Quantum Eigenvalue Transformations for Arbitrary Matrices

arXiv:2604.19688 · quant-ph · Submitted 2026-04-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Quantum Eigenvalue Transformations for Arbitrary Matrices".

Paper summary: Kai: So, we're looking at this paper called "Quantum Eigenvalue Transformations for Arbitrary Matrices." It tackles a problem where you can use Quantum Signal Processing or QSP to implement polynomials for block-encoded matrices, which is cool because it gives you a systematic way to design quantum algorithms.

Mira: That sounds useful right away, but the authors point out some big limitations with existing methods. QSP only works for unitary matrices when transforming eigenvalues, and QSVT only gives you singular values for Hermitian matrices. So, they can't directly do this for a general non-Hermitian matrix that isn't diagonalizable.

Lev: That means these established techniques just don't apply to the kind of arbitrary non-Hermitian matrix we often deal with in real hardware simulations, which is where most of our work is focused.

Kai: Exactly. The paper says neither QSP nor QSVT can be used directly for the polynomial transformation of a general non-Hermitian matrix that isn't diagonalizable; it's a direct roadblock for those applications.

Mira: But they propose a way around this by introducing an n-regular block encoding and then extending the idea to arbitrary square matrices by acting on their eigenvalues instead.

Lev: So, the core claim is that if you have this n-regular block encoding, applying QSP to any unitary with that property is equivalent to applying a polynomial of degree at most n to the block-encoded matrix, regardless of its internal structure.

Kai: And they show they can build this kind of regular encoding from any arbitrary block encoding using only about log n ancillary qubits and some operations scaling like n log over epsilon.

Mira: That's the construction part, which is important because it suggests we don't need a perfectly regular starting point to get this powerful eigenvalue transformation capability.

Lev: It also shows that if you start with the k-th power of a unitary block-encoding being the k-th power of another matrix, then regularity holds for those powers, which is what allows them to link QSP polynomials to the full matrix function P(A).

Kai: So, they aren't just transforming individual eigenvalues; they are getting the full matrix function P(A), which involves the derivatives of that polynomial on each Jordan block, not just mapping lambda to P of lambda.

Mira: That means we're getting something much more complete than just a simple eigenvalue map. The paper mentions that this works even for non-diagonalizable matrices, which is a big deal.

Lev: From an engineering standpoint, the complexity they claim is better in the short time regime where time t is small, specifically when t equals O of one. For example, the number of ancillary qubits they use for this method is significantly lower compared to methods like LCHS for simulating time-independent linear differential equations.

Kai: That low qubit count is what makes it appealing for near-term quantum hardware because we’re working with limited resources right now.

Paper summary: Mira: And another point they make, which I think you'll appreciate, is that the complexity of constructing this transformation for a matrix A doesn't depend on the eigenvector conditioning kappa of S. That means even defective generators at exceptional points don't cause a penalty in terms of construction time or resources.

Lev: That removes a major hurdle because those exceptional points are often where other methods break down entirely when you try to run them on real hardware, and this method handles them gracefully.

Kai: There is one limitation they flag, though, which is that the Quantum Signal Processing approach requires functions that are analytic on the entire unit disk D. Even if the numerical range of your matrix A is safely in a smaller subset of that disk, you still need analyticity across the whole thing.

Mira: They suggest you can get around this by using a conformal map M from where your eigenvalues lie to the unit circle, and then applying the function f to M of z instead.

Lev: If we use that mapping trick, they say the complexity for implementing something like e to minus t A in the long time regime drops down to something like O of t squared plus log over epsilon.

Kai: That scaling for long time exponentiation sounds much more manageable than what we might expect from other simulation techniques right now.

Mira: So, putting it all together, the paper "Quantum Eigenvalue Transformations for Arbitrary Matrices" proposes a way to apply a polynomial to the eigenvalues, or more generally the Jordan blocks, of an arbitrary block-encoded square matrix A.

Lev: They show that if you have this k-th power U k block-encodes A k, regularity holds and you can get P(A) by applying P as a QSP polynomial to U. This construction has lower query complexity and extra two qubit gates compared to other methods in the short time regime.

Kai: The main implication is that this method is applicable to any square matrix, including non-diagonalizable ones, which opens up a new path for simulating these complex systems on quantum computers.

