---
title: Hamiltonian Singular Value Transformation
url: https://www.emergentmind.com/topics/hamiltonian-singular-value-transformation
type: topic
---

# Hamiltonian Singular Value Transformation

Hamiltonian singular value transformation (HSVT) is the extension of quantum singular value transformation (QSVT) from the setting of a matrix embedded as a block of a unitary to the setting of a matrix embedded as a block of a Hamiltonian. In this formulation, polynomial transformations of the singular values are implemented by alternating the application of accessible Hamiltonians for chosen intervals, in a purely Hamiltonian context; the same framework also yields inverse block encoding and leads to procedures for matrix multiplication and for solving differential equations on quantum information processors [2104.01410]. The underlying singular-value-transformation paradigm comes from QSVT, which applies polynomial transformations to the singular values of a block of a unitary and subsumes optimal Hamiltonian simulation and related matrix algorithms [1806.01838].

## 1. Formal setting and spectral reduction

The basic HSVT setting assumes the ability to apply a Hamiltonian
$$
H=\begin{pmatrix}
\ast & A\\
A^\dagger & \ast
\end{pmatrix},
$$
where the $\ast$ blocks are arbitrary, the main off-diagonal blocks are $A$ and $A^\dagger$, and $A^\dagger A \leq I$ [2104.01410]. The construction also assumes the Pauli-$Z$-type Hamiltonian
$$
Z=\begin{pmatrix}
I & 0\\
0 & -I
\end{pmatrix}.
$$
By alternating application of $\pm H$ and $\pm Z$, as in echo/refocusing sequences, the diagonal elements of $H$ can be averaged out, leaving an effective Hamiltonian with only off-diagonal blocks [2104.01410].

A closely related spectral picture appears in the quantum polar decomposition algorithm, which uses the block Hamiltonian
$$
\begin{pmatrix}
0 & A^\dagger\\
A & 0
\end{pmatrix}
$$
acting on the direct sum $\mathcal{H}_L \oplus \mathcal{H}_R$ [2006.00841]. If
$$
A=\sum_j \sigma_j |l_j\rangle\langle r_j|,
$$
then the eigenvectors of this block Hamiltonian are combinations of left and right singular vectors,
$$
|\pm_j\rangle=\frac{1}{\sqrt{2}}\left(|l_j\rangle \oplus \pm |r_j\rangle\right),
$$
with eigenvalues $\pm \sigma_j$ [2006.00841]. This structure is the spectral mechanism by which transformations on singular values are converted into Hamiltonian dynamics.

## 2. Alternating-Hamiltonian implementation

Within a two-dimensional subspace corresponding to a singular value $\sigma_j$ of $A$, the effective Hamiltonian acts as
$$
H_j=\sigma_j X,
$$
where $X$ is the Pauli-$X$ matrix [2104.01410]. HSVT therefore reduces to controlled motion on a family of two-level systems indexed by singular values.

The central construction alternates evolution under $H$ and under $Z$, possibly conjugated by phase shifts $\phi_i$, through sequences such as
$$
e^{-iG_0 t_k}\cdots e^{-iG_0 t_1},
$$
where
$$
G_0=e^{-i\phi}
\begin{pmatrix}
0 & A\\
A^\dagger & 0
\end{pmatrix}
e^{i\phi},
$$
described as a Hamiltonian with axis in the $xy$-plane [2104.01410]. Through this sequence one can generate, for each subspace indexed by $j$, polynomial transformations of the form $P(\cos \sigma_j)$ and $Q(\cos \sigma_j)$, namely polynomials of the cosines and sines of $\sigma_j$ times evolution intervals [2104.01410].

This is the Hamiltonian analogue of QSVT: instead of alternating a block-encoded unitary with phase gates, it alternates accessible Hamiltonians for selected time intervals. The 2021 formulation explicitly identifies this as an example of the Quantum Alternating Operator Ansatz and as a generalized QAOA construction [2104.01410].

