Sliding-Window Decoding
Decoding a moving window of syndrome rounds lets a decoder emit corrections continuously instead of waiting for a computation to finish.
Why not decode everything at once
A quantum computation may run for millions of syndrome rounds. Waiting until the end to decode the whole spacetime history is impossible: the data would be enormous and no feed-forward gate could ever get its correction in time. Sliding-window decoding solves this by decoding a fixed-size window of recent rounds and advancing it as new rounds arrive.
Commit and overlap
Within each window the decoder makes a decision, but only commits the corrections in the older part of the window, leaving a fresh region undecided because errors near the window's leading edge may connect to future rounds. The window then slides forward, overlapping the previous one, so every round is eventually decoded with enough context on both sides. The committed region is final; the buffer region is revisited.
- A fixed window bounds memory and latency regardless of total runtime.
- Only the settled, older part of each window is committed.
- Overlap between windows preserves accuracy at the seams.
- Window size trades latency against the risk of splitting a long error chain.
The window size is the key parameter. Too small a window can cut through a long error string and misdecode it; too large a window increases latency and memory. In practice a window a few times the code distance, with an overlap of about one distance, retains near-batch accuracy while keeping latency bounded.
Sliding-window decoding is what makes decoders like matching and union-find usable in real time on an ongoing computation. It pairs naturally with parallel hardware decoders, since separate windows or spatial regions can be decoded concurrently and stitched together, distributing the load and meeting the per-round deadline.