Clique Decoder in Quantum Information

• Clique Decoder is a term used in quantum information to denote both a cryogenic surface-code local decoder and a permutation-invariant quantum circuit for hidden clique labeling.
• In the surface-code context, Clique_L1 and Clique_L2 reduce syndrome data and lower off-chip I/O demands with constant-time, neighbor-limited logic achieving up to 18.4× bandwidth reduction.
• In the graph learning context, the decoder leverages Sₙ symmetry to efficiently identify hidden cliques using a variational, symmetry-restricted quantum circuit that outperforms standard ansätze.

Searching arXiv for the cited papers and closely related entries to ground the article. {"query":"id:(Jia et al., 15 Jul 2025) OR id:(Mansky et al., 3 Jun 2025) clique decoder", "max_results": 10} In the 2025 arXiv literature, the term Clique Decoder appears in two distinct quantum-information settings. In surface-code quantum error correction, it denotes a cryogenic, local decoder family—the original Clique decoder, retrospectively denoted Clique_L1, and its extension Clique_L2—that performs constant-time, neighbor-limited syndrome processing and forwards only residual complex patterns to a global decoder (Jia et al., 15 Jul 2025). In quantum machine learning for graph problems, Clique Decoder denotes a permutation-invariant quantum circuit that labels the vertices of a hidden $k$-clique by exploiting covariance under node relabeling and by measuring all qubits in the $Z$ basis after a symmetry-restricted variational evolution (Mansky et al., 3 Jun 2025). The shared label is therefore terminologically unified but operationally heterogeneous.

1. Terminological scope and disambiguation

The two usages differ in task, representation, and algorithmic objective. In the surface-code setting, the object being decoded is a syndrome pattern on a planar lattice of parity qubits, and the principal design goal is to reduce cryo$\to$room-temperature I/O bandwidth while preserving local correction coverage for common error events (Jia et al., 15 Jul 2025). In the graph-learning setting, the object being decoded is a vertex-membership label vector $\ell \in \{0,1\}^n$ for a hidden clique in an undirected graph $G=(V,E)$, and the principal design goal is to align the variational hypothesis class with the problem’s $S_n$ symmetry (Mansky et al., 3 Jun 2025).

Usage	Task	Defining mechanism
Clique_L1 / Clique_L2	Surface-code syndrome decoding	Local parity checks on neighboring syndrome nodes
Permutation-invariant Clique Decoder	Hidden-clique labeling in graphs	$S_n$-invariant variational circuit and $Z$-basis readout

A recurrent source of confusion is that the phrase decoder has different meanings in these two contexts. In the first, it refers to a real-time QEC component that maps syndrome information to correction actions or off-chip escalation. In the second, it refers to a classifier-like procedure that maps an encoded graph state to clique-membership bits. The common name reflects structural selectivity rather than a shared implementation.

2. Clique_L1 as a local surface-code decoder

The original Clique decoder, later referred to as Clique_L1, was introduced as a cryogenic, local length-1 decoder for the surface code (Jia et al., 15 Jul 2025). Surface-code syndromes are arranged on a 2D grid of ancilla (“parity”) qubits. When a data-qubit error occurs, it flips the measurement outcome of the neighboring parity checks, producing syndrome “nodes.” A global Minimum-Weight Perfect Matching (MWPM) decoder would construct a graph whose vertices are active syndromes and whose edge weights are the shortest-path distances between them; Clique_L1 instead handles the common case of isolated, length-1 error chains using local combinational logic.

Its core rule is a local parity-check rule. For each parity qubit $a$ with outcome $s_a=1$, the decoder collects the four neighboring parity outcomes $Z$0. If exactly one neighbor is also active, the error is assumed to lie on the single data qubit adjacent to both active checks, and a Pauli correction is applied to that qubit. If there are 0 or 2 active neighbors, the pattern is treated as complex and its syndrome bits are forwarded to a full decoder. Boundary handling is inherited from the same logic, with fewer neighbors at edges and corners (Jia et al., 15 Jul 2025).

The architectural point of Clique_L1 is not to replace MWPM, but to support it by eliminating easy cases at cryogenic temperature. The design is stated to be implementable in Single-Flux-Quantum (SFQ) or cryo-CMOS logic at $Z$1, using $Z$2 simple gates per syndrome qubit plus minimal glue logic. Under this operating model, Clique_L1 eliminates $Z$3–$Z$4 of syndrome signatures from the cryo$Z$5room-temperature I/O stream, with the exact fraction depending on the physical error rate $Z$6 and code distance $Z$7 (Jia et al., 15 Jul 2025).

