Readout Dependent Parameter Identifiability of Quantum Kernels

AI summary

Overview

Research area: Quantum machine learning theory — specifically the identifiability of parameters in quantum kernels built from measured Pauli expectation values, and how the choice of readout (which observables are measured) and the choice of metric on those features determine which circuit parameters the kernel can actually "see."

Technical level: Advanced. The paper is written for readers comfortable with quantum states, Pauli observables, Gaussian/fidelity kernels, Jacobians and rank arguments, and Gram-matrix algebra. The prose is theorem-and-proof driven with a substantial appendix structure; there is no tutorial scaffolding.

Scope (one sentence): The paper characterizes exactly when two parameter settings of a quantum feature map produce the same projected Gaussian kernel, bounds how much of the lost parameter dependence a fixed metric can recover, and proves an exact four-qubit example where only a nonlinear feature lift — one appended scalar coordinate — breaks a rank ceiling of three to reach rank four.

What This Paper Is About

A circuit parameter can change the prepared quantum state and its measured expectation values while leaving every pairwise value of the resulting Gaussian kernel unchanged — so counting parameters, or even finding all directions that change measured features, does not tell you the dimension of the kernel family you are actually training. The authors ask which parameter distinctions survive a specified readout and a specified positive definite metric, and how dependence that is lost at the kernel level can be recovered. Their answer splits into three separate questions: an exact algebraic description of kernel equivalence, hard limits on what any single fixed metric (or fixed affine feature expansion) can recover, and — kept deliberately apart — finite-shot statistical resolution and compiled execution cost.

Key Contributions

1. A coefficient Gram criterion for kernel equivalence (Theorem 1). For features of the form f_θ(x) = C(θ)h(x) + u(θ) with analytic coefficients, a fixed metric M ≻ 0, fixed bandwidth γ > 0, and a data basis {1, h_1, …, h_A} that is affinely independent on the actual input domain, the kernel agrees on all input pairs if and only if G(θ) = C(θ)ᵀ M C(θ) agrees. The same object certifies global coordinate inactivity (∂_j G ≡ 0) and gives the generic differential rank bound r ≤ min{K, A(A+1)/2}. A corollary sharpens this to r ≤ min{K, Aρ − ρ(ρ−1)/2} with ρ = min(m, A), where m is the number of measured features.

2. An exact decision procedure on a restricted subclass (Corollary 3), for coefficient entries that are finite Laurent polynomials in e^{iθ_j/q} over an effective algebraic number field, plus an explicit sufficient circuit grammar (fixed algebraic unitaries and Pauli rotations with rational angle coefficients). The authors are explicit that this is a termination result, not a polynomial-time guarantee.

3. A sharp separation between metric recovery and representation recovery (Propositions 6 and 7, Theorem 8, Corollary 9). Proposition 6 shows a direction is invisible to every fixed positive definite metric exactly when its feature derivative is input-independent. In the exact four-qubit Hubregtsen/local Z family, diagonal metrics give generic rank two, dense metrics reach but cannot exceed three, and no fixed affine expansion of any output dimension crosses three — while one appended nonlinear scalar coordinate (f_0f_2) with an explicit nondiagonal metric, or the scalar sum f_0f_2 + f_1f_3 with a positive diagonal metric, attains rank four.

4. Operational checks kept separate from identifiability. Exact rank certificates for 54 selected circuit/readout settings; matched finite-shot comparisons on previously frozen panels; a variance result for conditional product estimation under a factorized Z measurement distribution (Corollary 17); and execution audits that compare certified redundancies with compiled work after global normalization.

Main Findings

• Kernel equivalence is a Gram-matrix statement, not a state or unitary statement. Equality of G(θ) = CᵀMC characterizes equality of the kernel on all input pairs, and its generic differential rank gives the locally identifiable dimension. Linear independence of h alone is insufficient — constant modes must be separated into the offset u, since h(x) = (1, x)ᵀ with G = diag(1,1) versus G′ = diag(2,1) gives identical distances.

• State geometry does not imply kernel sensitivity. For |ψ_θ(x)⟩ = R_y(θ)R_y(x)|0⟩, the SLD-normalized QFI entry is F_θθ(x) = 1 while the fidelity kernel cos²[(x−x′)/2] is globally independent of θ. The general relation is a one-way inclusion, ker F_𝒳 ⊆ ker J_𝒳 and rank J_𝒳 ≤ rank F_𝒳, which is local and sample-specific. Switching to Z readout makes the squared distance d_Z² = [cos(θ+x) − cos(θ+x′)]², whose θ-derivative equals 2 at (θ, x, x′) = (0, 0, π/2) — sensitivity restored with no change in the prepared-state QFI.

