arxiv: 2604.16779 · v3 · submitted 2026-04-18 · 🪐 quant-ph · cs.LG

Recognition: unknown

Q-SINDy: Quantum-Kernel Sparse Identification of Nonlinear Dynamics with Provable Coefficient Debiasing

Samrendra Roy, Syed Bahauddin Alam

Authors on Pith no claims yet

Pith reviewed 2026-05-10 07:38 UTC · model grok-4.3

classification 🪐 quant-ph cs.LG

keywords Q-SINDyquantum kernelSINDycoefficient cannibalizationorthogonalizationnonlinear dynamicssparse regressionquantum feature maps

0 comments

The pith

Orthogonalizing quantum kernel features against polynomials removes exact cannibalization bias from Q-SINDy equation recovery.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Q-SINDy by augmenting sparse identification of nonlinear dynamics with quantum feature maps, then isolates a failure mode in which quantum terms pull coefficient weight away from the polynomial basis. It derives the precise bias formula and demonstrates that projecting the quantum features orthogonal to the polynomial column space at fitting time sets this bias to zero. The result holds to machine precision on standard test systems and preserves SINDy’s structural recovery rates while naive quantum augmentation destroys them. A dynamics-specific diagnostic predicts when the bias appears, and controls confirm the effect is not merely from added feature count.

Core claim

Quantum features in Q-SINDy induce coefficient cannibalization bias given exactly by Δξ_P = (P^T P)^{-1} P^T Q ξ̂_Q; orthogonalizing the quantum matrix Q against the polynomial matrix P at fit time eliminates the bias identically, restoring the polynomial coefficients that vanilla SINDy would have found.

What carries the argument

The orthogonalization step that projects quantum feature columns Q perpendicular to the polynomial basis P before the sparse regression, which nulls the derived cannibalization term.

If this is right

Orthogonalized Q-SINDy recovers the same true-positive rates as vanilla SINDy across Duffing, Van der Pol, Lorenz, Lotka-Volterra, cubic oscillator and Rössler systems.
Uncorrected quantum augmentation lowers true-positive recovery by as much as 100 percent on the same systems.
The bias correction remains effective under depolarizing noise up to 2 percent per gate and applies unchanged to Burgers’ equation.
An R^2_Q diagnostic computed on the derivative data correlates with cannibalization severity at Pearson r = 0.70.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same orthogonalization step can be inserted into any kernel-augmented sparse regression to protect the original basis coefficients.
The R^2_Q diagnostic offers a practical gatekeeper for deciding whether quantum features add value on a given dataset.
Because the correction is linear-algebraic rather than hardware-specific, it remains available even if future quantum feature maps become more expressive.

Load-bearing premise

The polynomial basis is treated as the reference explanatory space whose coefficients must be recovered without distortion.

What would settle it

On any system whose true governing equation is known, run orthogonalized Q-SINDy and check whether the recovered polynomial coefficients differ from those of vanilla SINDy by more than 10^{-12}.

Figures

Figures reproduced from arXiv: 2604.16779 by Samrendra Roy, Syed Bahauddin Alam.

**Figure 2.** Figure 2: Theorem 1 verified to machine precision. For Duffing (left) and Lotka-Volterra (right, noise-free), predicted bias (orange) matches observed bias (blue) per polynomial feature to relative error < 10−12. Orthogonalized coefficients (green, barely visible) coincide with vanilla to < 10−14, confirming Theorem 3. 5.1 Dense noise sweep: the headline finding [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Dense noise sweep (N=5 trials). Across Duffing, Van der Pol, and Lorenz, orthogonalized Q-SINDy (green) overlays vanilla SINDy (blue) at all noise levels, while naive Q-augmentation (orange) degrades significantly. RBF-augmented (red) baselines fail more severely, ruling out feature count as the mechanism. σ = 0.05 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Three additional systems (Lotka-Volterra, cubic oscillator, R¨ossler). Pattern persists: orthogonalized Q-SINDy (green) tracks vanilla (blue) while naive Q-augmentation (orange) degrades. Lotka-Volterra shows cannibalization at σ = 0 itself; cubic oscillator from σ = 0.01 onward, confirming the failure mode is coefficient-geometric rather than noise-driven [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

