Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2610.01466

Zero- Versus Infinite-Temperature Damping in Variational Quantum Circuits: Feature Scale, Sampling Cost, and Frame Gauge

Vu-Quoc-Minh Nguyen, Tuan-Vu Truong, Hoang-Long Nguyen, Trung-Khanh Le

quant-phSubmitted Oct 1, 2026• Announced Oct 2, 2026• 0 citations
arXivHTMLPDF

Abstract

The Pauli twirl of amplitude damping (AD) is generalized amplitude damping at infinite temperature: it keeps the contraction of AD and removes its non-unital term, so comparing the two in variational circuits isolates the zero-temperature bias, which acts mainly through the scale of the features. For random parameters, features under AD settle on a floor, which at strong damping is set by the last layer, has a closed form, and at fixed $T_1$ falls with temperature as $\tanh(\hbarω/2k_BT)$; under the twirls they shrink by a constant factor per layer, up to eight qubits. A trainable output scale removes most of the resulting accuracy differences, leaving AD ahead of its twirls by at most about three percentage points in our simulations; what it removes reappears as a cost in measurement shots: trained and tested with $10^3$ shots per image, a four-qubit classifier under AD at $p=0.3$ stays within 1.5 points of noiseless accuracy, while the twirled classifiers lose up to 33. At weaker damping the separation depth grows roughly as $(np)^{-1}\ln(1/p)$. The damping direction is a gauge when the damping follows complete entangling layers and the circuit boundaries are trainable; in an eigensolver it becomes physical inside a decomposed two-qubit gate.

AI Summary

The paper asks whether differences between amplitude damping (AD), its unital twirls, and reversed damping reflect changes in variational expressivity, feature scale, sampling cost, or circuit symmetries. It combines frame-reparametrization arguments, exact parameter-averaged second-moment calculations, and density-matrix simulations of single-qubit regression, four-qubit image classifiers, and a three-qubit eigensolver. A trainable affine readout removes most accuracy differences between AD and its twirls with exact expectation values; residual AD advantages reach about three percentage points. Before training, AD features approach a nonzero, approximately width-independent floor, whereas twirled features decay exponentially with depth, tested through eight qubits. A closed-form last-layer contribution explains the floor at strong damping; at fixed $T_1$, it scales approximately as $\tanh(\hbar\omega/2k_BT)$. Rescaling amplifies measurement noise: with $10^3$ shots per image in training and testing, at damping strength $p=0.3$, AD classifiers remain within 1.5 percentage points of noiseless accuracy, while twirled classifiers lose up to 33 points. The separation depth scales roughly as $(np)^{-1}\ln(1/p)$, reaching about 1500 layers for four qubits at $p=10^{-3}$. Sufficient Pauli-frame or time-reversal conditions make damping reversal an exact reparametrization. In the eigensolver, moving damping from inside decomposed two-qubit gates to after complete gates eliminates the direction effect. Training studies are limited to four qubits, small datasets, and eight classifier initializations. Finite-shot training uses Gaussian feature noise without interqubit correlations or parameter-shift gradient noise. Hardware validation, accompanying dephasing, and trained temperature dependence remain untested; reduced input sensitivity is dataset-dependent.

No sign of AI-generated textIncrementalStrong assumptions

Error found: The reported finite-shot training accuracies use a Gaussian feature-noise surrogate with exact differentiation, not fully shot-based circuit training, despite the abstract's unqualified wording. The abstract's damping-direction gauge condition omits gate-family and frame-propagation restrictions needed for the proved equivalence, although the specific demonstrated circuits satisfy them. The abstract presents a tanh temperature dependence for the feature floor that is exact for its last-layer contribution but only approximate for the full floor.

Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.

