Projected Hybrid Kernel

• Projected hybrid kernel is a hybrid method that fuses classical feature maps with a finite set of quantum observables, ensuring positive semidefiniteness and a bounded representation.
• It avoids the dimensionality blow-up of full quantum state overlaps by projecting onto selected Hermitian observables and controlling the effective rank.
• The approach enables efficient learning with explicit PAC bounds and compresses high-order interactions, making it suitable for scalable quantum-classical hybrid models.

The projected hybrid kernel is a kernel construction for hybrid quantum–classical learning in which a classical feature space is augmented by a finite-dimensional quantum-projected readout rather than by a full quantum state overlap. In the formulation introduced for active quantum subspace data-encoding, a classical input $x$ is mapped both to a classical feature vector and to a quantum state $\rho_x$, after which only the expectation values of a fixed set of bounded Hermitian observables are retained as quantum features. The resulting kernel has the additive form $$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$, where $K_C$ is a classical kernel, $K_Q^{\rm proj}$ is the projected quantum kernel, and $\lambda \ge 0$ controls the quantum contribution. This construction is designed to preserve positive semidefiniteness while avoiding the dimension blow-up associated with naive global quantum kernels (Bang et al., 30 May 2026).

1. Formal construction

The projected hybrid kernel is defined by combining a classical feature map with a projected quantum feature map. The classical component is

$$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$

The quantum component begins with a data-encoding circuit $E(x)$ on $\kappa$ qubits, producing

$$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$

A finite collection of bounded Hermitian observables $\rho_x$0, each satisfying $\rho_x$1, is then fixed. The projected quantum feature map is

$\rho_x$2

and the projected quantum kernel is

$\rho_x$3

The hybrid feature map is the direct sum

$\rho_x$4

which induces

$\rho_x$5

This definition makes the role of projection explicit: the quantum sector is not represented by a full-state similarity, but by a controlled list of expectation values. A plausible implication is that the kernel inherits the inductive bias of the chosen observables $\rho_x$6, rather than that of the entire encoded Hilbert-space state (Bang et al., 30 May 2026).

2. Positive semidefiniteness and dimension control

A central structural property is that the projected hybrid kernel is positive semidefinite by construction. On a sample $\rho_x$7, let $\rho_x$8 be the $\rho_x$9 matrix with entries

$$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$0

Then

$$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$1

Thus the hybrid kernel remains a valid Mercer kernel whenever the classical component is itself psd.

The same finite-observable structure yields a rank bound: $$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$2 This is paired with a bound on the finite-sample regularized dimension

$$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$3

In the cited analysis, this is presented as a mechanism by which the “dimension blow-up of naive global kernels is avoided” (Bang et al., 30 May 2026).

These bounds distinguish the construction from kernels based on unrestricted state overlaps. The projected hybrid kernel enlarges the effective hypothesis class only by the number $$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$4 of projected observables, not by the full dimension of the quantum Hilbert space. For hybrid-learning design, this means that representational enrichment is explicitly budgeted.

3. Residual-information criterion for hybrid benefit

The projected hybrid kernel does not guarantee an improvement over a purely classical predictor merely by adding quantum features. The relevant question is whether the projected quantum sector contributes information that is simultaneously outside the classical span and aligned with the classical residual.

Let $$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$5 denote the closed span of the classical features, $$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$6 the finite span of the quantum features, and $$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$7. Let $$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$8 be the orthogonal projection of the target $$K_H(x,x') = K_C(x,x') + \lambda K_Q^{\rm proj}(x,x')$$9 onto $K_C$0, and $K_C$1 the projection onto $K_C$2. Define the classical residual

$K_C$3

Then the optimal hybrid risk satisfies

$K_C$4

More sharply, for any $K_C$5, with

$K_C$6

one has

$K_C$7

Strict improvement occurs if and only if there exists $K_C$8 such that

$K_C$9

This criterion rules out a common overgeneralization: quantum augmentation is not beneficial merely because the augmented model is larger. The projected quantum features must survive orthogonalization against the classical span and must correlate with what the classical predictor leaves unexplained. In this sense, the projected hybrid kernel formalizes a residual-information test for genuine hybrid benefit (Bang et al., 30 May 2026).

4. Statistical learnability in the noisy-oracle setting

The same framework provides a PAC-style learnability analysis under noisy labels. The setting assumes a target $K_Q^{\rm proj}$0 and observed labels $K_Q^{\rm proj}$1 satisfying

$K_Q^{\rm proj}$2

If $K_Q^{\rm proj}$3 minimizes the empirical noisy error $K_Q^{\rm proj}$4, then three statements are established.

First, for every $K_Q^{\rm proj}$5,

$K_Q^{\rm proj}$6

where $K_Q^{\rm proj}$7.

Second, on any sample where $K_Q^{\rm proj}$8,

$K_Q^{\rm proj}$9

Third, uniform-convergence arguments imply that if

$\lambda \ge 0$0

where $\lambda \ge 0$1, then $\lambda \ge 0$2 with probability $\lambda \ge 0$3. In particular,

$\lambda \ge 0$4

A corollary given in the same analysis states that if the hybrid model has feature dimension $\lambda \ge 0$5 and worst-case oracle reliability $\lambda \ge 0$6, then it is PAC-learnable in polynomially many samples. This links the finite-observable design of the projected hybrid kernel to explicit sample-complexity control: bounded representation size and inverse-polynomial reliability are sufficient for polynomial learnability (Bang et al., 30 May 2026).