## 3. Resultant unitary and inverse block encoding

The general transformation achievable by HSVT has the form
$$
U_f=i\begin{pmatrix}
\sqrt{I-f(A^\dagger)f(A)} & f(A)\\
-f(A^\dagger) & \sqrt{I-f(A)f(A^\dagger)}
\end{pmatrix},
$$
where $f(A)$ is a polynomial function of the singular values of $A$ with the same singular vectors, constructed from Chebyshev polynomial expansions as in QSVT [2104.01410]. In this sense, HSVT inherits the function-calculus role of QSVT while replacing circuit-level block-encoding access by Hamiltonian access.

A special case is obtained by setting $f(x)=x$, which yields
$$
U=i\begin{pmatrix}
\sqrt{I-A^\dagger A} & A\\
-A^\dagger & \sqrt{I-AA^\dagger}
\end{pmatrix}.
$$
This is called inverse block encoding: given a Hamiltonian with $A$ as a block, one deterministically implements a unitary with $A$ as a block [2104.01410]. The paper characterizes this as a “Hamiltonian-to-unitary” conversion and treats it as pivotal for subsequent algorithms.

The main procedures introduced around HSVT can be summarized as follows.

| Procedure | Input requirement | Output or effect |
|---|---|---|
| HSVT | Hamiltonian $H$ with block $A$ | Unitary block with any $f(\cdot)$ function of singular values |
| Inverse block encoding | As above, set $f(x)=x$ | Unitary with block $A$ |
| Matrix multiplication/state prep. | State $|v\rangle$, Hamiltonian $H$ | State proportional to $A|v\rangle$ |
| Differential equation solver | $B$, initial state | State $e^{Bt}|v\rangle$ |
| Matrix inversion/history states | As above | State proportional to $(I-A^\dagger A)^{-1}\cdots|v\rangle$ |

## 4. Algorithmic procedures and applications

For matrix multiplication, inverse block encoding acts on an input state $|v\rangle$ as
$$
U
\begin{pmatrix}
|v\rangle\\
0
\end{pmatrix}
=
i
\begin{pmatrix}
\sqrt{I-A^\dagger A}\,|v\rangle\\
A|v\rangle
\end{pmatrix},
$$
so measurement of the second register yields $A|v\rangle$ [2104.01410]. The same presentation states that amplitude amplification can quadratically speed up successful state preparation [2104.01410].

For first-order linear ordinary differential equations
$$
\frac{d}{dt}|\psi(t)\rangle = B|\psi(t)\rangle,
$$
the construction sets $A=I+B\Delta t$ and iterates the inverse-block-encoding procedure [2104.01410]. The resulting state has the form
$$
\left(\prod_{k=0}^{n-1}\sqrt{I-A_k^\dagger A_k}\right)A_{n-1}\cdots A_0 |v\rangle,
$$
which approximates the Euler forward solution, and measurement yields the evolved state $e^{Bt}|v\rangle$ [2104.01410]. The same formulation states that this is valid for dissipative systems with $\operatorname{Re}[\operatorname{eig}(B)]<0$, as required by $A^\dagger A \leq I$ [2104.01410].

By additional manipulation, the method also enables probabilistic quantum matrix inversion and obtains the “history state” for ODE solutions as required in quantum simulation algorithms [2104.01410]. These applications are significant because they are presented as purely Hamiltonian procedures rather than gate-decomposition-first implementations.

## 5. Relation to QAOA, polar decomposition, and recursive constructions

HSVT is explicitly described as generalized QAOA: instead of variationally optimizing a fixed-cost and mixer Hamiltonian, it alternates application of two or more accessible Hamiltonians for selected time intervals [2104.01410]. The same source states that, as with QAOA, the Lie algebra generated by the set of accessible Hamiltonians can densely span the full space of unitaries, providing both universality and flexibility for variational optimization [2104.01410].