Mira: It seems like they've found a way to bridge the gap between QSP and general matrix functions without needing super-specific matrix structures like unitarity or Hermiticity.

Lev: The comparison table shows that this GQSP plus n-regular block encoding method achieves better complexity for arbitrary square matrices, hitting O of log over epsilon log log over epsilon extra gates and O of log over log over epsilon ancillas for the general case.

Kai: That outperforms methods like LCHS, which requires super quadratic extra gates, which is a big win for practical implementation efficiency.

Mira: And it's particularly favorable for short-time exponentiation at exceptional points where other standard methods just don't work anymore. This feels like a very targeted improvement for certain types of quantum simulation problems.

Lev: It shows that even though the underlying math is complex, if you can find the right block encoding structure, you can get a more efficient quantum implementation than what we currently have available.

Conclusion: Kai: So this paper is about extending those quantum polynomial methods to any square matrix, even ones that aren't nice and tidy like Hermitian or unitary matrices.

Mira: Yeah, the authors are showing you how to get a full matrix function by using a regular block encoding property instead of being stuck with just singular values or eigenvalues.

Kai: They introduce this idea of n-regular block encoding which lets you apply these quantum signal processing tools to general matrices, period.

Mira: It’s neat because they show how you can build that regular structure from any starting point using only a few extra qubits and some manageable operations.

Kai: And the main result is that this method lets you compute the full matrix function P of A, which includes all those weird derivatives on the Jordan blocks, not just a simple map from lambda to P of lambda.

Mira: That means for any general square matrix, you can get that complete function using these quantum techniques.

Kai: It’s applicable to non-diagonalizable matrices too, which is a big deal for things like time evolution or solving differential equations where the matrix isn't diagonalizable.

Mira: Yeah, so the implication here is that we might not be as restricted about which types of problems we can even tackle with these quantum algorithms.

Kai: It really opens up doors for simulating complex systems where you’re dealing with non-trivial dynamics on arbitrary matrices.

Mira: It sets a new benchmark for how far this type of transformation can be pushed in the context of quantum algorithms.

Kai: But there are still limitations, like the function having to be analytic everywhere on the unit disk, which they address with some clever mapping tricks.

Mira: So it’s not a magic fix for every single limitation, but it shows a much more robust framework for general matrix problems overall.

Kai: We'll be talking about how this actually performs compared to other methods in the next part of the show.

Department of Physical Chemistry, University of the Basque Country UPV/EHU · Faculty of Informatics – Universitá della Svizzera Italiana · Basque Center for Applied Mathematics (BCAM)

quant-ph

Submitted: 2026-04-21

Updated: 2026-09-01

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: Quantum Signal Processing (QSP) and Quantum Singular Value Transformation (QSVT) provide an efficient framework for implementing polynomials of block-encoded matrices…

Key concepts

n-regular block encoding
This is a special type of block encoding where raising the encoded matrix to any power up to n reproduces the result of applying a polynomial of degree n to the original matrix. This property allows QSP polynomials to operate on this structure, regardless of the encoding's internal details.
Quantum Eigenvalue Transformation
This technique enables finding solutions related to a matrix's characteristic polynomials (eigenvalues) for any square matrix, even those that are not diagonalizable. It achieves this by leveraging the regularity property of block encodings and applying QSP polynomials to transform the encoded representation.
GQSP polynomial
This refers to a specific type of polynomial used in Quantum Signal Processing that is adapted for this transformation. It allows one to obtain the full matrix function P(A) by applying it as a quantum operation to the block encoding, which involves handling derivatives across Jordan blocks.

Terminology

Summary

Quantum Signal Processing (QSP) and Quantum Singular Value Transformation (QSVT) provide an efficient framework for implementing polynomials of block-encoded matrices, and thus offer a systematic approach to quantum algorithm design<ref:2604. p. 0

Where the paper says it

Quantum Eigenvalue Transformations for Arbitrary Matrices Xabier Guti´errez Department of Physical Chemistry, University of the Basque Country UPV/EHU, Apartado 644, 48080 Bilbao, Spain and EHU Quantum Center, University of the Basque Country UPV/EHU, Apartado 644, 48080 Bilbao, Spain Lorenzo Laneve Faculty of Informatics — Universit`a della Svizzera Italiana, 6900 Lugano, Switzerland Mikel Sanz Department of Physical Chemistry, University of the Basque Country UPV/EHU, Apartado 644, 48080 Bilbao, Spain EHU Quantum Center, University of the Basque Country UPV/EHU, Apartado 644, 48080 Bilbao, Spain and Basque Center for Applied Mathematics (BCAM), Alameda de Mazarredo 14, 48009 Bilbao, Spain (Dated: September 2, 2026) Quantum Signal Processing (QSP) and Quantum Singular Value Transformation (QSVT) provide an efficient framework for implementing polynomials of block-encoded matrices, and thus offer a systematic approach to quantum algorithm design. However, despite a number of recent advances, important limitations remain.

