ruvector-neural-decoder

Neural Quantum Error Decoder (NQED) - GNN-based quantum error correction with O(d²) Mamba architecture

Overview

ruvector-neural-decoder implements a neural network-based quantum error decoder that combines Graph Neural Networks (GNN) with Mamba state-space models for efficient syndrome decoding on surface codes.

Key Features

• Graph Attention Encoder: Multi-head attention over syndrome graphs with learned positional encodings
• Mamba O(d²) Decoder: Linear-time sequence modeling via selective state spaces (vs O(d⁴) transformers)
• Min-Cut Feature Fusion: Integrates structural coherence signals from ruvector-mincut
• Pure Rust: No external ML frameworks - uses ndarray for tensor operations
• WASM Compatible: Designed for browser and edge deployment

Installation

Add to your Cargo.toml:

[dependencies]
ruvector-neural-decoder = "0.1"

Quick Start

use ruqu_neural_decoder::{NeuralDecoder, DecoderConfig};

// Create decoder for distance-5 surface code
let config = DecoderConfig {
    distance: 5,
    embed_dim: 64,
    hidden_dim: 128,
    num_gnn_layers: 3,
    num_heads: 4,
    ..Default::default()
};

let mut decoder = NeuralDecoder::new(config);

// Decode a syndrome (9 stabilizers for d=5 rotated surface code)
let syndrome = vec![true, false, true, false, false, false, false, false, true];
let correction = decoder.decode(&syndrome)?;

println!("X corrections: {:?}", correction.x_corrections);
println!("Confidence: {:.2}%", correction.confidence * 100.0);

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     NQED Pipeline                           │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Syndrome    ┌──────────────┐    ┌──────────────┐          │
│  Round ────► │ Syndrome→    │───►│ GNN Encoder  │          │
│              │ DetectorGraph│    │ (GraphRoPE)  │          │
│              └──────────────┘    └──────┬───────┘          │
│                                         │                   │
│                                         ▼                   │
│              ┌──────────────┐    ┌──────────────┐          │
│              │ Min-Cut      │───►│ Feature      │          │
│              │ Engine       │    │ Fusion       │          │
│              └──────────────┘    └──────┬───────┘          │
│                                         │                   │
│                                         ▼                   │
│                                  ┌──────────────┐  ┌──────┐│
│                                  │ Mamba        │─►│Corr- ││
│                                  │ Decoder      │  │ection││
│                                  └──────────────┘  └──────┘│
└─────────────────────────────────────────────────────────────┘

Module Overview

use ruqu_neural_decoder::{DecoderConfig, NeuralDecoder};

let config = DecoderConfig {
    distance: 7,              // Surface code distance
    embed_dim: 128,           // Node embedding dimension
    hidden_dim: 256,          // Hidden layer dimension
    num_gnn_layers: 4,        // Number of GNN layers
    num_heads: 8,             // Attention heads
    mamba_state_dim: 64,      // Mamba hidden state size
    use_mincut_fusion: true,  // Enable min-cut features
    dropout: 0.1,             // Training dropout rate
};

let mut decoder = NeuralDecoder::new(config);

use ruqu_neural_decoder::{GraphBuilder, NodeType};

// Create a d=3 surface code graph
let graph = GraphBuilder::new()
    .with_distance(3)
    .add_node(0, 0, NodeType::XStabilizer)?
    .add_node(0, 1, NodeType::ZStabilizer)?
    .add_node(1, 0, NodeType::ZStabilizer)?
    .add_node(1, 1, NodeType::XStabilizer)?
    .connect_grid()?  // Auto-connect adjacent stabilizers
    .build()?;

println!("Graph has {} nodes, {} edges", graph.num_nodes(), graph.num_edges());

// Sequential decoding for time-series syndromes
let syndromes = vec![
    vec![false, true, false, false],   // Round 1
    vec![true, true, false, false],    // Round 2
    vec![true, false, false, false],   // Round 3
];

for (round, syndrome) in syndromes.iter().enumerate() {
    let correction = decoder.decode(syndrome)?;
    println!("Round {}: {} corrections, confidence {:.1}%",
        round,
        correction.x_corrections.len(),
        correction.confidence * 100.0
    );

    // Reset state between rounds if needed
    decoder.reset();
}

use ruqu_neural_decoder::GNNEncoder;

let gnn = GNNEncoder::new(gnn_config);
let graph = GraphBuilder::from_surface_code(5)
    .with_syndrome(&syndrome)?
    .build()?;

// Get node embeddings before Mamba decoding
let embeddings = gnn.encode(&graph)?;
println!("Embedding shape: {:?}", embeddings.shape());

// Get attention weights from last layer
let attention_weights = gnn.last_attention_weights();

[dependencies]
ruvector-neural-decoder = { version = "0.1", features = ["ruqu-integration"] }

use ruqu::{QuantumFabric, SyndromeRound};
use ruqu_neural_decoder::NeuralDecoder;

// Create fabric with neural decoder backend
let fabric = QuantumFabric::new()
    .with_neural_decoder(NeuralDecoder::new(config))?;

// Process syndrome cycle with coherence assessment
let syndrome_round = SyndromeRound::from_measurements(&measurements)?;
let decision = fabric.process_cycle_neural(&syndrome_round)?;

match decision.gate {
    GateDecision::Permit => apply_correction(decision.correction),
    GateDecision::Defer => request_human_review(),
    GateDecision::Deny => abort_computation(),
}

cargo bench --package ruvector-neural-decoder --bench neural_decoder_bench

cargo test -p ruvector-neural-decoder

API Reference

Core Types

/// Decoder configuration
pub struct DecoderConfig {
    pub distance: usize,
    pub embed_dim: usize,
    pub hidden_dim: usize,
    pub num_gnn_layers: usize,
    pub num_heads: usize,
    pub mamba_state_dim: usize,
    pub use_mincut_fusion: bool,
    pub dropout: f32,
}

/// Correction output
pub struct Correction {
    pub x_corrections: Vec<usize>,  // Bit flip locations
    pub z_corrections: Vec<usize>,  // Phase flip locations
    pub confidence: f64,            // 0.0 to 1.0
    pub decode_time_ns: u64,        // Latency
}

Error Handling

All operations return Result<T, NeuralDecoderError>:

pub enum NeuralDecoderError {
    DimensionMismatch { expected: usize, got: usize },
    EmptyGraph,
    InvalidSyndrome(String),
    EncodingError(String),
    DecodingError(String),
}

References

1. AlphaQubit: Neural decoders for quantum error correction
2. Mamba: Linear-Time Sequence Modeling
3. GNNs for Quantum Error Correction

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.