5. Active quantum subspaces, noise, and explicit feature compression

The projected hybrid kernel is presented in conjunction with active quantum subspace data-encoding, where only an information-bearing subset of variables is lifted to the quantum representation and the remaining variables remain classical. In the canonical Clifford example, an active subset $\lambda \ge 0$7 of size $\lambda \ge 0$8 is selected. For $\lambda \ge 0$9, the preparation is

$$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$0

while for $$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$1 one prepares $$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$2 in the computational basis. A Clifford circuit $$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$3 is applied to all $$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$4 qubits, followed by measurement of a Pauli string $$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$5. Equivalently, one studies

$$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$6

If

$$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$7

with $$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$8 and $$\Phi_C:\;x\mapsto \phi_C(x)\in\mathbb R^{D_C}, \qquad K_C(x,x')=\langle \phi_C(x),\phi_C(x')\rangle_{\mathbb R^{D_C}}.$$9, then the ideal expectation factorizes as

$E(x)$0

The score vanishes if $E(x)$1 contains any $E(x)$2 on an active qubit or any $E(x)$3 on a context qubit.

Under local dephasing noise after each Clifford gate $E(x)$4,

$E(x)$5

the Heisenberg action is

$E(x)$6

If $E(x)$7 denotes the total number of relevant dephasing sites, then

$E(x)$8

If $E(x)$9 and $\kappa$0, then $\kappa$1; more generally, if $\kappa$2, then $\kappa$3. The oracle reliability becomes

$\kappa$4

so if $\kappa$5 and $\kappa$6, then $\kappa$7 (Bang et al., 30 May 2026).

An explicit sixty-four-qubit example illustrates high-order-interaction compression. With only qubits $\kappa$8 carrying phases, qubits $\kappa$9 carrying context bits, and the remaining $$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$0 qubits prepared but unused in the readout, one chooses $$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$1 so that

$$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$2

The single projected feature is then

$$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$3

Under the stated product distribution, this monomial is orthogonal to every “interaction-only” polynomial of total degree $$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$4 in the raw variables $$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$5. Hence no classical model built from such degree-$$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$6 polynomials can capture it, whereas a single quantum-projected observable can (Bang et al., 30 May 2026).

6. Relation to projected quantum kernels and multiple-kernel methods

The projected hybrid kernel is closely related to earlier projected quantum-kernel constructions, but it is not identical to them. In projected quantum kernels based on reduced density operators, one starts from

$$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$7

and defines kernels such as

$$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$8

or

$$\rho_x=E(x)\,\bigl|0^{\otimes\kappa}\bigr\rangle\!\bigl\langle 0^{\otimes\kappa}\bigr|\,E(x)^\dagger.$$9

For alternating layered ansatzes, the variance of a projected term $\rho_x$00 depends on circuit depth, local block size, and initial-state entanglement. Under global random-circuit assumptions, the variance remains exponentially small in $\rho_x$01; under shallow alternating layered ansatzes with product initial states, one obtains $\rho_x$02, while highly entangled initial states recover vanishing behavior. Edge positions decay more slowly than middle positions, with $\rho_x$03 versus $\rho_x$04 bounds (Suzuki et al., 2023). This places the projected hybrid kernel within a broader effort to avoid vanishing similarity by using local or projected quantum information rather than global fidelity.

A second neighboring line of work is quantum multiple kernel learning. There, one forms a dictionary of projected quantum kernels, fidelity quantum kernels, and classical kernels, and uses a nonnegative linear combination

$\rho_x$05

typically with $\rho_x$06. In the projected-kernel component of that framework, the one-body reduced states $\rho_x$07 are estimated from single-qubit Pauli expectations, and the kernel is

$\rho_x$08

This approach requires $\rho_x$09 circuit executions for an $\rho_x$10-point training set, in contrast to $\rho_x$11 for fidelity kernels, and the circuit depth is approximately halved because no compute–uncompute step is required. In the reported experiments, this depth reduction was described as the key to scaling from approximately $\rho_x$12 to up to $\rho_x$13 qubits on hardware (Miyabe et al., 2023).

The distinction is therefore structural. In multiple-kernel learning, “hybrid” refers to a weighted combination of separate kernels, potentially classical and quantum. In the projected hybrid kernel, by contrast, the hybridization occurs at the feature-map level through the direct sum $\rho_x$14. The two perspectives are compatible, but they address different design questions: one concerns kernel composition, the other concerns how much quantum information is injected into a single composite representation.

7. Conceptual significance and common misunderstandings

The projected hybrid kernel occupies a specific position in the landscape of quantum-enhanced kernel methods. It is neither a full-state fidelity kernel nor merely a classical kernel with quantum preprocessing. Its defining feature is that the quantum sector is explicitly projected onto a bounded set of observables before kernelization. This makes the kernel psd, rank-controlled, and sample-dimension controlled by construction (Bang et al., 30 May 2026).

One misconception is that hybrid quantum advantage in kernel methods requires full quantum data-encoding into a highly superposed state. The active-subspace results were introduced precisely to test whether advantage can persist without such full encoding, and the stated conclusion is that scalable hybrid advantage can be obtained “without full quantum data-encoding” when the projected quantum sector contributes residual information not present in the classical span (Bang et al., 30 May 2026). Another misconception is that projected methods automatically evade trainability issues. The projected-kernel analysis for alternating layered ansatzes shows instead that trainability depends on circuit depth, block size, subsystem position, and especially the entanglement of the initial state (Suzuki et al., 2023).

Taken together, these results present the projected hybrid kernel as a controlled hybridization mechanism: it restricts the quantum contribution to a finite, observable-defined subspace; it admits explicit rank and sample-complexity bounds; and it can, in structured settings, compress high-order interactions into a low-dimensional augmentation of a classical model. A plausible implication is that its principal use is not maximal quantum expressivity, but selective quantum feature injection under representation, trainability, and hardware constraints.