**Figure 5.** Figure 5: RBF hyperparameter sweep. Across 20 configurations, the best RBF-augmented variant achieves TPR=0.50 vs. vanilla’s 0.92. Most configurations give TPR= 0. Feature count and bandwidth are not the mechanism. regime, correctly predicting that naive augmentation and orthogonalization will both recover ground truth—which they do (TPR= 1.0 for both). This is the first validation of the R2 Q diagnostic’s predictio… view at source ↗

**Figure 6.** Figure 6: Dynamics-aware R2 Q diagnostic vs. column-space overlap. Across 10 (system, feature-map) combinations, the refined R2 Q diagnostic (right) predicts cannibalization severity with Pearson r = 0.70 (p = 0.023), versus r = 0.55 (p = 0.098) for column-space overlap (left). The previously anomalous Duffing/IQP case (square marker, low cannibalization despite high overlap) aligns with the R2 Q prediction. the bot… view at source ↗

**Figure 7.** Figure 7: Hardware-noise robustness. Under depolarizing channels up to 2% per gate (realistic NISQ regime), orthogonalized Q-SINDy retains vanilla TPR; naive augmentation retains its cannibalization. cannot settle. The hardware-noise results (Sec. 5.6) demonstrate the method’s portability to current NISQ devices; a real-hardware demonstration at 2–3 qubits is the natural near-term validation step. Limitations. (1) … view at source ↗

read the original abstract

Quantum feature maps offer expressive embeddings for classical learning tasks, and augmenting sparse identification of nonlinear dynamics (SINDy) with such features is a natural but unexplored direction. We introduce \textbf{Q-SINDy}, a quantum-kernel-augmented SINDy framework, and identify a specific failure mode that arises: \emph{coefficient cannibalization}, in which quantum features absorb coefficient mass that rightfully belongs to the polynomial basis, corrupting equation recovery. We derive the exact cannibalization-bias formula $\Delta\xi_P = (P^\top P)^{-1}P^\top Q\,\hat\xi_Q$ and prove that orthogonalizing quantum features against the polynomial column space at fit time eliminates this bias exactly. The claim is verified numerically to machine precision ($<10^{-12}$) on multiple systems. Empirically, across six canonical dynamical systems (Duffing, Van der Pol, Lorenz, Lotka-Volterra, cubic oscillator, R\"ossler) and three quantum feature map architectures (ZZ-angle encoding, IQP, data re-uploading), orthogonalized Q-SINDy consistently matches vanilla SINDy's structural recovery while uncorrected augmentation degrades true-positive rates by up to 100\%. A refined dynamics-aware diagnostic, $R^2_Q$ for $\dot X$, predicts cannibalization severity with statistical significance (Pearson $r=0.70$, $p=0.023$). An RBF classical-kernel control across 20 hyperparameter configurations fails more severely than any quantum variant, ruling out feature count as the cause. Orthogonalization remains robust under depolarizing hardware noise up to 2\% per gate, and the framework extends without modification to Burgers' equation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Q-SINDy shows orthogonalization removes quantum feature bias from SINDy coefficients exactly via a standard linear algebra identity, with the uncorrected version degrading recovery in practice.