19688#pg3>. In this work, we propose a simple yet powerful method to extend these ideas to arbitrary square matrices by acting on their eigenvalues<ref:2604. p. 0
Where the paper says it

As a consequence, these techniques do not directly apply to an arbitrary non-Hermitian matrix that is not diagonalizable. In this work, we propose a simple yet powerful method to extend these ideas to arbitrary square matrices by acting on their eigenvalues.

19688#pg4>.

Limitations of Existing Frameworks

Existing techniques suffer from important limitations because QSP can only transform unitary matrices, while QSVT is a singular-value transformation and thus one can only obtain the polynomial of Hermitian matrices<ref:2604.19688#pg3>. Consequently, these techniques do not directly apply to an arbitrary non-Hermitian matrix that is not diagonalizable<ref:2604. p. 0

Where the paper says it

In particular, QSP can only transform unitary matrices, by applying a polynomial to their eigenvalues, while QSVT is a singular-value transformation and thus one can only obtain the polynomial of Hermitian matrices. As a consequence, these techniques do not directly apply to an arbitrary non-Hermitian matrix that is not diagonalizable.

19688#pg4>. Specifically, neither method can be used directly to implement the polynomial transformation of a general non-Hermitian matrix<ref:2604. p. 1
Where the paper says it

2 formation of a unitary matrix, or a singular value transformation of a block-encoded matrix A, but they do not directly provide an eigenvalue transformation of a blockencoded matrix, except in the Hermitian case, where singular values and eigenvalues coincide up to sign. Consequently, for a general non-Hermitian matrix, neither method can be used directly to implement its polynomial transformation.

19688#pg3>.

The Proposed Regular Block Encoding

To extend these ideas, the authors introduce the notion of an n-regular block encoding, which is a block encoding whose k-th power reproduces the k-th power of the encoded matrix for every 0 ≤ k ≤ n<ref:2604.19688#pg4>. They show that applying QSP to any unitary with this property is equivalent to applying a polynomial of degree at most n to the block-encoded matrix, independently of its internal structure<ref:2604. p. 0

Where the paper says it

To this end, we introduce the notion of an n-regular block encoding, namely, a block encoding whose k-th power reproduces the k-th power of the encoded matrix for every 0 ≤ k ≤ n. We show that applying QSP to any unitary with this property is equivalent to applying a polynomial of degree at most n to the block-encoded matrix, independently of its internal structure.

19688#pg4>. Furthermore, they provide a simple construction that transforms any block encoding into an n-regular one using only O(log n) ancillary qubits and O(n log(n/ϵ)) operations<ref:2604. p. 0
Where the paper says it

We show that applying QSP to any unitary with this property is equivalent to applying a polynomial of degree at most n to the block-encoded matrix, independently of its internal structure. Moreover, we provide a simple construction that transforms any block encoding into an n-regular one using only O(log n) ancillary qubits and O(n log(n/ϵ)) operations.

19688#pg4>.

The Quantum Eigenvalue Transformation

The proposed technique allows for the transformation of eigenvalues, understood as solutions to the characteristic polynomials of a matrix, of an arbitrary square matrix, even non-diagonalizable ones<ref:2604.19688#pg4>. This is achieved by showing that if the k-th power U k block-encodes A k, a property called regularity holds<ref:2604. p. 4

Where the paper says it

Conclusions — In this Article we propose a method to apply a polynomial to the eigenvalues (or, more generally, to the Jordan blocks) of a block-encoded square matrix A. We have shown that, if the k-th power U k block-encodes Ak , a property we call regularity, then P(A) can be obtained by simply applying P as a GQSP polynomial to U, and that any block-encoding can be turned into an n-regular one by equipping it with a quantum incrementer.