Related papers

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

Vicente Humberto Monteverde

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

•Oct 11, 2026•0 citations

The Encoding Gauge of Fermionic Variational Quantum Algorithms

S. M. Yousuf Iqbal Tomal, Abdullah Al Shafin

Whether a fermionic variational quantum algorithm offers a genuine quantum advantage depends on whether its outputs are classically hard to compute. We show that for such algorithms classical simulability is not an intrinsic property of the circuit. It depends on the fermion-to-qubit encoding. Standard encodings are related by Clifford transformations, so they form a gauge group, and we classify the standard simulability resources by how they transform under it. The Pauli weight, the causal-cone width, and the entanglement that sets the cost of tensor-network contraction are gauge-covariant: a generic interacting Hamiltonian whose truncated Pauli-propagation cost grows linearly in the system size under the Jordan-Wigner encoding becomes logarithmic under a tree encoding, and we exhibit the resulting change of verdict at a fixed physical ansatz. The dynamical-Lie-algebra dimension and the non-stabilizerness are gauge-invariant. From this split we prove that the encoding-optimized simulation cost isbounded by the invariant resources, so genuine, encoding-robust hardness requires both the dynamical Lie algebra and the magic to be large; the covariant weight is necessary but never sufficient. These routes are sufficient certificates of simulability, not a complete characterization, so the floor is an upper bound and a full classification of representation-invariant cost functionals is left open. We confirm both invariant simulation routes numerically, and we show that the dynamical-Lie-algebra dimension separates simulable from hard instances where the Pauli weight cannot; we also construct an explicit one-parameter interacting family for which all three certificate routes fail at once, so the floor is achieved. We then establish a sharp dichotomy with trainability. Because the cost landscape is encoding-invariant, the gradient and the barren-plateau variance are exactly gauge-invariant, which we verify to machine precision; we show further that this invariance follows from faithfulness rather than Clifford-equivalence and is robust even to symmetry tapering, so trainability can be altered only by changing what the dynamics explores, such as restricting the ansatz, never by relabeling the qubits. One may therefore re-encode a fermionic circuit to make it easier to simulate, but never to make it easier to train. These results reframe advantage claims for fermionic ansatze: an advantage that isnot gauge-invariant is an artifact of the encoding

•Jan 14, 2026•0 citations

The Encoding Gauge of Fermionic Variational Quantum Algorithms: Classical Simulability is Encoding-Relative, Trainability is Invariant

S. M. Yousuf Iqbal Tomal, Abdullah Al Shafin

Abstract Whether a fermionic variational quantum algorithm offers genuine advantage depends on whether its outputs resist classical computation. We show that classical simulability is encoding-dependent, not intrinsic to the circuit. Standard encodings are Clifford-related, forming a gauge group, and we classify resources by how they transform under it. Pauli weight, causal-cone width, and the entanglement setting tensor-network cost are gauge-covariant: an interacting Hamiltonian's truncated Pauli-propagation cost is linear under Jordan-Wigner but logarithmic under a tree encoding, changing the verdict at fixed ansatz. Dynamical-Lie-algebra dimension and non-stabilizerness are gauge-invariant. We prove the encoding-optimized simulation cost is bounded by the invariant resources, so encoding-robust hardness requires a large Lie algebra and large magic; covariant weight is necessary but never sufficient. These routes are sufficient certificates, not a complete characterization, so the floor is an upper bound; a full classification of representation-invariant costs remains open. We confirm both invariant routes numerically, show Lie-algebra dimension separates simulable from hard instances where weight cannot, and construct a one-parameter family where all three certificates fail at once, achieving the floor. We then establish a sharp dichotomy with trainability: because the cost landscape is encoding-invariant, gradient and barren-plateau variance are exactly gauge-invariant to machine precision. This follows from faithfulness, not Clifford-equivalence, and survives symmetry tapering, so trainability shifts only by changing what the dynamics explores, such as restricting the ansatz, never by relabeling qubits. One may re-encode a circuit to ease simulation but never to ease training. These results reframe advantage claims: an advantage that is not gauge-invariant is an artifact of the encoding.

•Jul 21, 2026•0 citations

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

Vicente Humberto Monteverde

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 corrects neither dephasing nor unequal T1 across qubits, although alone it loses less than one point up to a ±80% spread of 1/T1. 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. These models are classically simulable and no quantum advantage is claimed. The filter is a way to make noise-free training valid under relaxation, at a shot cost known in advance. The models are released as qang. qml (version 0.6.13, pip install qang).

•Oct 5, 2026•0 citations

Exact and Efficient Stabilizer Simulation of Thermal-Relaxation Noise for Quantum Error Correction

Sean R. Garner, Nathan M. Myers, Meng Wang +3 more

Stabilizer-based simulation of quantum error-correcting codes typically relies on the Pauli-twirling approximation (PTA) to render non-Clifford noise classically tractable, but PTA can distort the behavior of physically relevant channels such as thermal relaxation. Physically accurate noise simulation is needed to train decoders and understand the noise suppression capabilities of quantum error correction codes. In this work, we develop an exact and stabilizer-compatible model of qubit thermal relaxation noise and show that the combined amplitude damping and dephasing channel admits a fully positive probability decomposition into Clifford operations and reset whenever $T_2 \leqslant T_1$. For $T_2 > T_1$, the resulting decomposition is negative, but allows a smaller sampling overhead versus independent channels. We further introduce an approximated error channel with reset that removes the negativity of the decomposition while achieving higher channel fidelity to the true thermal relaxation than PTA, and extend our construction to finite temperature relaxation. We apply the exact combined model to investigate large surface codes and bivariate bicycle codes on superconducting platforms with realistic thermal relaxation error. The differing logical performances across code states further indicate that noise-model-informed decoders will be essential for accurately capturing thermal-noise structure in future fault-tolerant architectures.