read the letter

The main thing to know is that adding quantum kernel features to SINDy creates a coefficient bias that pulls weight away from the polynomial terms, and the paper gives an exact formula for it plus a fix that restores the original SINDy coefficients. They derive Δξ_P = (P^T P)^{-1} P^T Q ξ̂_Q from the joint least-squares equations and prove that projecting the quantum features orthogonal to the polynomial columns sets the cross term to zero, so the polynomial estimator matches the pure-polynomial case at every step of the thresholded iteration. This is verified to machine precision on Duffing, Van der Pol, Lorenz, Lotka-Volterra, cubic oscillator, and Rössler, and they show uncorrected augmentation drops true-positive rates by up to 100% while the orthogonalized version matches vanilla SINDy. The RBF classical-kernel control across 20 configurations rules out raw feature count as the driver, and the R^2_Q diagnostic correlates with bias severity at r=0.70. The noise test up to 2% depolarizing error is a practical plus, and the extension to Burgers' equation is noted without extra work. The derivation is standard linear algebra, so the machine-precision match and identical recovery are expected once you orthogonalize; they are confirmatory rather than independent proof. Still, showing that the bias actually harms structural recovery on these systems is useful, because it demonstrates the problem is not just theoretical. The implicit assumption that the polynomial basis must stay untouched is reasonable for interpretability but could be restrictive if quantum features might improve the model in other regimes. This is for researchers working on sparse nonlinear dynamics who are testing quantum or kernel augmentations. The algebra is clean, the experiments are consistent across systems and maps, and the controls are adequate, so it deserves a serious referee even if the core fix is algebraic.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces Q-SINDy, a quantum-kernel-augmented SINDy framework for nonlinear dynamics identification. It identifies the coefficient cannibalization failure mode in which quantum features absorb mass from the polynomial basis, derives the exact bias formula Δξ_P = (P^⊤ P)^{-1} P^⊤ Q ξ̂_Q, proves that orthogonalizing the quantum feature matrix Q against the polynomial column space P at fit time eliminates the bias exactly (by enforcing P^⊤ Q_perp = 0 and decoupling the normal equations), and verifies the elimination to machine precision (<10^{-12}) on multiple systems. Empirically, orthogonalized Q-SINDy matches vanilla SINDy structural recovery across six dynamical systems (Duffing, Van der Pol, Lorenz, Lotka-Volterra, cubic oscillator, Rössler) and three quantum maps (ZZ-angle encoding, IQP, data re-uploading), while uncorrected augmentation degrades true-positive rates by up to 100%; a dynamics-aware diagnostic R²_Q for Ẋ predicts cannibalization severity (Pearson r=0.70, p=0.023), an RBF classical-kernel control across 20 configurations fails more severely, and the method remains robust to depolarizing noise up to 2% per gate. The framework extends without modification to Burgers' equation.

Significance. If the central linear-algebra result holds, the work supplies a principled, exact debiasing technique that permits safe incorporation of expressive quantum features into sparse regression without corrupting the interpretable polynomial coefficients or the SINDy thresholding procedure. The derivation from standard least-squares identities, the explicit proof of bias elimination, and the machine-precision numerical verification (<10^{-12}) constitute clear strengths, as does the consistent empirical matching to the vanilla baseline and the control experiment that rules out feature-count explanations. The approach may generalize to other hybrid quantum-classical sparse identification tasks in physics and engineering.

minor comments (3)

The bias formula in the abstract uses the notation Δξ_P = (P^⊤ P)^{-1}P^⊤ Q ξ̂_Q; ensure identical symbol definitions and hat notation appear in the main-text derivation (likely §3) to avoid reader confusion on first reading.
The claim that orthogonality is preserved under SINDy’s sequential thresholded least-squares iterations is central to asserting identical recovery at every step; a one-sentence reference to the fact that column subsetting preserves the P^⊤ Q_perp = 0 property would improve clarity without lengthening the proof.
The R²_Q diagnostic is reported with Pearson r=0.70 (p=0.023); state the number of independent configurations or systems over which the correlation was computed so that the statistical significance can be assessed directly.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive and accurate summary of the manuscript, which correctly identifies the coefficient cannibalization issue, the exact bias formula, the orthogonalization proof, and the empirical results across the tested systems. We appreciate the recommendation to accept.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The central derivation applies the normal equations of ordinary least squares to the joint feature matrix [P Q] to obtain the exact bias formula Δξ_P = (PᵀP)⁻¹PᵀQ ξ̂_Q; this is a direct algebraic identity independent of SINDy or quantum features. Orthogonalization of Q against the column space of P sets PᵀQ_perp = 0 by the definition of the orthogonal complement, decoupling the estimators with no additional assumptions. Numerical checks to machine precision (<10^{-12}) simply confirm the identity and are expected by construction. Performance comparisons across six dynamical systems, three quantum feature maps, an RBF control, and noise robustness constitute external empirical validation rather than self-referential evidence. No self-citations, ansatzes, or renamings load-bear the proof.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard linear algebra for orthogonal projection and regression; no free parameters are introduced for the debiasing result itself, and no new entities are postulated.

