Generalized Quantum Singular Value Transformation

• GQSVT is an extension of standard QSVT that generalizes polynomial transforms to include arbitrary ancilla dimensions, matrix-valued outputs, and non-Hermitian operator treatments.
• It employs techniques like projected unitary encodings, Hermitianization, and multi-output generalizations to overcome standard parity and embedding limitations.
• The framework underpins advanced quantum algorithms for Hamiltonian simulation, phase estimation, and iterative linear solvers while addressing trade-offs and open research challenges.

Generalized quantum singular value transformation (GQSVT) denotes a family of extensions of quantum singular value transformation in which the original single-qubit, scalar-polynomial, finite-dimensional framework is enlarged along one or more axes. Depending on the formulation, the generalization may remove parity and real-coefficient restrictions, allow vector- or matrix-valued polynomial outputs, replace a single control qubit by an $N$-dimensional ancilla, move from standard block-encodings to projected unitary encodings, Hamiltonian embeddings, or Lie-group signal models, or target non-Hermitian superoperators such as quantum channels. Across these formulations, the common objective is to compile a coherent sequence whose designated block implements a prescribed transformation of singular values, typically by reducing the dynamics to invariant two-dimensional signal subspaces or by completing polynomial matrices to unitary objects (Sünderhauf, 2023, Lu et al., 2024, Laneve, 2023).

1. Standard QSVT as the baseline

The standard point of departure is the projected-unitary or block-encoding picture. For a matrix $A$, one embeds $A/\alpha$ into a unitary $U$ through projectors $\Pi_{\mathrm{in}}$ and $\Pi_{\mathrm{out}}$ so that

$$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$

If

$A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$

then the singular-value transform of a scalar function $f$ is

$$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$

Standard QSVT realizes bounded polynomial approximants to $A$0 by interleaving uses of $A$1 and $A$2 with phase-modulated reflections. The construction acts within invariant two-dimensional subspaces attached to singular pairs $A$3, so odd polynomials appear in in-to-out blocks and even polynomials in in-to-in or out-to-out blocks. The query complexity is linear in the polynomial degree, and the operator error is bounded by the uniform approximation error of the chosen polynomial on the singular-value domain (Gilyén et al., 2018).

This baseline already contains the essential algebraic constraints that later formulations modify rather than remove. In the standard setting, implementable polynomials must be bounded on $A$4 and obey parity conditions induced by the alternating sequence. Generalized formulations are therefore best understood as changing the admissible polynomial class, the signal register, or the embedding model, while retaining unitary consistency.

2. GQSP-derived extensions for arbitrary and Hermitian matrices

One major branch of GQSVT begins with generalized quantum signal processing (GQSP). In this formulation, the ancilla processing gates are arbitrary $A$5 rotations rather than the restricted single-axis rotations of standard QSP, and the signal operator is taken in a particularly simple controlled form,

$A$6

The main theorem states that for degree-$A$7 polynomials $A$8, there exists a sequence of ancilla rotations producing a unitary whose upper-left block is $A$9 if and only if $A/\alpha$0 on the unit circle. A complementary theorem shows that whenever $A/\alpha$1 on the unit circle, such a completion polynomial $A/\alpha$2 exists. In this sense, the practical parity and interlacing restrictions of standard QSP are replaced by a sole unitary norm bound on the circle. The same framework includes negative powers by mixing the signal with a complementary operator involving $A/\alpha$3, and it provides an explicit recursive angle synthesis when $A/\alpha$4 and $A/\alpha$5 are known. When only $A/\alpha$6 is known, the paper gives an FFT-based optimization based on the convolution identity for $A/\alpha$7, reporting random instances up to degree $A/\alpha$8 solved in under $A/\alpha$9 seconds on an NVIDIA A100 GPU (Motlagh et al., 2023).

