Quantum Singular Value Transformations

• Quantum Singular Value Transformations are a framework that implements polynomial approximations of singular values in block-encoded matrices.
• They employ block-encoding and quantum signal processing techniques to achieve near-optimal resource scaling in quantum algorithms.
• QSVT underpins efficient approaches for Hamiltonian simulation, matrix inversion, and quantum machine learning with significant speedups.

Quantum Singular Value Transformations (QSVT) constitute a powerful and unifying algebraic and algorithmic framework for implementing polynomial transformations of the singular values (or, in the Hermitian case, eigenvalues) of linear operators block-encoded into larger unitaries. QSVT and its extensions provide the formal backbone for many quantum algorithms achieving exponential or polynomial speedups in quantum simulation, linear-systems solving, optimization, and quantum signal processing, anchoring efficient quantum procedures for matrix inversion, Hamiltonian simulation, eigenvalue discrimination, amplitude amplification, quantum machine learning, and beyond. The core utility of QSVT lies in its capability to implement block-encodings of matrix functions—$P(A)$ for a chosen polynomial $P$—with optimal or near-optimal resource scaling, leveraging ancilla-efficient circuits that generalize quantum signal processing (QSP) and catalyze systematic algorithmic improvements in both fault-tolerant and near-term quantum computing regimes.

1. Foundational Framework: Block-Encoding and QSVT Circuits

At the heart of QSVT is the block-encoding paradigm: for a matrix $A$ with $\|A\| \leq \alpha$, an $(\alpha, a, \epsilon)$ block-encoding is a unitary $U$ on $n + a$ qubits such that

$$(\langle 0^a| \otimes I) U (|0^a\rangle \otimes I) = A / \alpha \pm \epsilon.$$

Given such a block-encoding and a real polynomial $P$ of degree $d$ with $P$0 on $P$1, the QSVT implements a block-encoding of $P$2 via an alternating sequence of $P$3, $P$4, and single-qubit phase rotations, using $P$5 instances of $P$6 or $P$7 and $P$8 single-qubit $P$9 rotations (Gilyén et al., 2018, Tang et al., 2023). The construction naturally lifts the single-qubit QSP protocol—applying $A$0 to SU(2) eigenvalues— to arbitrary block-encoded matrices via the Cosine–Sine Decomposition, decoupling the matrix action into invariant two-dimensional subspaces corresponding to the singular vectors of $A$1.

2. Polynomial Transformations, Circuit Synthesis, and Complexity Bounds

The QSVT protocol enables the implementation of $A$2 for a block-encoding $A$3 of $A$4, with $A$5 synthesized as a Chebyshev expansion or via Remez minimax optimization. Circuit depth is $A$6 for degree-$A$7 $A$8, with total gate complexity dominated by the block-encoding cost and the sequence of single- and two-qubit operations (Gilyén et al., 2018, Tang et al., 2023). For common analytic matrix functions:

• Hamiltonian simulation: $A$9, degree $\|A\| \leq \alpha$0.
• Matrix inversion: $P(x)