\left(\frac{p_Z}{p_{\text{th}}}\right)}^{d/2}+d\cdot p_X$

where the first term is exponentially suppressed by code distance and the second by the physical bias.

Overhead Comparison

Noise Model	Surface Code Threshold	Logical Qubits per Physical (at $p=10^{-3}$)
Depolarizing ($\eta =0.5$)	~1%	~1000:1
Biased ($\eta =100$)	~4% (XZZX)	~300:1
Biased ($\eta =10^4$)	~10% (repetition + cat)	~50:1
Erasure-dominated	~50% (erasure)	~30:1

Historical Context

• Aliferis & Preskill (2008) first analyzed fault-tolerant thresholds under biased noise.
• Tuckett et al. (2018) showed that the surface code threshold increases dramatically with noise bias.
• Guillaud & Mirrahimi (2019) proposed cat qubit + repetition code concatenation.
• Ataides et al. (2021) introduced the XZZX surface code optimized for biased noise.
• Experimental bias ratios $>10^4$ demonstrated in cat qubits (Lescanne et al. 2020).

Key relationships

• erasure-error-vs-pauli-error — companion note for flagged-location information rather than Pauli-direction asymmetry
• kerr-cat-qubit and cat-codes — canonical phase-flip-biased hardware and its tailored coding layer
• 0-pi-qubit — protected circuit whose remaining error channel, not just total error rate, matters
• bacon-shor-code and surface-code-logical-qubit — code families whose geometry or checks can be adapted to a declared bias model

References

• kerr-cat-qubit
• cat-codes
• erasure-qubit
• surface-code-logical-qubit