quant-pharXiv:2512.09189•Physical Review Research•Dec 11, 2025•2 citations

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

Vicente Humberto Monteverde

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 corrects neither dephasing nor unequal T1 across qubits. These models are classically simulable and no quantum advantage is claimed. The filter is a way to make noise—free training valid under relaxation, at a shot cost known in advance. The models are released as qang. qml (version 0.6.0, pip install qang).

•Oct 3, 2026•0 citations
The Pauli twirl of amplitude damping (AD) is generalized amplitude damping at infinite temperature: it keeps the contraction of AD and removes its non-unital term, so comparing the two in variational circuits isolates the zero-temperature bias, which acts mainly through the scale of the features. For random parameters, features under AD settle on a floor, which at strong damping is set by the last layer, has a closed form, and at fixed T1T_1T1​ falls with temperature as tanh⁡(ℏω/2kBT)\tanh(\hbarω/2k_BT)tanh(ℏω/2kB​T); under the twirls they shrink by a constant factor per layer, up to eight qubits. A trainable output scale removes most of the resulting accuracy differences, leaving AD ahead of its twirls by at most about three percentage points in our simulations; what it removes reappears as a cost in measurement shots: trained and tested with 10310^3103 shots per image, a four-qubit classifier under AD at p=0.3p=0.3p=0.3 stays within 1.5 points of noiseless accuracy, while the twirled classifiers lose up to 33. At weaker damping the separation depth grows roughly as (np)−1ln⁡(1/p)(np)^{-1}\ln(1/p)(np)−1ln(1/p). The damping direction is a gauge when the damping follows complete entangling layers and the circuit boundaries are trainable; in an eigensolver it becomes physical inside a decomposed two-qubit gate.
The paper asks whether differences between amplitude damping (AD), its unital twirls, and reversed damping reflect changes in variational expressivity, feature scale, sampling cost, or circuit symmetries. It combines frame-reparametrization arguments, exact parameter-averaged second-moment calculations, and density-matrix simulations of single-qubit regression, four-qubit image classifiers, and a three-qubit eigensolver.
A trainable affine readout removes most accuracy differences between AD and its twirls with exact expectation values; residual AD advantages reach about three percentage points. Before training, AD features approach a nonzero, approximately width-independent floor, whereas twirled features decay exponentially with depth, tested through eight qubits. A closed-form last-layer contribution explains the floor at strong damping; at fixed T1T_1T1​, it scales approximately as tanh⁡(ℏω/2kBT)\tanh(\hbar\omega/2k_BT)tanh(ℏω/2kB​T). Rescaling amplifies measurement noise: with 10310^3103 shots per image in training and testing, at damping strength p=0.3p=0.3p=0.3, AD classifiers remain within 1.5 percentage points of noiseless accuracy, while twirled classifiers lose up to 33 points. The separation depth scales roughly as (np)−1ln⁡(1/p)(np)^{-1}\ln(1/p)(np)−1ln(1/p), reaching about 1500 layers for four qubits at p=10−3p=10^{-3}p=10−3.
Sufficient Pauli-frame or time-reversal conditions make damping reversal an exact reparametrization. In the eigensolver, moving damping from inside decomposed two-qubit gates to after complete gates eliminates the direction effect.
Training studies are limited to four qubits, small datasets, and eight classifier initializations. Finite-shot training uses Gaussian feature noise without interqubit correlations or parameter-shift gradient noise. Hardware validation, accompanying dephasing, and trained temperature dependence remain untested; reduced input sensitivity is dataset-dependent.
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
Whether a fermionic variational quantum algorithm offers a genuine quantum advantage depends on whether its outputs are classically hard to compute. We show that for such algorithms classical simulability is not an intrinsic property of the circuit. It depends on the fermion-to-qubit encoding. Standard encodings are related by Clifford transformations, so they form a gauge group, and we classify the standard simulability resources by how they transform under it. The Pauli weight, the causal-cone width, and the entanglement that sets the cost of tensor-network contraction are gauge-covariant: a generic interacting Hamiltonian whose truncated Pauli-propagation cost grows linearly in the system size under the Jordan-Wigner encoding becomes logarithmic under a tree encoding, and we exhibit the resulting change of verdict at a fixed physical ansatz. The dynamical-Lie-algebra dimension and the non-stabilizerness are gauge-invariant. From this split we prove that the encoding-optimized simulation cost isbounded by the invariant resources, so genuine, encoding-robust hardness requires both the dynamical Lie algebra and the magic to be large; the covariant weight is necessary but never sufficient. These routes are sufficient certificates of simulability, not a complete characterization, so the floor is an upper bound and a full classification of representation-invariant cost functionals is left open. We confirm