Skip to main content
Back to All Articles
August 2026Multi-RoundBy Guillaume Lessard · 5 min read

Space-Time Decoding with Noisy Measurements

Detector differencing, temporal and spatial weights, measurement glitches, and streaming scope.

Author: Guillaume Lessard / qector.store
Source: QECTOR Decoder v3 Reference Manual v1.0.0
Date: August 2026
DOI: 10.5281/zenodo.21941046
Tags: space-time decoding, detector error model, measurement noise, streaming QEC

Abstract

With noisy syndrome extraction, one round of measurements is not the same object as a clean syndrome. Data faults and measurement faults must be represented together so the decoder can decide whether a detection event is spatial or temporal. QECTOR's SpaceTimeDecoder lifts the problem into a (2+1)-dimensional detector lattice; StreamingDecoder and SlidingWindowDecoder provide separate multi-round primitives. This post derives detector differencing, anisotropic edge weights, the lifted faithfulness equation, a measurement-glitch example, and the manual's explicit boundary on Python streaming claims.

1. Raw syndromes are histories

Let s_(c,t) be the observed syndrome bit for check c at round t. It mixes data-error effects with measurement noise. Instead of decoding each round as an isolated two-dimensional problem, form detector differences:

dc,t=sc,t+sc,t−1(mod2),d_{c,t} = s_{c,t} + s_{c,t-1} \pmod 2,

with the initial layer handled by the chosen boundary convention.

The difference records changes rather than absolute values. A persistent data fault creates events at the temporal boundaries of its interval. A measurement flip that appears and then disappears creates a pair of events at the same detector in adjacent layers. The decoder can then prefer a time-like explanation when the measurement prior makes it more likely.

The detector error model (DEM) is the object that enumerates these mechanisms. QECTOR exposes it through a small verified API:

from qector_decoder_v3 import dem

model = dem.from_stim(circuit)        # detector error model
decoder = model.make_decoder()        # routed backend
if model.is_graphlike:
    graph = model.collapse_to_graph() # MWPM-ready form

2. The lifted parity-check problem

The lifted detector matrix H_ST contains both data-fault columns and measurement-fault columns. Rows represent detectors at particular time layers. A space-like column changes detector locations according to the base code; a time-like column connects the same detector across adjacent rounds.

The correction is now a space-time vector c_ST, and the correctness equation is

HSTcST=d(mod2).H_{ST}c_{ST} = d \pmod 2.

For graphlike lifted models, paths in the detector lattice have the same boundary property as paths in the two-dimensional graph. The SpaceTimeDecoder therefore solves one matching problem over all rounds, instead of making independent decisions that confuse measurement faults with data faults.

3. Anisotropic weights

Data and measurement mechanisms need not share a probability. Assign

wspace=log⁡(1−pdatapdata),w_{space} = \log\left(\frac{1-p_{data}}{p_{data}}\right),

and

wtime=log⁡(1−pmeaspmeas).w_{time} = \log\left(\frac{1-p_{meas}}{p_{meas}}\right).

The decoder compares path costs under these priors. If measurement faults are more likely than data faults, a time-like path may be cheaper. If the measurement channel is cleaner, spatial paths are preferred. The weights belong to the detector model, not to a tuning knob.

For example, p_data = 0.01 gives w_space around 4.595, while p_meas = 0.03 gives w_time around 3.476. The example shows the direction of the preference. It is not a claim about logical-error performance.

4. A glitch that must not become a data correction

Assume clean rounds 1 and 2. At round 3, check 2 reports a spurious flip. At round 4 the measurement returns to its previous value. Differencing produces two detector events:

appearance at (check 2, round 3)
disappearance at (check 2, round 4)

The natural explanation is a pair of time-like edges. The projected spatial correction is empty because no data qubit changed. This is the regression behaviour described in the manual: a lone measurement glitch that reverts on the next round must not corrupt the spatial correction.

5. The lifted faithfulness theorem

Let d be the detector-event vector. If the decoder returns a space-time correction with

HSTcST=d,H_{ST}c_{ST} = d,

then every detector event is reproduced by the correction. Projecting away the measurement columns yields a spatial correction whose final-round syndrome differs from the raw final measurement only by the final-round measurement-error boundary term.

That projection statement is a property of the lifted chain complex. It is not a claim that measurement noise has been eliminated from the physical device. The logical test stays separate: once the final spatial correction is formed, its residual must be scored in the appropriate logical-observable or stabilizer quotient.

6. Full space-time versus streaming primitives

These names describe different scopes:

SurfaceDocumented role
SpaceTimeDecoderFull lifted detector-lattice decoding across T rounds
StreamingDecoderRust streaming primitive with accumulated history
SlidingWindowDecoderRust window with decay weighting
Python streaming layerPer-round/windowed workflow for batching and telemetry; not full 3D matching by itself

For a decayed window, the manual gives a geometric truncation bound. If 0 <= lambda < 1,

∥SW−S∞∥1≤λW1−λ∥s∥∞.\|S_W-S_{\infty}\|_1 \leq \frac{\lambda^W}{1-\lambda}\|s\|_{\infty}.

The expression describes the history approximation. It does not by itself establish a numerical logical-fidelity loss or a hardware-control guarantee.

7. Choosing between offline and online paths

Use the full lifted decoder when the experiment needs a circuit-level or multi-round decision that consumes the complete detector history. Use a streaming primitive when the workflow needs bounded state, incremental commits, or telemetry, and can accept the stated windowing semantics.

In either case, record:

number of rounds and boundary convention
data and measurement error priors
detector construction and DEM settings
decoder family and weight mode
window length and decay factor, if applicable
observable scoring, seed, environment, and artifact hash

Do not describe a per-round Python window as full circuit-level space-time matching. The manual explicitly reserves that claim for the lifted SpaceTimeDecoder path.

8. Why this matters to hardware teams

The detector lattice is an interface between measurement electronics and decoding software. It lets a decoder consume repeated stabilizer outcomes without forcing the hardware layer to declare every event a data fault. It also makes priors visible: calibration data can inform time-like and space-like weights, while the algebraic gate verifies the emitted detector correction.

This separation supports heterogeneous systems. The physical layer produces rounds, the DEM describes mechanisms, the space-time decoder handles correlated history, and the validation harness scores observables under a named noise model.

Takeaway

Noisy measurements do not invalidate syndrome decoding; they change the matrix. XOR differencing turns a history into detector events, anisotropic weights express the relative fault priors, and the lifted equation preserves the same syndrome-faithfulness idea in one higher dimension. Streaming is useful, but its scope must be stated precisely rather than marketed as full circuit-level matching.

Reference

Guillaume Lessard, QECTOR Decoder v3: Syndrome-Faithful Decoding - Foundations, Algorithms, and Architecture of a Fifteen-Backend Quantum Error Correction Engine, v1.0.0, August 2026. DOI: 10.5281/zenodo.21941046.