Training quantum neural networks for Tl-limited hardware: what the qg symmetry filter does and does not do

A quantum neural network (QNN) built from gates that conserve the Hamming weight keeps its data in one weight sector, so the shots that left the sector can be discarded at readout. This is the qg filter of the qang library [1, 2]. We show that, with equal amplitude damping (T1) on every qubit, the filtered readout of such a network is exactly the noiseless one in any weight sector. The kept fraction of shots is (1 − γ)k d for weight k and depth d, and training under T1 with the filter gives exactly the parameters of noiseless training. We then measure what this is worth on four small classification datasets. Every result is reported with the filter and without it, and every prediction was committed before its run, failures included. Over 60 runs per model, a network trained on a simulator and run under T1 gains +2.5 points from the filter at weight 1 (95% CI [0.8, 4.1]) and +16.5 points at weight 2 ([11.8, 21.2]), recovering its noiseless accuracy exactly. A network trained under the calibrated noise without the filter catches up (diflerences of 0.1 points), and the filter alone corrects neither dephasing nor unequal T1 across qubits, although it loses less than one point up to a ±80% spread of 1/T1. Unequal T1 is corrected by training UÏtl the filter under the calibrated decay rates: the kept state depends only on their ratios, so the model lands within 0.5 points of noiseless at a ±100% spread, ignores a 10% calibration error and, at weight 2, loses 0.7 points under a 1.5-fold drift of all T1 that costs the raw noise-aware model 10. With noise-aware training and a richer readout, the model without the filter was 1.2 points better in one study, but over 60 new runs that gap shrank to +0.5 points and was not significant. The weight-2 deficit comes from the pair-product encoding, not from the readout; with an encoding that fills the whole sector and a full-rank readout, weight 2 reaches the weight-1 accuracy. On noise models of IBM and IonQ devices the filter cuts the error of the measured qgZ by 2.7-3.6×; on the IonQ Forte-1 noise model over 189 inputs of four datasets the filtered model is 1.1 points more accurate than the raw one, and calibrating the errors the filter keeps with five echo circuits per model halves the remaining error of the decision value: decisions that difler from the noiseless model fall from 15 (raw) and 7 (filter) to 3 over 945 readings on five device noise models. With a faithful weight-2 encoding the advantage of the unfiltered model under noise-aware training disappears, and a classifier given both the filtered and the raw features gains nothing over the filter alone. Over 513 readout configurations, including kept fractions down to 2.6 ×10−7, unequal T1 and dephasing, the filter has the lower readout error whenever about 20 shots are kept, and a closed-form rule predicts the winner. The same holds beyond the first model: at 8 qubits a weight-2 network keeps its noiseless 0.965 with the filter and drops to 0.705 without it; a five-class classifier (one class per qubit) keeps its noiseless accuracy under T1 with the filter and loses 15-18 points without it on five-class digits; training from 100 or 1000 shots per step with the filter is 1.0-1.5 points ahead of training without it; and the filtered gradient equals the noiseless one, while the raw one is K times smaller at weight 1. On five device noise models the echo calibration turns class confusions back: flipped decisions of the five-class model fall from 51 (raw) and 24 (filter) to 15 over 500 readings. The echo helped only where the filtered error was well above shot noise, and it still removed a third of the filtered error with a 28 × 28 matrix at 8 qubits. These models are classically simulable and no quantum advantage is claimed. The filter is