19688#pg4>. Applying P as a GQSP polynomial to U yields a block encoding of P(A)<ref:2604. p. 5
Where the paper says it

We also note that unitaries not only block-encode themselves, but are also ∞-regular, which provides an alternative viewpoint on why applying GQSP to unitaries effectively yields an eigenvalue transformation. At the same time, if we consider a block-encoding of a matrix A that is close to a unitary, then the block-encoding will be approximately regular, since the failed branches multiply by blocks of small operator norm.

19688#pg4>. This results in obtaining the full matrix function P(A), which involves the derivatives of P on each Jordan block, not merely the map λ -> P(λ)<ref:2604. p. 0
Where the paper says it

Moreover, we provide a simple construction that transforms any block encoding into an n-regular one using only O(log n) ancillary qubits and O(n log(n/ϵ)) operations. Finally, we show that this construction induces the desired transformation on the eigenvalues associated with the Jordan normal form of the matrix; for non-diagonalizable matrices this yields the full matrix function P(A), whose action on each Jordan block involves the derivatives of P, and not merely the map λ 7→ P(λ).

19688#pg4>.

Performance and Comparison

The method achieves better complexity in every metric when considering the short-time regime, i.e., t = O(1)<ref:2604. p. 0

Where the paper says it

Despite their rapid development, these techniques still suffer from important limitations. In particular, the existing frameworks can perform either an eigenvalue transarXiv:2604.

19688#pg4>. For example, the number of ancillary qubits is significantly lower compared to LCHS methods for simulating time-independent homogeneous linear differential equations<ref:2604. p. 3
Where the paper says it

, implementing e tA for t = O(1)), this method requires fewer copies of A than the LCHS method for simulating time-independent homogeneous linear differential equations presented in [20], and it can be used to address unstable evolutions as well. The number of ancillary qubits is also significantly lower, see Table I.

19688#pg4>. The complexity of the construction for a matrix A is independent of the eigenvector conditioning κ(S), which applies without penalty to defective generators at exceptional points<ref:2604. p. 1
Where the paper says it

In particular, since the cost of our construction for a matrix A is independent of the eigenvector conditioning κ(S), where A = SJS−1 is the Jordan decomposition of the matrix, it applies without penalty to defective generators (i. , with high κ(S)) at exceptional points, a regime of intense current interest in non-Hermitian photonics, PT-symmetric and open quantum systems that is out of reach of previous techniques.

19688#pg4>.

Domain Manipulation

A limitation is that GQSP requires functions that are analytic on the entire unit disk D, even if the numerical range of A can be safely assumed to be in a proper subset of it<ref:2604.19688#pg4>. This limitation can be circumvented by manipulating the domain of the function via a conformal map M(z) from the region in which the eigenvalues lie to the unit circle<ref:2604. p. 5

Where the paper says it

A limitation of this method lies in the rigidity of the domain: GQSP requires the construction of functions that are analytic on the entire unit disk D, even if we can assume that the eigenvalues of A lie in a proper subset of it (for example, the stability region for the simulation of non-unitary dynamics). We can circumvent this problem via a conformal map M(z) from the region in which the eigenvalues lie to the unit circle: we first construct M−1 (A), and apply the function f(M(z)) to it, with M(z) chosen such that the composite function is easier to handle.

19688#pg4>. This involves constructing M−1(A) and applying f(M(z)) to it<ref:2604. p. 0
Where the paper says it

Moreover, we provide a simple construction that transforms any block encoding into an n-regular one using only O(log n) ancillary qubits and O(n log(n/ϵ)) operations. Finally, we show that this construction induces the desired transformation on the eigenvalues associated with the Jordan normal form of the matrix; for non-diagonalizable matrices this yields the full matrix function P(A), whose action on each Jordan block involves the derivatives of P, and not merely the map λ 7→ P(λ).

19688#pg4>. The resulting complexity for implementing e−tA in the long-time regime is O(t squared + log 1/ϵ)<ref:2604. p. 5
Where the paper says it

We can circumvent this problem via a conformal map M(z) from the region in which the eigenvalues lie to the unit circle: we first construct M−1 (A), and apply the function f(M(z)) to it, with M(z) chosen such that the composite function is easier to handle. We provide an example for the exponential in the Supplementary Material section G, with a total complexity O(t 2 + log 1/ϵ) in the long-time regime.