The matrix generalization of this idea yields what one paper explicitly names generalized quantum singular value transformation together with generalized quantum eigenvalue transformation. Here the starting point is a projected unitary encoding

$U$0

with two principal constructions. The first is Hermitianization,

$U$1

which turns singular-value problems for arbitrary $U$2 into eigenvalue problems for a Hermitian dilation. The second is multiplication, which works with a block-encoding of $U$3. In the Hermitian case, generalized quantum eigenvalue transformation allows complex coefficients and indefinite parity. For arbitrary matrices, the Hermitianization-based GQSVT simultaneously produces odd and even singular-value transforms in different blocks, while the multiplication-based variant reduces the degree by passing to functions of $U$4 and then, in the odd case, left-multiplying by $U$5. A distinctive feature of this line is the need to control not just the Chebyshev polynomial $U$6 on $U$7 but also the monomial polynomial $U$8 on the unit circle. The resulting downscaling factor

$U$9

is proved to satisfy $\Pi_{\mathrm{in}}$0 in the worst case. The same work reports $\Pi_{\mathrm{in}}$1 for matrix inversion polynomials up to degree $\Pi_{\mathrm{in}}$2, while emphasizing that Hermitianization can roughly double query complexity relative to standard QSVT, whereas the multiplication route recovers a comparable overall query count with one extra qubit and a measure-early variant that can avoid even that extra qubit when the product is immediately measured (Sünderhauf, 2023).

3. Multi-output and matrix-valued generalizations on $\Pi_{\mathrm{in}}$3 and $\Pi_{\mathrm{in}}$4

A second branch of GQSVT replaces scalar outputs by matrix- or vector-valued polynomial transforms. In the $\Pi_{\mathrm{in}}$5 formulation, the ancilla is $\Pi_{\mathrm{in}}$6-dimensional, and the circuit implements an entire polynomial matrix $\Pi_{\mathrm{in}}$7 rather than a single scalar polynomial. The forward theorem states that a sequence with $\Pi_{\mathrm{in}}$8 calls to the signal unitary implements a block-encoding of a matrix of complex polynomials of degree at most $\Pi_{\mathrm{in}}$9. The backward theorem gives a necessary-and-sufficient style condition: if a polynomial matrix $\Pi_{\mathrm{out}}$0 of degree at most $\Pi_{\mathrm{out}}$1 has singular values in $\Pi_{\mathrm{out}}$2 for $\Pi_{\mathrm{out}}$3, then it can be block-encoded with $\Pi_{\mathrm{out}}$4 controlled-$\Pi_{\mathrm{out}}$5 calls. For singular-value transformation, the admissibility condition becomes

$\Pi_{\mathrm{out}}$6

and the output is a block-encoding of the matrix-valued transform $\Pi_{\mathrm{out}}$7. The construction uses polynomial matrix spectral factorization and a recursive projector-selection rule based on the column space of the top-degree coefficient. This same framework supports a bi-variate Laurent-polynomial construction $\Pi_{\mathrm{out}}$8 under separable vector-norm constraints and gives an amplitude-estimation formulation in which every valid degree-$\Pi_{\mathrm{out}}$9 outcome distribution is polynomial in the amplitude. In that application, the paper states Heisenberg scaling $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$0, a lower bound $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$1, and a matching achievable constant $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$2 using a $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$3-QSVT realization of the QPE sine-state distribution (Lu et al., 2024).

The $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$4 formulation takes a different route. Here $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$5 control qubits give an ancilla of dimension $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$6, and the signal operator is

$$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$7

Interleaving $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$8 with $$\Pi_{\mathrm{out}} U \Pi_{\mathrm{in}}=\frac{A}{\alpha}.$$9 synthesizes a vector polynomial

$A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$0

subject to the normalization

$A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$1

The main existence theorem states that any such normalized vector of degree at most $A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$2 can be realized. In the singular-value interpretation, the effective operator satisfies

$A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$3

so measuring the ancilla produces multiple bits of information in a single run via the distribution $A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$4. An exponential-steps variant replaces $A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$5 by

$A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$6

implemented through controlled powers $A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$7. Under the explicit assumption that these powers are easily implementable, one step can reach degree $A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$8. This viewpoint recovers phase estimation as a one-step $A=\sum_k \sigma_k\, |u_k\rangle\langle v_k|,$9-QSP protocol with $f$0, and it rederives Shor’s discrete-logarithm algorithm from the same perspective (Laneve, 2023).

