Quantum Mean Embeddings

• Quantum Mean Embeddings are generalizations of kernel mean embeddings that represent probability distributions as normalized quantum states or operator-valued measures.
• They formulate classical measures in infinite-dimensional Hilbert spaces and reproducing kernel Hilbert modules, enabling quantum overlap and inner product estimation.
• Applications include quantum tomography and anomaly detection with practical challenges in scalable state preparation and robust overlap estimation.

Quantum mean embeddings are quantum and operator-valued generalizations of kernel mean embeddings. In the current literature, the expression covers two closely related constructions. One represents a probability distribution as a normalized pure state in an infinite-dimensional Hilbert space so that inner products of classical mean embeddings become quantum overlaps (Kübler et al., 2019). Another generalizes classical scalar embeddings from probability measures to von Neumann-algebra-valued measures and Positive Operator-Valued Measures (POVMs) by replacing reproducing kernel Hilbert spaces (RKHSs) with reproducing kernel Hilbert modules (RKHMs), and, in a tensorized formulation, by integrating rank-one kernel operators against operator-valued measures (Hashimoto et al., 2020, Mayer et al., 24 May 2026). Both viewpoints preserve the role of kernel embeddings as linear representations of measures, but they differ in whether the target object is a unit vector, an element of a Hilbert module, or an operator on a tensor product space.

1. Classical kernel mean embeddings and the scope of the term

Classical kernel mean embedding starts from a measurable space equipped with a probability measure and a positive definite kernel. With feature map $\varphi(x)=k(\cdot,x)$ and RKHS $H_k$, the embedding is

$\mu_P=\int_X \varphi(x)\,dP(x),$

and the induced inner product satisfies

$$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$

The maximum mean discrepancy (MMD) is the RKHS norm of the difference of two embeddings, and for characteristic kernels it vanishes if and only if the underlying measures coincide. Universality refers to density of the RKHS in a target function space such as $C_0(X)$, and on many spaces characteristicness and universality are closely related or equivalent (Kübler et al., 2019).

Within this classical template, the quantum literature has developed several non-identical extensions.

Construction	Embedded object	Target space
Classical KME	Probability measure	RKHS
Pure-state QME	Probability distribution	Unit sphere of a Hilbert space
Operator-valued QME / QCE	POVM or von Neumann-algebra-valued measure	RKHM or $R(K)\otimes Q$

The pure-state construction retains scalar kernels but realizes the embedding as a quantum state. The operator-valued construction changes the measure itself: scalar measures are replaced by operator-valued measures, the inner product becomes operator-valued, and scalarization is introduced only when required. A further tensorized development, the Quantum Covariance Embedding (QCE), embeds POVMs through operators of the form $k_xk_x^*$ rather than through first-order feature vectors. This makes “quantum mean embedding” a family of closely allied constructions rather than a single formalism (Hashimoto et al., 2020, Mayer et al., 24 May 2026).

2. Pure-state realizations in infinite-dimensional Hilbert spaces

In the pure-state formulation, a quantum feature map assigns to each $x\in X$ a normalized state $|\varphi(x)\rangle$ in a Hilbert space $H$, inducing the kernel

$H_k$0

with $H_k$1 because of normalization. For a probability distribution $H_k$2, the quantum mean embedding is

$H_k$3

where

$H_k$4

The relation

$H_k$5

shows that classical kernel inner products can be recovered from quantum overlaps together with the known normalization factors. The empirical version is

$H_k$6

with

$H_k$7

This gives an explicit state-vector realization of an embedding that is otherwise handled only implicitly through the kernel trick (Kübler et al., 2019).

Injectivity is inherited from universality. If $H_k$8 is universal over $H_k$9, then $\mu_P=\int_X \varphi(x)\,dP(x),$0 is injective over Borel probability measures. The canonical continuous-variable realization uses coherent states,

$\mu_P=\int_X \varphi(x)\,dP(x),$1

for which

$\mu_P=\int_X \varphi(x)\,dP(x),$2

A bandwidth parameter $\mu_P=\int_X \varphi(x)\,dP(x),$3 is included by mapping $\mu_P=\int_X \varphi(x)\,dP(x),$4. Because the Gaussian kernel is universal, this coherent-state model yields an injective quantum mean embedding. The empirical QME is then a superposition of coherent states, described in the paper as a “cat state” (Kübler et al., 2019).