both invariant simulation routes numerically, and we show that the dynamical-Lie-algebra dimension separates simulable from hard instances where the Pauli weight cannot; we also construct an explicit one-parameter interacting family for which all three certificate routes fail at once, so the floor is achieved. We then establish a sharp dichotomy with trainability. Because the cost landscape is encoding-invariant, the gradient and the barren-plateau variance are exactly gauge-invariant, which we verify to machine precision; we show further that this invariance follows from faithfulness rather than Clifford-equivalence and is robust even to symmetry tapering, so trainability can be altered only by changing what the dynamics explores, such as restricting the ansatz, never by relabeling the qubits. One may therefore re-encode a fermionic circuit to make it easier to simulate, but never to make it easier to train. These results reframe advantage claims for fermionic ansatze: an advantage that isnot gauge-invariant is an artifact of the encoding
Abstract Whether a fermionic variational quantum algorithm offers genuine advantage depends on whether its outputs resist classical computation. We show that classical simulability is encoding-dependent, not intrinsic to the circuit. Standard encodings are Clifford-related, forming a gauge group, and we classify resources by how they transform under it. Pauli weight, causal-cone width, and the entanglement setting tensor-network cost are gauge-covariant: an interacting Hamiltonian's truncated Pauli-propagation cost is linear under Jordan-Wigner but logarithmic under a tree encoding, changing the verdict at fixed ansatz. Dynamical-Lie-algebra dimension and non-stabilizerness are gauge-invariant. We prove the encoding-optimized simulation cost is bounded by the invariant resources, so encoding-robust hardness requires a large Lie algebra and large magic; covariant weight is necessary but never sufficient. These routes are sufficient certificates, not a complete characterization, so the floor is an upper bound; a full classification of representation-invariant costs remains open. We confirm both invariant routes numerically, show Lie-algebra dimension separates simulable from hard instances where weight cannot, and construct a one-parameter family where all three certificates fail at once, achieving the floor. We then establish a sharp dichotomy with trainability: because the cost landscape is encoding-invariant, gradient and barren-plateau variance are exactly gauge-invariant to machine precision. This follows from faithfulness, not Clifford-equivalence, and survives symmetry tapering, so trainability shifts only by changing what the dynamics explores, such as restricting the ansatz, never by relabeling qubits. One may re-encode a circuit to ease simulation but never to ease training. These results reframe advantage claims: an advantage that is not gauge-invariant is an artifact of the encoding.
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 corrects neither dephasing nor unequal T1 across qubits, although alone it loses less than one point up to a ±80% spread of 1/T1. 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. These models are classically simulable and no quantum advantage is claimed. The filter is a way to make noise-free training valid under relaxation, at a shot cost known in advance. The models are released as qang. qml (version 0.6.13, pip install qang).
Stabilizer-based simulation of quantum error-correcting codes typically relies on the Pauli-twirling approximation (PTA) to render non-Clifford noise classically tractable, but PTA can distort the behavior of physically relevant channels such as thermal relaxation. Physically accurate noise simulation is needed to train decoders and understand the noise suppression capabilities of quantum error correction codes. In this work, we develop an exact and stabilizer-compatible model of qubit thermal relaxation noise and show that the combined amplitude damping and dephasing channel admits a fully positive probability decomposition into Clifford operations and reset whenever T2⩽T1T_2 \leqslant T_1T2​⩽T1​. For T2>T1T_2 > T_1T2​>T1​, the resulting decomposition is negative, but allows a smaller sampling overhead versus independent channels. We further introduce an approximated error channel with reset that removes the negativity of the decomposition while achieving higher channel fidelity to the true thermal relaxation than PTA, and extend our construction to finite temperature relaxation. We apply the exact combined model to investigate large surface codes and bivariate bicycle codes on superconducting platforms with realistic thermal relaxation error. The differing logical performances across code states further indicate that noise-model-informed decoders will be essential for accurately capturing thermal-noise structure in future fault-tolerant architectures.
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 corrects neither dephasing nor unequal T1 across qubits. These models are classically simulable and no quantum advantage is claimed. The filter is a way to make noise—free training valid under relaxation, at a shot cost known in advance. The models are released as qang. qml (version 0.6.0, pip install qang).