Formulation	Output object	Core admissibility condition
$f$1-QSVT	Polynomial matrix $f$2	$f$3 on $f$4
$f$5-QSP/QSVT	Vector-valued amplitudes $f$6	$f$7 on the signal domain

These two families generalize different aspects of standard QSVT. The $f$8 picture is a matrix-valued extension of scalar polynomial synthesis, whereas the $f$9 picture is an ancilla-amplitude extension in which multiple outputs are read out directly from the control register.

4. Operator-theoretic, Hamiltonian, and continuous-variable formulations

Another line of work reframes QSVT at the level of bounded operators on general Hilbert spaces. Instead of a finite-dimensional top-left block, one starts with closed subspaces $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$0, orthogonal projections $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$1, and a unitary $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$2 on a possibly infinite-dimensional Hilbert space, and defines the embedded operator by compression,

$$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$3

In this setting the singular-value transform is basis-independent: the polynomial acts on the operator itself through the functional calculus of $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$4 and $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$5, not through a chosen matrix representation. The alternating products of $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$6 and subspace rotations $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$7, $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$8 again reduce to $$f^{\mathrm{SV}}(A)=\sum_k f(\sigma_k)\,|u_k\rangle\langle v_k|.$$9 signal blocks on singular subspaces, yielding even transforms as polynomials in $A$00 and odd transforms as $A$01. The formulation is valid for arbitrary bounded operators embedded in a unitary on possibly infinite-dimensional Hilbert spaces, and the paper emphasizes that the transformation applies to the whole operator rather than to a basis-dependent block picture (Kiumi et al., 2024).

A continuous-time generalization replaces unitary block-encodings by Hamiltonian embeddings. With an off-diagonal Hamiltonian

$A$02

and an auxiliary Hamiltonian

$A$03

alternating evolutions under $A$04 and $A$05 generate a Hamiltonian singular value transformation. In each singular subspace, $A$06 acts as $A$07, while $A$08 rotates the effective generator in the $A$09-plane. The synthesized unitary has the form

$A$10

provided $A$11. The special case $A$12 gives an inverse block encoding, namely a unitary whose off-diagonal block is $A$13. The paper interprets the resulting continuous-time circuit as an instance of generalized QAOA and states a segment count scaling $A$14 for analytic targets (Lloyd et al., 2021).

A further generalization leaves compact groups entirely and moves to $A$15, motivated by continuous-variable systems. There the signal variable is hyperbolic,

$A$16

and the QSP-like product has the functional form

$A$17

Because $A$18 is non-compact, the relevant invariant two-dimensional objects live in operator space rather than finite-dimensional state space, and the “singular values” are hyperbolic gains rather than ordinary singular values in $A$19. Controlled squeezing then plays the role of a block-encoding analogue. The paper establishes density of achievable transforms on compact $A$20-intervals but also stresses non-Riesz-basis behavior and potentially exponential growth of protocol length with interval size (Rossi et al., 2023).

5. Phase synthesis, algorithmic realization, and applications

The practical usefulness of GQSVT depends heavily on phase or parameter synthesis. In the single-qubit $A$21 setting, a 2025 result proves that GQSP is exactly equivalent to the nonlinear Fourier transform over $A$22. The analytic product

$A$23

is identified with an $A$24 transfer matrix $A$25, and the inverse nonlinear Fourier transform yields canonical GQSP phases. The same paper gives a Riemann-Hilbert-Weiss algorithm, accelerated by a Half-Cholesky procedure, with complexity

$A$26

and proves canonical uniqueness of phase factors under the condition

$A$27

This provides a unified and provably stable route for computing phase factors in QSP variants covered by the single-qubit $A$28 formalism (Laneve, 4 Mar 2025).