Inner products and MMD can be estimated quantumly through overlap estimation. For states $\mu_P=\int_X \varphi(x)\,dP(x),$5 and $\mu_P=\int_X \varphi(x)\,dP(x),$6, the SWAP test produces

$\mu_P=\int_X \varphi(x)\,dP(x),$7

When overlaps are nonnegative, as for Gaussian kernels, one recovers $\mu_P=\int_X \varphi(x)\,dP(x),$8. The paper further gives a linear-time normalization trick based on a reference state $\mu_P=\int_X \varphi(x)\,dP(x),$9 and discusses a conjectured reduction from classical $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$0 kernel summation to $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$1 time provided that state preparation costs $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$2 and overlap estimation is constant-time in $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$3. The same work emphasizes that scalable preparation of $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$4 and universal overlap estimation in continuous-variable systems remain open challenges (Kübler et al., 2019).

3. Von Neumann-algebra-valued measures and reproducing kernel Hilbert modules

A more structural notion of quantum mean embedding starts from a von Neumann algebra $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$5 acting on a Hilbert space, together with an $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$6-valued measure $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$7. Because $$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$8 is generally noncommutative, integration must distinguish left and right versions,

$$\langle \mu_P,\mu_Q\rangle_{H_k}=\iint k(x,y)\,dP(x)\,dQ(y).$$9

and the construction in the paper uses the right integral. An $C_0(X)$0-valued positive definite kernel is a map $C_0(X)$1 such that $C_0(X)$2 and

$C_0(X)$3

for finite sets $C_0(X)$4 and coefficients $C_0(X)$5. The associated feature map is $C_0(X)$6, and the RKHM $C_0(X)$7 is the completion of finite sums $C_0(X)$8 under the $C_0(X)$9-valued inner product

$R(K)\otimes Q$0

The reproducing property becomes

$R(K)\otimes Q$1

Here the inner product itself carries operator information instead of collapsing immediately to scalars (Hashimoto et al., 2020).

For a bounded $R(K)\otimes Q$2-kernel and a finite regular $R(K)\otimes Q$3-valued Borel measure, the embedding is

$R(K)\otimes Q$4

It is characterized by

$R(K)\otimes Q$5

and the inner product between two embeddings is

$R(K)\otimes Q$6

The module also carries an $R(K)\otimes Q$7-valued absolute value via $R(K)\otimes Q$8. To obtain scalar distances, one applies a faithful normal state $R(K)\otimes Q$9 on $k_xk_x^*$0,

$k_xk_x^*$1

which expands analogously to classical MMD. Without scalarization, the RKHM analogue of MMD over the unit ball is

$k_xk_x^*$2

This preserves operator structure and allows direct treatment of POVMs, density operators, and higher-order interactions among variables (Hashimoto et al., 2020).

4. Injectivity, universality, and discrepancy functionals

For the RKHM construction, injectivity is formulated by the condition

$k_xk_x^*$3

The paper gives sufficient conditions in the finite-dimensional operator case $k_xk_x^*$4 with $k_xk_x^*$5. If $k_xk_x^*$6 with a positive semidefinite $k_xk_x^*$7-valued measure $k_xk_x^*$8 satisfying $k_xk_x^*$9, then $x\in X$0 is injective. If

$x\in X$1

with positive definite $x\in X$2-valued measure $x\in X$3 and $x\in X$4, then $x\in X$5 is also injective. Concrete covered kernels include matrix-valued diagonal kernels whose diagonal entries are Gaussian, Laplacian, B-spline, or inverse multiquadratic kernels. In the same finite-dimensional setting, $x\in X$6 is injective if and only if $x\in X$7 is dense in $x\in X$8 in the sup-norm topology. For general von Neumann algebras, density of $x\in X$9 in $|\varphi(x)\rangle$0 is a sufficient condition for injectivity, while the converse remains open (Hashimoto et al., 2020).

The tensorized QCE line gives an operator-level discrepancy for POVMs. For a $|\varphi(x)\rangle$1-kernel $|\varphi(x)\rangle$2 with RKHS $|\varphi(x)\rangle$3 and POVM $|\varphi(x)\rangle$4, there exists a unique bounded, non-negative definite operator

$|\varphi(x)\rangle$5

characterized by

$|\varphi(x)\rangle$6

This induces the Quantum Maximum Discrepancy

$|\varphi(x)\rangle$7

If $|\varphi(x)\rangle$8 is characteristic for scalar measures on $|\varphi(x)\rangle$9, then it is characteristic for POVMs as well, and $H$0 becomes a metric on the space of POVMs. The conditional version for a fixed state $H$1 is

$H$2

and the operator-level discrepancy dominates the supremum of conditional discrepancies over states. In this framework, the classical first-order mean embedding is replaced by a covariance-type operator $H$3, reflecting the covariance-like structure of density operators (Mayer et al., 24 May 2026).