An earlier Hamiltonian-level route to singular-value manipulation is the quantum polar decomposition algorithm. It shows that access to
$$
\begin{pmatrix}
0 & A^\dagger\\
A & 0
\end{pmatrix}
$$
translates into deterministic implementation of the unitary or isometry $U=A(A^\dagger A)^{-1/2}$ through the sign function, and of $e^{-iBt}$ for the positive part $B=(A^\dagger A)^{1/2}$ through functions of $|x|$; the same work states that it can perform a Hamiltonian version of QSVT [2006.00841]. This places HSVT in direct continuity with polar decomposition, pretty good measurements, the quantum Procrustes problem, and positive-part constructions [2006.00841].

A later development is recursive QET/QSVT, which organizes complicated matrix functions by recursively composing low-degree transformations with analytically chosen parameters [2304.13330]. For the matrix sign function, an analytically obtained parameter set composed of only $8$ different values is sufficient for executing QET with an arbitrarily small error $\varepsilon$ [2304.13330]. The same framework generalizes to singular-value transformations, with the recursion
$$
X_{n+1}=\frac{X_n}{8}\left(15-10X_n^\dagger X_n+3(X_n^\dagger X_n)^2\right),
$$
converging to $A(\sqrt{A^\dagger A})^{-1}$, the unitary part of the polar decomposition [2304.13330]. This suggests an analytic route to HSVT-style parameter synthesis in regimes where high-degree phase finding is numerically unstable.

## 6. Scope, complexity, and adjacent frameworks

The 2021 HSVT formulation emphasizes a purely Hamiltonian implementation that avoids the need for deep quantum circuits or explicit logical gate decomposition and is accessible to systems controlled via time-dependent semiclassical fields, including analog and near-term quantum devices [2104.01410]. It also states that the number of alternations required is
$$
O(\log(1/\epsilon)/\sigma_{\min})
$$
for error $\epsilon$, matching or improving upon circuit-based quantum algorithms [2104.01410].

For Hermitian Hamiltonian simulation, later QSVT-based analyses state that QSVT achieves the minimum simulation time and that, within that framework, oblivious amplitude amplification gives query complexity $\mathcal{O}(t+\log(1/\epsilon))$, whereas fixed-point amplitude amplification gives $\mathcal{O}(t\log(1/\epsilon)+\log^2(1/\epsilon))$ [2304.08937]. Block-encoding-free variants developed later eliminate block encoding, need only a single ancilla qubit, and use direct Hamiltonian simulation with circuit depth
$$
\widetilde{O}\!\left(L(d\lambda_{\mathrm{comm}})^{1+o(1)}\right)
$$
[2504.02385]. Randomized QSVT methods similarly avoid block encodings, use only a single ancilla qubit, and achieve gate complexity independent of the number of Hamiltonian terms, with quadratic dependence on the degree of the target polynomial [2510.06851]. These developments are not HSVT in the narrow 2021 sense, but they continue the broader shift toward Hamiltonian-native singular-value manipulation.

A recurring misconception is that singular-value transformation already covers arbitrary eigenvalue transformation. Later work distinguishes the two: eigenvalue transformations are distinct, especially for non-normal matrices, and are stated to be beyond the reach of the QSVT framework [2411.04010; 2601.11959]. Laplace-transform-based and contour-integral-based methods therefore target functions such as $A^{-k}$, $e^{-A^{-1}}$, and general holomorphic $f(A)$ for dissipative or non-normal matrices, whereas HSVT remains a singular-value framework [2411.04010; 2601.11959].

Another boundary concerns classical simulability. For sparse matrices, low-degree polynomial QSVT can be classically simulated efficiently with arbitrarily small constant precision, and the same paper states that the dequantization technique applies also to Hamiltonian singular value transformation when the function is an even low-degree polynomial, the Hamiltonian is sparse, and access to a guiding vector is samplable [2111.09079]. In the same source, inverse-polynomial precision becomes BQP-complete [2111.09079]. This does not diminish HSVT’s algorithmic role; it clarifies that the regime of claimed quantum advantage is sensitive to precision, access model, and the degree of the transformed polynomial.

Source: https://www.emergentmind.com/topics/hamiltonian-singular-value-transformation