Quantum Singular Value Transformation (QSVT)

A grand unification: apply any suitable polynomial to the singular values of a block-encoded matrix, recovering most quantum algorithms as special cases.

The general statement

Given a block-encoding of a matrix A (not necessarily Hermitian or square) with singular value decomposition A = sum_i sigma_i |u_i>QSVT (Gilyen, Su, Low, Wiebe, 2019) produces a block-encoding of P(A), where P is applied to each singular value: P^(SV)(A) = sum_i P(sigma_i) |u_i>

Hermitian case: eigenvalue transformation

When A is Hermitian, singular values relate to eigenvalues and QSVT applies P directly to the spectrum: it produces a block-encoding of P(H/alpha). This is exactly the operator-level version of quantum signal processing and is what qubitization uses for time evolution.

Algorithms recovered as special cases

• Hamiltonian simulation: P approximates e^(-i alpha t x).
• Quantum linear systems (HHL improved): P approximates 1/x.
• Amplitude amplification and Grover search: low-degree odd polynomials.
• Ground-state preparation and eigenstate filtering: P approximates a step or window function.
• Fixed-point search and phase estimation refinements.

Why it is a big deal

Before QSVT these algorithms had separate analyses and constructions. QSVT shows they are all the same primitive: choose a polynomial, bound its degree, apply it to a block-encoded operator. Designing an algorithm reduces to a classical approximation-theory problem, find a low-degree polynomial that approximates your target function on the relevant interval.

Costs

The circuit uses one call to the block-encoding (or its inverse) per polynomial degree, plus one ancilla for the phase rotations, on top of the ancillas the block-encoding itself needs. Numerically stable methods now compute the QSVT phase angles for high-degree polynomials, which had been an early bottleneck. QSVT is the modern lens through which resource estimates for simulation and chemistry are derived.