• Full Pauli readout is not a rank maximizer. With all nonidentity n-qubit Pauli words at unit weight, d² = 2ⁿ Tr[(ρ − ρ′)²], equal to 2^{n+1}(1 − F) for pure states. But kernel differential rank is not monotone under adding observables: at two layers ChebyshevRx has r_Z = 6 > r_XYZ = 4, while one-layer ParamZ has r_Z = 0 < r_XYZ = 2. Adding rows to C adds positive semidefinite terms to G, which does not order Jacobian ranks.

• A fixed metric has a hard ceiling in the four-qubit family. For f_q(x) = −sin θ_q cos x_{q mod 2}, q = 0,…,3, diagonal metrics give generic rank two, and every dense metric gives rank at most three because the coefficient Gram on (cos x_0, cos x_1) is a symmetric 2×2 matrix with only three entries. The dense metric M = (32I + uuᵀ)/158 with u = (1,2,3,4)ᵀ is positive definite with unit trace and attains three, certified by the nonzero rational minor −835727837184/978271035584375 at a rational-phase point.

• One appended scalar suffices, and is minimal under the stated restrictions. With f_0f_2 appended and M_1/7 (entries: 2,1,0,0,0 / 1,2,0,0,0 / 0,0,1,0,0 / 0,0,0,1,0 / 0,0,0,0,1), the eigenvalues are 3,1,1,1,1; the exact minor is det D_(a,b,c,d) s = −8a²cd(2a² − c²), which at θ_0 = ⋯ = θ_3 = π/6 gives the angle Jacobian minor −9/128, i.e. −9/307328 for the unit-trace metric. Zero additions cannot exceed three, so one is both sufficient and necessary when the original four features are retained and the metric is fixed independently of inputs and trainable parameters. The authors stress this is not minimality over replacement representations, parameter-dependent metrics, or kernel families.

• Coordinate count and product count are different things. Any single pairwise product with a positive diagonal metric gives rank at most three; the single scalar f_0f_2 + f_1f_3 nonetheless attains rank four with a positive diagonal metric (Appendix T), and appending the two products separately is another sufficient construction. The products equal ZZ means in this family, so no additional ideal measurement information is required.

• Fixed affine expansions are capped. Corollary 9 shows any f̃ = Lf + b with arbitrary finite output dimension, fixed L, b and fixed M̃ ≻ 0 still gives generic rank at most three — so the recovery in Theorem 8 is genuinely nonlinear, not a relabeling.

• Metric design recovers a lot elsewhere. On ten previously studied source circuits from the YZ-CX, Hubregtsen and Kyriienko constructors, fixed diagonal local XYZ weights evenly spaced from 1 to 4 attain generic rank K, certified by exact Fourier Gram minors; the two- and four-qubit Kyriienko ranks change from 4 to 6 and from 10 to 12. These are redesigned kernels on existing circuits, not additions to the 54-setting census. A broader 30-setting metric survey is numerical only, and random or uniform-basis controls already attain many of the same ranks, so no superiority of a sophisticated selector is claimed.

• Minimal representation need not mean best sensitivity. Matched finite-shot comparisons on previously frozen panels, using common measurement outcomes and shot budgets, show that among the tested fixed designs the smallest sufficient representation does not have the best sensitivity. Under a factorized Z measurement distribution, conditional product estimation cannot increase variance at a fixed design (Corollary 17) — but equal basis count alone does not imply equal estimation precision.

• Identifiability is not cost savings. Execution audits compare certified redundancies against compiled work after global normalization, and the authors state plainly that rank recovery alone establishes neither predictive improvement nor execution savings.

Methodology in Plain English

The authors work at the level of features rather than states. Each feature is the ideal expectation value of a Pauli observable on the prepared state; the kernel is an ordinary Gaussian kernel applied to the metric-weighted difference of feature vectors. Because the Gaussian is injective with nonzero derivative, comparing kernels reduces to comparing the coefficient Gram matrix — the object that encodes all pairwise squared distances in the metric. That reduction converts "do these two parameter settings give the same kernel?" into finite-dimensional linear algebra and lets the authors count identifiable dimensions as a Jacobian rank.

They then construct an exact, hand-checkable counterexample family: a one-layer four-qubit Hubregtsen circuit whose local Z means are simply −sin θ_q cos x_{q mod 2}, with a diagonal controlled-rotation tail that does not touch those means. Within this family they can write the coefficient Gram in closed form for diagonal metrics, show it is a 2×2 symmetric matrix (hence rank ≤ 3 for any dense metric), and then test candidate lifts by computing exact rational Jacobian minors at rational parameter points. A nonzero rational minor is a proof, so the rank claims are certificates rather than numerical observations. The metric chosen for Theorem 8 is deliberately simple — a small integer matrix divided by seven to give unit trace — so the certificate can be checked by hand.

