Singular value transformation for unknown quantum channels

Ryotaro Niwa, Zane Marius Rossi, Philip Taranto, Mio Murao · 2025

Read Singular value transformation for unknown quantum channels by Ryotaro Niwa, Zane Marius Rossi, Philip Taranto, Mio Murao (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-24112.txt.

Introduction

Ryotaro Niwa and co-authors transform singular values of unknown quantum channels in 2025. A quantum channel describes how a system changes under noise or a device. Users may be able to run the channel without knowing its inner workings. The team builds an algorithm that reshapes its singular values anyway.

Their trick encodes the channel in a Hermitian form from black box access alone. Standard singular value tools can then apply polynomial reshaping. Cost bounds are proven from above and below for the encoding step. A sample use reads power sums of the singular values of a mystery channel.

Catalog record

CREATORRyotaro Niwa, Zane Marius Rossi, Philip Taranto, Mio Murao
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-24112

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Singular value transformation for unknown quantum channels Ryotaro Niwa, Zane Marius Rossi, Philip Taranto, Mio Murao Abstract Given the ability to apply an unknown quantum channel acting on a d d -dimensional system, we develop a quantum algorithm for transforming its singular values. The spectrum of a quantum channel as a superoperator is naturally tied to its Liouville representation, which is in general non-Hermitian. Our key contribution is an approximate block-encoding scheme for this representation in a Hermitized form, given only black-box access to the channel; this immediately allows us to apply polynomial transformations to the channel’s singular values by quantum singular value transformation (QSVT). We then demonstrate an O ⁡ ( d 2 / δ ) O(d^{2}/\delta) upper bound and an Ω ⁡ ( d / δ ) \Omega(d/\delta) lower bound for the query complexity of constructing a quantum channel that is δ \delta -close in diamond norm to a block-encoding of the Hermitized Liouville representation. We show our method applies practically to the problem of learning the q q -th singular value moments of unknown quantum channels for arbitrary q > 2 , q ∈ ℝ q>2,q\in\mathbb{R} , which has implications for testing if a quantum channel is entanglement breaking. Introduction.— To date, various quantum algorithms have demonstrated that the laws of quantum mechanics can be leveraged to efficiently process information. A unifying framework encompassing many important quantum algorithms such as search, matrix inversion, and Hamiltonian simulation, is the quantum singular value transformation (QSVT) Gilyén et al. 2019 . Generalizing the results of quantum signal processing (QSP) Low et al. 2016 , it allows one to directly apply polynomial functions to the singular values of linear operators encoded into sub-blocks of a larger unitary. While QSVT is a powerful framework for designing quantum algorithms, the cost of block-encoding a linear operator (Def. 1 ) can sometimes dominate the overall algorithmic cost. To efficiently construct block-encodings, many works have restricted to specific cases, e.g., where the input matrix is sparse Camps et al. 2023 or has special structure Sünderhauf et al. 2024 . When the matrix of interest has physical meaning, such as a density matrix or a quantum channel, it is often assumed that one has access to a purification of the density matrix Gilyén et al. 2019 or the Kraus operators of the channel, respectively Chen et al. 2023 . However, these access models may be too strong from the viewpoint of physical implementability; a purification of a density matrix or the Kraus operators of a channel cannot always be obtained given the quantum states or channels themselves. Figure 1: (a) Approximate block-encoding scheme for the Liouville representation \mathcal{E}_{A} . By applying an unknown channel \mathcal{E} multiple times we construct a quantum channel \widetilde{\mathcal{U}} that approximates a unitary block-encoding U_{\mathcal{E}_{A}} of the Hermitized Liouville representation. (b) Circuit for quantum singular value transformation (QSVT). Here, R_{z}^{(n)}=e^{-i\phi_{n}\sigma_{z}} are single qubit rotations while grey boxes are projection-controlled-NOT operations locating \mathcal{E}_{A} as a sub-block of U_{\mathcal{E}_{A}} . In this Letter, we consider the fundamental task of transforming the singular values of quantum channels in the most general and physical access model: one where the quantum channel can only be applied as a black box . As we aim to investigate the singular values of the quantum channel—a superoperator acting on density matrices—we first have to make sense of density matrices as vectors. Under such vectorization , the action of a quantum channel is given by a matrix called the Liouville representation Wallman and Flammia 2014 , which is in general non-Hermitian. This representation translates channel application into matrix multiplication, making it particularly suited for discussing spectral properties of a channel, for instance, the mixing time of Lindblad dynamics Lindblad 1976 dictated by its spectral gap. This unique feature is absent in other matrix representations of the channel such as the Choi representation Jamiołkowski 1972 ; Choi 1975 . Although the non-Hermiticity of the Liouville representation seemingly complicates its physical implementation, we show we can construct a quantum channel arbitrarily close to a unitary encoding this matrix as a sub-block. We emphasize that our scheme is achieved under the black-box channel access model, avoiding assumptions such as having access to the channel’s Kraus operators. With this block-encoded Liouville representation at hand, we can freely apply standard quantum algorithmic tools such as QSVT, which will allow us to transform the singular values of block-encoded matrices without having to know their singular values or singular vectors (see FIG. 1 ). As a concrete application of our scheme, we consider the task of learning the q -th singular value moments of unknown quantum channels for arbitrary q>2,q\in\mathbb{R} , based on a polynomial approximation to f(x)=\frac{1}{2}x^{q-2} . This extends previous works on entanglement detection Cai and Song 2008 ; Liu et al. 2022 , which discuss the special case of even integer q , and builds upon recent work for estimating the trace of quantum state powers \Tr\rho^{q} for q\in\mathbb{R} Liu and Wang 2025 . We also show that our exponentiation scheme is useful for testing if a quantum channel is entanglement breaking based on the reshuffling criterion of entanglement Rudolph 2005 ; Chen and Wu 2003 . Our work not only extends recent work Wei et al. 2024 discussing exponentiation of Hermiticity-preserving maps to the realm of non-Hermiticity preserving maps, but also provides a foundation for universally manipulating black-box quantum channels with standard techniques such as QSVT, allowing us to directly evaluate functions of unknown quantum channels without reducing to tomography and/or classical postprocessing. Preliminaries.— Let \mathfrak{S}(\mathcal{H}),\mathfrak{L}(\mathcal{H}) respectively denote the set of density matrices and linear operators on the Hilbert space \mathcal{H} , and let \mathfrak{C}(\mathcal{H}) denote the set of completely positive trace-preserving (CPTP) maps acting on \mathfrak{L}(\mathcal{H}) . A quantum channel \mathcal{E}\in\mathfrak{C}(\mathcal{H}) has various representations; two particularly important ones relevant to our manuscript are the Liouville representation and the Choi representation 1 1 1 The Liouville representation \mathcal{E}_{A} and the Choi state \mathcal{E}_{B} are sometimes called the A-form and the B-form matrices, respectively (as they were originally derived in Ref. Sudarshan et al. 1961 ) , as we describe below. (A) Liouville representation: When \rho\in\mathfrak{S}(\mathcal{H}) is represented by its vectorized form |\rho\rangle\rangle\in\mathcal{H}\otimes\mathcal{H} as \displaystyle\rho=\sum_{ij}\rho_{ij}|i\rangle\langle j|\rightarrow|\rho\rangle\rangle:=\sum_{ij}\rho_{ij}|i\rangle\otimes|j\rangle, (1) the action of a quantum channel \mathcal{E}\in\mathfrak{C}(\mathcal{H}) becomes the matrix multiplication \displaystyle|\mathcal{E}(\rho)\rangle\rangle=\mathcal{E}_{A}|\rho\rangle\rangle. (2) The matrix \mathcal{E}_{A}\in\mathfrak{L}(\mathcal{H}\otimes\mathcal{H}) is called the Liouville representation Wallman and Flammia 2014 of the channel \mathcal{E} , which is in general non-Hermitian . This suggests that \mathcal{E}_{A} itself cannot be obtained as a quantum state. On the other hand, an important consequence of Eq. ( 2 ) is that \mathcal{E}_{A} is naturally associated with the spectrum of the channel \mathcal{E} as a superoperator. Indeed, if \displaystyle\mathcal{E}(\rho_{n})=\lambda_{n}\rho_{n}\quad(\lambda_{n}\in\mathbb{C}) (3) for some \rho_{n}\in\mathfrak{L}(\mathcal{H}) (not necessarily \rho_{n}\in\mathfrak{S}(\mathcal{H}) ), then \displaystyle\mathcal{E}_{A}|\rho_{n}\rangle\rangle=|\mathcal{E}(\rho_{n})\rangle\rangle=\lambda_{n}|\rho_{n}\rangle\rangle. (4) Thus the eigenvalues of the non-Hermitian matrix \mathcal{E}_{A} coincide with the eigenvalues of the quantum channel (see Appendix A ). Analogously, the singular values of the quantum channel \mathcal{E} can be defined as the square root of the eigenvalues of the map \mathcal{E}^{\dagger}\circ\mathcal{E} . Since map composition becomes matrix multiplication in the Liouville representation due to ( 2 ), the singular values of the quantum channel coincide with those of \mathcal{E}_{A} : \displaystyle\mathcal{E}_{A}=U\Sigma V^{\dagger},\Sigma=\mathrm{diag}[\sigma_{1},\sigma_{2},\cdots,\sigma_{d^{2}}], (5) where U,V\in\mathfrak{L}(\mathcal{H}\otimes\mathcal{H}) are

Read the complete text