Interpretation of Union-Find Decoder on Weighted Graphs and Application to XZZX Surface Code

Yue Wu $^\dagger$

Namitha Liyanage $^\dagger$

Shruti Puri $^\ddagger$

Lin Zhong $^\dagger$

$^\dagger$

$^\ddagger$

Evidence of Similarity

  • MWPM decoder: Minimum-Weight Perfect Matching Decoder
  • Union-Find decoder: similar threshold on the same weighted decoding graph
    • 2.6% (UF) compared to 2.9% (MWPM) under code capacity noise model in CSS surface code
      Delfosse N, Nickerson N H. Almost-linear time decoding algorithm for topological codes[J]. Quantum, 2021, 5: 595.
    • 0.76% (UF) compared to 0.90% (MWPM) under circuit-level noise model in CSS surface code
      Huang S, Newman M, Brown K R. Fault-tolerant weighted union-find decoding on the toric code[J]. Physical Review A, 2020, 102(1): 012419.

Why are They Similar?

  • MWPM decoder: Minimum-Weight Perfect Matching Decoder
  • Union-Find decoder: similar threshold on the same weighted decoding graph
    • 2.6% (UF) compared to 2.9% (MWPM) under code capacity noise model in CSS surface code
      Delfosse N, Nickerson N H. Almost-linear time decoding algorithm for topological codes[J]. Quantum, 2021, 5: 595.
    • 0.76% (UF) compared to 0.90% (MWPM) under circuit-level noise model in CSS surface code
      Huang S, Newman M, Brown K R. Fault-tolerant weighted union-find decoding on the toric code[J]. Physical Review A, 2020, 102(1): 012419.
  • What's the relationship between them?
  • Under what conditions UF decoder has the same accuracy as MWPM decoder?

CSS Surface Code

code distance d = 3

Data qubits
Z stabilizers
(detect Pauli X errors on data qubits)
X stabilizers
(detect Pauli Z errors on data qubits)

Focus on Z Stabilizers

code distance = {{ d }}

Data qubits
(dotted lines ..............)
Z stabilizers
(detect Pauli X errors on data qubits)
Non-trivial measurements

MWPM Decoder Finds the Most Likely Error Pattern

i.i.d. errors: the most likely error pattern $\iff$ the lowest amount of errors

more Likely (9 errors)
less Likely (10 errors)

Decoder Fails when Logical Error Occurs

A logical X error occurs: an error chain connecting the left and right boundaries

all syndromes are canceled out

Union-find Decoder

Cluster: a connected region of exploratory region

odd (even) cluster contains odd (even) number of nontrivial measurements

Odd clusters
expand evenly, merge others
Even clusters
stop expanding
Touching Boundary
terminate

Union-find Decoder Solves an Error Pattern

Peeling algorithm: find errors only inside each cluster to cancel out syndrome

these errors correspond to a perfect matching (usually not minimum-weighted)

Blossom Algorithm as a MWPM Solver

Dual variables: drawn as the "radius" or "width" of exploratory region

some exploratory region grows while some may shrink (depends on internal structure)

Blossom
considered as a single vertex
Temporary matches
stop expanding
Alternating tree
expand + shrink
Touching Boundary
terminate
Dual variables:
$y_{A}$ $=$ {{ display_dual(dual_A) }}
$y_{B}$ $=$ {{ display_dual(dual_B) }}
$y_{C}$ $=$ {{ display_dual(dual_C) }}
$y_{D}$ $=$ {{ display_dual(dual_D) }}
$y_{E}$ $=$ {{ display_dual(dual_E) }}
$y_{\{A,B,C\}}$ $=$ {{ display_dual(dual_ABC) }}
$y_{\{A,B,C,D,E\}}$ $=$ {{ display_dual(dual_ABCDE) }}

Blossom Algorithm Solves a Perfect Matching

Blossom algorithm: two vertices cannot be matched if they are in different clusters

any minimum-weighted perfect matching corresponds to an error pattern

Error Pattern $\iff$ Perfect Matching

Both UF and MWPM Decoder: all errors inside clusters

UF Decoder
Blossom Algorithm

Same Constraint and Similar Clusters

Both UF and MWPM Decoder: all errors inside clusters

In this example, their error patterns only differ by several stabilizer operators

UF Decoder
MWPM Decoder

Relationship Between UF and MWPM Decoder

What's the relationship between them?

