Feature Request: Bond Dimension Cutoff for Tensor Network Decoder

[takh04]

Summary

Add bond dimension cutoff capability to enable approximate decoding with polynomial computational cost instead of the current exponential exact solving.

Current Behavior

The tensor network decoder currently solves problems exactly with exponential computational cost, making it impractical for larger quantum error correction codes and error cycles.

Requested Feature

Add a bond_dimension or max_bond_dim parameter to enable approximate decoding with polynomial computational cost.

Proposed API:

decoder = qec.get_decoder(
    "tensor_network_decoder",
    H,
    logical_obs=logicals,
    noise_model=noise_model,
    contract_noise_model=True,
    bond_dimension=64  # New parameter
)

Benefits

• Enable scaling to larger quantum codes
• Allow cost vs. performance trade-offs

Use Cases

• Benchmarking decoder accuracy vs. bond dimension
• Large-scale quantum error correction simulations

Environment

• Package: cudaq-qec[tensor-network-decoder]

[npancotti]

Thanks @takh04! This can be currently done by using quimb + contengra as a backend. It requires to dig a little into the code, but the quimb.tensor.TensorNetwork object you would like to approximately contract is basically decoder.full_tn. We are considering to support approximate contractions in the near future via cuTensorNet.