3. Clique_L2 and the extension to length-2 space errors

Clique_L2 extends the original local-decoding rule to correct length-2 error chains in space (Jia et al., 15 Jul 2025). The extension is motivated by the observation that such chains become non-trivial at higher physical error rates and higher code distances, and that clustered errors can invalidate a decoder tuned only to isolated length-1 events.

The key change is a relaxed activation rule combined with an even-parity criterion. In addition to the original length-1 rule, Clique_L2 permits local decoding when the central syndrome is $Z$8 and exactly two of the four neighbors are active. According to the paper summary, this captures all non-overlapping length-2 chains in the horizontal, vertical, and diagonal orientations. The added logic is explicitly described as low-cost, and the decoder is presented as an incremental rather than wholesale redesign (Jia et al., 15 Jul 2025).

The implementation is organized as a four-stage decoding pipeline. Stage 1 performs the original Clique_L1 length-1 decoding on active cliques with odd neighbor parity. Stage 2 performs the new length-2 decoding using parity-two patterns, flipping the two data qubits adjacent to the two active checks. Stage 3 applies the inherited edge and corner rules. Stage 4 marks any remaining active syndromes as complex and forwards them off chip. To prevent intersecting cliques from decoding simultaneously, the lattice is 4-colored, so that no two neighboring cliques decode in the same cycle; in consequence, at most 4 sub-cycles are required to cover all patches (Jia et al., 15 Jul 2025).

This design preserves the original local-decoder philosophy while relaxing the original activation condition. A plausible implication is that Clique_L2 treats the local decoder not as a single-pattern recognizer but as a small family of recognizers for the lowest-weight error patterns that dominate practical operation.

4. Mathematical structure, complexity, and hardware characteristics

The surface-code formulation is expressed using an implicit syndrome graph $Z$9, where $\to$0 is the set of parity nodes. In the global MWPM formulation, the decoder assigns an edge weight

$\to$1

where $\to$2 is the graph distance, and seeks a minimum-weight perfect matching $\to$3. Clique_L1 and Clique_L2 do not construct or solve this global optimization; instead, they inspect each node $\to$4 and its direct neighborhood $\to$5 of size $\to$6 or fewer on boundaries (Jia et al., 15 Jul 2025).

The complexity contrast is explicit. Global MWPM is reported as $\to$7 or $\to$8 optimizations over $\to$9 syndrome nodes, whereas Clique_L1/L2 performs constant-time $\ell \in \{0,1\}^n$0 work per node, giving $\ell \in \{0,1\}^n$1 total per syndrome round (Jia et al., 15 Jul 2025). The significance of this comparison lies not only in asymptotics but also in implementation locality: Clique_L2’s logic is neighbor-limited and highly parallel.

The reported hardware overhead is correspondingly modest. Clique_L1 uses $\ell \in \{0,1\}^n$2 simple logic gates per parity qubit, mainly XORs and ANDs. Clique_L2 adds only 2–4 more gates per clique for the even-parity length-2 check. The 4-stage pipeline and 4-color schedule impose a worst-case latency of 4 clock cycles per syndrome evaluation round, which remains $\ell \in \{0,1\}^n$3. The paper summary further states that the total additional area is $\ell \in \{0,1\}^n$4 relative to Clique_L1, and that the added cryogenic power remains in the low-$\ell \in \{0,1\}^n$5 regime per logical qubit, far below typical $\ell \in \{0,1\}^n$6 cooling-power constraints (Jia et al., 15 Jul 2025).

The integration consequence is an explicit reduction in off-fridge traffic. Clique_L2 is said to eliminate up to $\ell \in \{0,1\}^n$7 of out-of-fridge transmissions, thereby relaxing I/O wiring requirements and reducing the load on room-temperature decoders. This suggests an architectural role in large-scale systems where cryogenic locality and interconnect scarcity dominate practical design decisions.