5. Quantum measurements, tomography, and higher-order structure

The operator-valued viewpoint is motivated by quantum mechanics and by multivariate dependence structure. A POVM $H$4 is an $H$5-valued measure, and for a density operator $H$6 the measure $H$7 defined by $H$8 captures outcome probabilities together with operator weights. A concrete finite-dimensional example uses

$H$9

together with the POVM

$H_k$00

for an orthonormal basis $H_k$01. Then, for density matrices $H_k$02,

$H_k$03

so a standard Hilbert–Schmidt similarity appears as a special case of the RKHM construction. The same paper also defines matrix-valued cross-covariance measures for multivariate random variables and shows, for separable kernels, that the RKHM distance recovers the Hilbert–Schmidt difference of matrices of cross-covariance operators. This keeps pairwise interaction structure explicit rather than collapsing all comparisons to scalars (Hashimoto et al., 2020).

In the QCE framework, the embedding is applied directly to tomography. Density estimation is reformulated as tensorized kernel regression, leading to the QUAntum Regression with Kernels (QUARK) estimator. The paper states that this accommodates the spectral geometry of physical implementations, derives a central limit theorem and concentration inequalities, and proves exactness of trace-preserving projections. It also develops a geometric design theory for quantum Gram superoperators in which unitary designs are strictly E-optimal experimental designs and are therefore statistically superior to Pauli observables. For structure-free estimation, it derives the exact minimax lower bound and proves that the tensorized estimators achieve the optimal rate. Under mutually unbiased bases, efficient estimation is obtained through the fast Walsh-Hadamard transform (Mayer et al., 24 May 2026).

These applications clarify a central distinction from scalar quantum kernel methods. In the RKHM and QCE approaches, the kernel and the embedding are operator-valued, so POVMs, density operators, and multivariate relations are handled within one noncommutative framework. Scalarization via a faithful normal state or via trace is optional and task-dependent rather than intrinsic to the definition (Hashimoto et al., 2020).

6. Empirical methods, computational prospects, and limitations

Empirical operator-valued embeddings are obtained from weighted Dirac approximations. Given pairs $H_k$04 with $H_k$05,

$H_k$06

and the empirical inner product is

$H_k$07

The same framework defines operator-valued two-sample statistics and kernel PCA in RKHM. On climate data, the $H_k$08-valued MMD using a diagonal Gaussian matrix kernel achieved better acceptance rates when the samples came from the same type and lower acceptance when they differed than classical scalar MMDs. For quantum anomaly detection with density matrices containing phase or amplitude errors, RKHM kernel PCA achieved higher AUCs across eight error types than prior methods based on Hilbert–Schmidt inner products (Hashimoto et al., 2020).

In the pure-state setting, the main computational prospect is that overlap estimation depends on the size of the prepared quantum states rather than on explicit $H_k$09 kernel summation. The cited proposal shows how inner products and empirical MMD can be reduced to state preparation, overlap estimation, and normalization recovery; it also emphasizes that scalable preparation of superpositions of many nonorthogonal states, universal SWAP-test implementations in continuous-variable systems, and robustness to decoherence, photon loss, detector inefficiencies, and phase noise are unresolved practical constraints (Kübler et al., 2019).

Related developments broaden the surrounding algorithmic landscape. “Q-means using variational quantum feature embedding” constructs characteristic cluster quantum states

$H_k$10

and minimizes inter-cluster overlap, equivalently maximizing Hilbert–Schmidt distance between pure cluster representatives (Menon et al., 2021). “Entaglement-Based Quantum Mean Estimator Circuit” proposes a QRAM-based circuit for estimating empirical means with stated complexity $H_k$11 (Tamirat, 2019). “Near-Optimal Quantum Algorithms for Multivariate Mean Estimation” gives covariance-sensitive Euclidean guarantees in a binary oracle model, proves matching lower bounds up to logarithmic factors, and shows that the phase-oracle model is strictly weaker for mean estimation (Cornelissen et al., 2021). These works address coherent cluster representatives and mean-estimation primitives rather than the RKHM or QCE definitions themselves, but they illuminate the computational subroutines and access models that quantum mean-embedding methods may rely on.

Across the literature, the principal limitations are consistent. The operator-valued line requires bounded $H_k$12-kernels, finite regularity assumptions, and care with left and right integrals in noncommutative settings; injectivity-equivalence with universality is fully established only in finite-dimensional operator algebras (Hashimoto et al., 2020). The tensorized tomography line is developed in finite-dimensional quantum systems and requires additional structure to extend to infinite dimensions (Mayer et al., 24 May 2026). The pure-state line depends on efficient state preparation and overlap estimation that have not yet been realized at scale in continuous-variable hardware (Kübler et al., 2019). Quantum mean embeddings therefore form a mathematically coherent family of constructions with strong representational and metric properties, but their algorithmic advantages remain conditional on demanding physical and computational assumptions.