Applications already span several algorithmic domains. In the GQSP-derived framework, one paper uses generalized synthesis to simplify Hamiltonian simulation, derives an $A$29-approximate fractional query algorithm with query complexity

$A$30

and gives a normal-matrix and convolution framework in which a length-$A$31 filter can be implemented in

$A$32

one- and two-qubit gates (Motlagh et al., 2023). In the $A$33 and $A$34 multi-output settings, the examples include amplitude estimation with optimal asymptotic query complexity, phase location by one-shot multi-bit ancilla readout, phase estimation as a one-step protocol, and discrete logarithm via tensorized phase-estimation polynomials (Lu et al., 2024, Laneve, 2023).

The framework has also been extended to objects for which no explicit matrix block-encoding is available. For an unknown quantum channel $A$35, one 2025 paper constructs an approximate block-encoding of a Hermitized Liouville representation from black-box access to $A$36 alone. The abstract states an $A$37 upper bound and an $A$38 lower bound for producing a channel that is $A$39-close in diamond norm to a block-encoding of the Hermitized Liouville operator, after which standard QSVT applies polynomial transformations to its singular values. The same construction is used to estimate $A$40-th singular value moments for arbitrary real $A$41, with an application to testing whether a channel is entanglement breaking (Niwa et al., 30 Jun 2025).

A more directly linear-algebraic application is a quantum bi-conjugate gradient method. There GQSVT is used to realize the polynomial recurrences underlying BiCG for non-Hermitian systems, while swap tests estimate the inner products needed for the BiCG coefficients. The paper emphasizes relatively shallow circuit depth and a small number of ancilla qubits, and it gives the maximum circuit depth in terms of the convergence rate parameters $A$42 and condition numbers as

$A$43

That result illustrates a recurring theme in the GQSVT literature: once the admissible polynomial class is enlarged, Krylov-style iterative methods become more natural to encode because mixed-parity transforms no longer need to be split across separate QSVT gadgets (Liu et al., 29 Aug 2025).

6. Admissibility conditions, trade-offs, and open questions

A persistent misconception is that “generalized” means that the unitary constraints of QSVT disappear. They do not. What changes is the form of the admissibility condition. In scalar GQSP-derived formalisms the key condition is $A$44 on the unit circle, or equivalently $A$45 after completion. In $A$46-QSVT the condition becomes the matrix positivity constraint $A$47. In $A$48 vector-valued constructions it is the normalization $A$49. In the $A$50 continuous-variable setting it becomes the hyperbolic relation $A$51. This suggests that GQSVT is not a single theorem with a universal feasibility test, but rather a family of signal-processing frameworks sharing a common singular-value objective.

The trade-offs are correspondingly framework-dependent. Hermitianization-based GQSVT increases expressivity but may require downscaling relative to standard QSVT and can roughly double the number of oracle calls, although its faster phase-factor computation and direct treatment of complex or mixed-parity polynomials can offset that cost in concrete tasks (Sünderhauf, 2023). Multi-qubit $A$52 schemes enable multi-bit readout and, when powers $A$53 are cheaply available, exponential degree per step, but efficient synthesis of general $A$54 layers is only guaranteed when the ancilla dimension is polynomially bounded, and the paper explicitly notes that a complete multi-qubit Chebyshev/QSVT theorem remains open (Laneve, 2023). In the black-box channel setting, the current upper and lower bounds still leave a gap in dimension dependence, so direct two-sided block-encodings of unknown non-Hermitian superoperators remain an open structural question (Niwa et al., 30 Jun 2025). In non-compact Lie-group extensions such as $A$55, approximation theory is fundamentally less benign: closure and density survive on compact intervals, but uniform approximation can be badly conditioned and sequence lengths may grow exponentially with interval size (Rossi et al., 2023).

Across the literature, the most stable interpretation is therefore methodological rather than terminological. GQSVT refers to the extension of QSVT beyond the original scalar, single-qubit, parity-constrained setting, but the extension can target different bottlenecks: polynomial expressivity, output multiplicity, signal model, embedding architecture, or the class of operators to which singular-value methods apply. The field’s current structure is plural rather than fully standardized, and that plurality is itself one of the defining features of the subject.