5. Reported performance of Clique_L1 and Clique_L2 under multiple noise models

The principal reported metric is the fraction $\ell \in \{0,1\}^n$8 of syndrome events forwarded off-chip, where lower values indicate broader local coverage (Jia et al., 15 Jul 2025). The summarized numerical results are:

Scenario	Clique_L1	Clique_L2
Data-Qubit-Only, $\ell \in \{0,1\}^n$9, $G=(V,E)$0	$G=(V,E)$1	$G=(V,E)$2
Uniform Noise, 2 rounds, $G=(V,E)$3, $G=(V,E)$4	$G=(V,E)$5	$G=(V,E)$6
Gaussian-Clustered Noise, $G=(V,E)$7, $G=(V,E)$8	$G=(V,E)$9	$S_n$0

For the Data-Qubit-Only setting at $S_n$1 and $S_n$2, the summary gives a $S_n$3 reduction in forwarded events. For Uniform Noise at $S_n$4 and $S_n$5, the corresponding improvement is $S_n$6. For Gaussian-Clustered Noise at $S_n$7 and $S_n$8, the reported improvement is $S_n$9 (Jia et al., 15 Jul 2025).

The paper also reports aggregate bandwidth-reduction figures. Under data-qubit-only errors and uniformly random noise, Clique_L2 achieves up to $S_n$0 decoding bandwidth reduction over Clique_L1. Under clustered errors and longer error chains, it achieves up to $S_n$1 decoding bandwidth reduction over Clique_L1; in the Dual-Error Model at $S_n$2, $S_n$3, Clique_L1 is described as having $S_n$4 forwarded fraction, while Clique_L2 retains small $S_n$5 through $S_n$6, yielding up to $S_n$7 reduction (Jia et al., 15 Jul 2025).

A stated trend across all models is that, as $S_n$8 or as errors become more clustered, the relative advantage of Clique_L2 grows markedly. This is consistent with the decoder’s explicit expansion from isolated length-1 events to non-overlapping length-2 spatial chains.

6. The permutation-invariant quantum-circuit Clique Decoder for hidden-clique labeling

A second object called Clique Decoder is introduced in the paper “Clique detection using symmetry-restricted quantum circuits” (Mansky et al., 3 Jun 2025). Here the problem is not QEC but graph inference. Given an undirected graph $S_n$9 with $Z$0, one assigns one qubit per node and prepares the graph-dependent state

$Z$1

followed by Controlled-$Z$2 gates along every edge:

$Z$3

The embedding therefore places the graph’s connectivity into the entanglement pattern of the initial state (Mansky et al., 3 Jun 2025).

The variational circuit is constrained to be permutation-invariant under the full symmetric group $Z$4. For any permutation $Z$5 and ansatz unitary $Z$6, the symmetry condition is

$Z$7

with $Z$8 the qubit-permutation operator. One layer is built from globally shared generators:

$Z$9

This layer is repeated $a$0 times; the reported experiments use $a$1, giving $a$2 trainable parameters, which is stated to be on par with the comparator ansätze (Mansky et al., 3 Jun 2025).

The target is a label vector $a$3 whose entries indicate whether each node belongs to the hidden $a$4-clique. The loss is implemented by a bit-wise penalty Hamiltonian

$a$5

and the variational objective is

$a$6

After preparing

$a$7

all qubits are measured in the $a$8 basis, with

$a$9

The readout convention is $s_a=1$0 for “node $s_a=1$1 in clique” and $s_a=1$2 for “node $s_a=1$3 not in clique” (Mansky et al., 3 Jun 2025).

Training uses the quantum natural gradient

$s_a=1$4

where $s_a=1$5 is the quantum Fisher information matrix. The symmetry restriction reduces the parameter count from the standard $s_a=1$6 angles per layer to 3 shared angles per layer, described as a factor-of-$s_a=1$7 reduction that sharply limits barren-plateau effects (Mansky et al., 3 Jun 2025).

On balanced Erdős–Rényi graphs, the reported final test-set accuracies after 50 epochs are as follows:

Ansatz	6 qubits	8 qubits
Permutation-invariant $s_a=1$8	$s_a=1$9	$Z$00
Cyclic-invariant $Z$01	$Z$02	$Z$03
Standard (strongly entangling)	$Z$04	$Z$05

The tasks were 4-clique among 6 nodes and 5-clique among 8 nodes, with means and 95\% CI reported over 10 random seeds. Figure 1 of the paper summary states that the $Z$06-invariant ansatz converges within $Z$07–$Z$08 epochs, the $Z$09 variant takes $Z$10–$Z$11, and the standard ansatz remains at random-guess levels (Mansky et al., 3 Jun 2025).

The theoretical rationale is that the clique problem has an intrinsic symmetry: both the input graph and the target label transform covariantly under node permutations,

$Z$12

Enforcing $Z$13 invariance therefore restricts optimization to a variational subspace matched to the task’s true symmetry. A plausible implication is that, in this second usage of the term, Clique Decoder denotes less a specialized decoding primitive than a symmetry-aligned quantum model for structured label recovery.