For the broader census they use exact Fourier-type Gram minors where the circuit grammar falls inside the decision subclass, and numerical surveys otherwise, flagging which is which. Finally, they run finite-shot simulations on frozen design panels to ask whether the extra identifiable directions are actually resolvable at realistic shot budgets, and separate audits of compiled circuit cost to ask whether redundant parameters translate into saved work.

Why This Matters

The paper draws a line the quantum-kernel literature tends to blur: a kernel's trainable dimension is a property of the readout-plus-metric pair, not of the circuit's parameter count or the state manifold's dimension. This has direct consequences for how people report and prune quantum models — a parameter that changes expectation values, and even one that changes the state, may be exactly invisible to the kernel being trained.

• Quantum kernel model design: choosing which Pauli observables to measure and what weights to assign, with an a priori bound on how much parameter dependence that choice can expose.
• Redundancy removal in variational circuits: distinguishing parameters that are dead for the task from those merely dead for the state, which matters for pruning claims that are otherwise overstated.
• Readout selection under hardware constraints: the one-qubit R_y example shows a readout change can restore kernel sensitivity without any change to the state-preparation circuit, which is relevant when the state circuit is fixed by compilation.
• Compiler and execution auditing: the paper's insistence that identifiability ≠ compiled savings is a caution for anyone marketing redundancy elimination as a runtime win.

Industry relevance: the results are directly usable by teams building kernel-based quantum classifiers or variational models on near-term hardware, and by compiler/tooling groups (the paper explicitly contrasts its targets with QuCLEAR-style Clifford absorption and Paulihedral-style scheduling and synthesis) that need to state precisely what their transformations preserve.

Future Directions

• Characterizing null directions that are not deletable coordinates. Corollary 4's bound comes with null directions that may depend on θ; whether these correspond to globally removable parameters, and how to find them constructively, is open.
• Efficient certification beyond the exact subclass. Corollary 3 is a termination result with potentially prohibitive minor enumeration; the ChebyshevRx and ParamZ examples are explicitly outside it and need separate methods. Practical algorithms for rank certification remain unaddressed.
• Optimizing metrics for finite-shot precision, not just rank. The authors show minimal representation size need not give the best sensitivity, and that their comparisons do not optimize over all metrics within each feature class — so a precision-aware metric design objective is a natural next step.
• Extending the minimality question. Theorem 8's minimality is restricted to retaining the original features with a fixed metric; minimal appended dimension over replacement representations, parameter-dependent metrics, or different kernel families is left open, as is the question of whether two scalar additions are ever necessary for a diagonal metric (the paper shows one scalar sum already suffices).

Target Audience

Quantum machine learning theorists and graduate students working on kernel methods, quantum feature maps, and expressivity/identifiability questions; researchers studying parameter redundancy in variational circuits who need rigorous certificates rather than sampled derivatives; and quantum compiler and tooling developers who want to know exactly which properties a circuit transformation does and does not preserve. The paper assumes comfort with rank arguments and Gram matrices — beginners will find the four-qubit worked example and Table 2 the most accessible entry points.

Note on figures: the paper does not report wall-clock speedups, dataset sizes, or predictive-accuracy gains from the proposed lifts; the 54-setting certificate census, the 30-setting numerical survey and the ten-circuit metric study are the only quantitative scope figures given.

Authors’ abstract

A circuit parameter can change measured features without changing their projected Gaussian kernel. We characterize this lost dependence and distinguish recovery by metric design from nonlinear feature expansion. With analytic feature coefficients, a fixed finite data basis affinely independent on the input domain, a fixed positive definite metric, and fixed bandwidth, the coefficient Gram matrix determines all-input kernel equivalence, global coordinate inactivity, and generic differential rank. An exact four-qubit family exhibits a sharp boundary: diagonal metrics give rank two, dense metrics reach but cannot exceed three, and no fixed affine expansion crosses this ceiling. One appended nonlinear scalar coordinate suffices for rank four and is minimal when the original features are retained. A single pairwise product achieves this with an explicit nondiagonal metric; a scalar sum of two products achieves it with a diagonal metric. Both use existing ideal local means. Exact certificates establish ranks in 54 selected circuit/readout settings. Matched finite-shot comparisons show that minimal representation size need not give the best sensitivity among the tested designs. Under a factorized measurement distribution, conditional product estimation cannot increase variance at a fixed design. Execution audits further separate kernel identifiability from compiled cost.