axioms (1)

standard math Standard properties of orthogonal projection in linear algebra (P^⊤Q = 0 after orthogonalization implies no bias transfer)
Invoked in the derivation of Δξ_P and the proof that orthogonalization eliminates bias exactly.

pith-pipeline@v0.9.0 · 5615 in / 1279 out tokens · 30803 ms · 2026-05-10T07:38:09.220569+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

7 extracted references · 4 canonical work pages · 1 internal anchor

[1]

Harnessing quantum algorithms in nuclear engineering

Samrendra Roy, Kazuma Kobayashi, Sajedul Talukder, and Syed Bahauddin Alam. Harnessing quantum algorithms in nuclear engineering. InNuclear Plant Instrumentation, Control & Human-Machine Interface Technology (NPIC & HMIT 2025). American Nuclear Society, 2025a. Vojtˇ ech Havl´ ıˇ cek, Antonio D. C´ orcoles, Kristan Temme, Aram W. Harrow, Abhinav Kandala, J...

2025
[2]

Sparse identification of quantum hamiltonian dynamics via quantum circuit learning.arXiv:2602.14556,

Yusei Tateyama and Yuzuru Kato. Sparse identification of quantum hamiltonian dynamics via quantum circuit learning.arXiv:2602.14556,

work page arXiv
[3]

Niklas Heim, Atiyo Ghosh, Oleksandr Kyriienko, and Vincent E. Elfving. Quantum model-discovery. arXiv:2111.06376,

work page arXiv
[4]

Protocols for solving nonlinear ODEs without grid evalua- tion,

Annie E. Paine, Vincent E. Elfving, and Oleksandr Kyriienko. Physics-informed quantum machine learning: Solving nonlinear differential equations in latent spaces without costly grid evaluations.arXiv:2308.01827,

work page arXiv
[5]

PennyLane: Automatic differentiation of hybrid quantum-classical computations

Ville Bergholm et al. PennyLane: Automatic differentiation of hybrid quantum-classical computations. arXiv:1811.04968,

work page internal anchor Pith review arXiv
[6]

Foundation models for dynamic nuclear systems: Energy-efficient surrogates via deep operator learning

Samrendra Roy, Kazuma Kobayashi, Souvik Chakraborty, Sajedul Talukder, and Syed Bahauddin Alam. Foundation models for dynamic nuclear systems: Energy-efficient surrogates via deep operator learning. Transactions of the American Nuclear Society, 133(1):158–161, 2025b. Kazuma Kobayashi, Farid Ahmed, Samrendra Roy, Mehedi H. Tusar, Sajedul Talukder, and Syed...

2025
[7]

System Equations IC Poly deg

13 A System specifications Table 2: Systems, parameters, and library configurations. System Equations IC Poly deg. STLSQ thresh. Duffing ˙x=y,˙y=−x−0.3x 3 −0.1y[1,0] 3 0.05 Van der Pol ˙x=y,˙y=µ(1−x 2)y−x, µ= 1 [2,0] 3 0.05 Lorenz [Lorenz, 1963] ˙x= 10(y−x),˙y=x(28−z)−y,˙z=xy− 8 3 z[1,1,1] 2 0.1 Lotka-Volterra ˙x= 2 3 x− 4 3 xy,˙y=xy−y[1,1] 2 0.05 Cubic o...

1963