UF decoder is an approximate MWPM decoder using the same weighted graph

  • UF decoder drops internal details of clusters, the blossom algorithm keeps them
    • UF decoder: better time complexity, lower accuracy
    • MWPM decoder: worse time complexity, higher accuracy
  • Combining UF and MWPM decoder could potentially balance between decoding time and accuracy
    • e.g. by dropping internal details of clusters only when the cluster is too large

Conditions of UF Equally Accurate as MWPM Decoder

Under what conditions UF decoder has the same accuracy as MWPM decoder?

Claim: weighted union-find decoder is equally accurate as MWPM decoder if

  1. All clusters have the same shape in UF decoder and (one possible solution of) blossom algorithm
UF Decoder
MWPM Decoder
(Blossom Algorithm)
MWPM Decoder
(Blossom Algorithm)
MWPM Decoder
(Blossom Algorithm)

Conditions of UF Equally Accurate as MWPM Decoder

Under what conditions UF decoder has the same accuracy as MWPM decoder?

Claim: weighted union-find decoder is equally accurate as MWPM decoder if

  1. All clusters have the same shape in UF decoder and (one possible solution of) blossom algorithm
  2. If any cluster touches both boundaries (corner cases), this syndrome inherently confuses both of them
UF Decoder: Confused
UF Decoder: Confused
MWPM Decoder: Deterministic

XZZX Surface Code: UF as Accurate as MWPM Decoder

Claim: weighted union-find decoder is equally accurate as MWPM decoder if

  1. All clusters have the same shape in UF decoder and (one possible solution of) blossom algorithm
  2. If any cluster touches both boundaries (corner cases), this syndrome inherently confuses both of them

Application: XZZX surface code with infinite noise bias under perfect measurements

  1. The shape of clusters is the same
  2. The second condition stands for any code distance
UF Decoder
MWPM Decoder

Numerical Evidence: UF as Accurate as MWPM Decoder

  • Noise model: Pauli errors only on data qubits, single round of perfect measurement
    Bonilla Ataides J P, Tuckett D K, Bartlett S D, et al. The XZZX surface code[J]. Nature communications, 2021, 12(1): 1-12.
    • Noise bias: $\eta$, total error rate: $p=7\%$ (fixed) [physical error rate $p_Z = \frac{\eta p}{1+\eta}$, $p_X = p_Y = \frac{p}{2(1+\eta)}$]
  • Result: when $\eta \xrightarrow[]{} \infty$, the accuracy loss vanishes

UF Decoder Approaching MWPM Decoder in Accuracy

  • Noise model: circuit-level noise model, noise bias: $\zeta$
    Darmawan A S, et al. Practical quantum error correction with the XZZX code and Kerr-cat qubits[J]. PRX Quantum, 2021, 2(3): 030345.
  • Result: UF decoder close to MWPM decoder at $\zeta = 100$
    • threshold = 0.90% (UF decoder) v.s. 0.98% (MWPM decoder)

Conclusion: UF Decoder Approximates MWPM Decoder

Analytical and numerical results: UF is equally (closely) accurate as MWPM decoder

Future work: combining UF and MWPM decoder

UF Decoder
MWPM Decoder
(Blossom Algorithm)

XZZX Surface Code: Circuit-level Noise Model

  • Noise model: correlated Pauli errors during two-qubit gates and single-qubit errors
    Darmawan A S, et al. Practical quantum error correction with the XZZX code and Kerr-cat qubits[J]. PRX Quantum, 2021, 2(3): 030345.
    • Initialization and measurement: $p_X = p_Y = p_Z / \zeta$
    • CPHASE: $p_{Z\otimes I} = p_{I\otimes Z} = p_Z$, $p_{\text{others}} = p_Z / \zeta$
    • Standard CNOT: $p_{Z\otimes I} = p_Z$, $p_{I\otimes Z} = p_{Z\otimes Z} = \frac{3 p_Z}{8}$, $p_{I\otimes Y} = p_{Z\otimes Y} = \frac{p_Z}{8}$, $p_{\text{others}} = p_Z / \zeta$
    • Bias-preserving CNOT: $p_{Z\otimes I} = p_Z$, $p_{I\otimes Z} = p_{Z\otimes Z} = \frac{p_Z}{2}$, $p_{\text{others}} = p_Z / \zeta$

UF Decoder Approaching MWPM Decoder in Accuracy

  • Noise model: correlated Pauli errors during two-qubit gates and single-qubit errors
  • Result: UF decoder is only close to but not equally accurate as MWPM decdoer
    • In circuit-level noise model, the decoding graph is no longer 1-dimensional