19688#pg4>.

Conclusion

The authors propose a method to apply a polynomial to the eigenvalues (or, more generally, to the Jordan blocks) of a block-encoded square matrix A<ref:2604.19688#pg4>. They have shown that if the k-th power U k block-encodes A k, regularity holds and P(A) can be obtained by applying P as a GQSP polynomial to U<ref:2604. p. 4

Where the paper says it

Conclusions — In this Article we propose a method to apply a polynomial to the eigenvalues (or, more generally, to the Jordan blocks) of a block-encoded square matrix A. We have shown that, if the k-th power U k block-encodes Ak , a property we call regularity, then P(A) can be obtained by simply applying P as a GQSP polynomial to U, and that any block-encoding can be turned into an n-regular one by equipping it with a quantum incrementer.

19688#pg4>. This construction has lower query complexity and extra two-qubit gates compared to other methods in the short-time regime<ref:2604. p. 3
Where the paper says it

Since our construction never references the eigenvector basis, it implements e tA in short-time regime at the exceptional point with exactly the same cost O log 1/ϵ log log 1/ϵ as in the diagonalizable case, whereas LCHS is inapplicable (A+A† ⪰̸ 0) and the Chebyshev/Faber guarantees degrade with κ(S) → ∞. Besides having no condition on the transformed matrix, for short-time exponentiation the method presented here has lower query complexity, extra two-qubit gates and extra ancillas compared with the rest of the methods (see Table D.

19688#pg4>. The method is applicable to arbitrary square matrices, including non-diagonalizable ones<ref:2604. p. 1
Where the paper says it

This leads to small success probabilities and, consequently, to further overhead in the overall complexity. In this Article, we propose a method to carry out transformations of eigenvalues, understood as solutions to the characteristic polynomials of a matrix, of an arbitrary square matrix, even non-diagonalizable ones.

19688#pg4>.

Table II Comparison

The comparison table shows that the work (GQSP + n-regular BE) achieves better complexity for arbitrary square matrices<ref:2604.19688#pg4>. Specifically, it achieves O(log 1/ϵ log log 1/ϵ) extra gates and O(log log 1/ϵ) ancillas for the general case<ref:2604. p. 3

Where the paper says it

Since our construction never references the eigenvector basis, it implements e tA in short-time regime at the exceptional point with exactly the same cost O log 1/ϵ log log 1/ϵ as in the diagonalizable case, whereas LCHS is inapplicable (A+A† ⪰̸ 0) and the Chebyshev/Faber guarantees degrade with κ(S) → ∞. Besides having no condition on the transformed matrix, for short-time exponentiation the method presented here has lower query complexity, extra two-qubit gates and extra ancillas compared with the rest of the methods (see Table D.

19688#pg4>. This outperforms methods like LCHS, which requires super-quadratic extra gates<ref:2604. p. 8
Where the paper says it

The count of all the extra gates, dominated by the complexity of the two QFTs, is thus O(n 2 ) = O(log2 1 ϵ ) [35]. Note that, on the other hand, LCHS requires super-quadratic extra gates, mainly due to the implementation of the state-preparation oracle for the LCU [20, Theorem 4].

19688#pg4>. The method is particularly favorable for short-time exponentiation at exceptional points where other methods are inapplicable<ref:2604. p. 3
Where the paper says it

Since our construction never references the eigenvector basis, it implements e tA in short-time regime at the exceptional point with exactly the same cost O log 1/ϵ log log 1/ϵ as in the diagonalizable case, whereas LCHS is inapplicable (A+A† ⪰̸ 0) and the Chebyshev/Faber guarantees degrade with κ(S) → ∞. Besides having no condition on the transformed matrix, for short-time exponentiation the method presented here has lower query complexity, extra two-qubit gates and extra ancillas compared with the rest of the methods (see Table D.

19688#pg4>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Arbitrary matrix polynomial transformation capability: The improved system can implement any polynomial transformation P(A) of an arbitrary square matrix A, including non-diagonalizable ones, by utilizing n-regular block encodings and Quantum Signal Processing (QSP).

  2. Jordan form function implementation: The system can perform a full matrix function P(A) by transforming the Jordan blocks according to P(Ji), where P(Ji) is an upper triangular Toeplitz matrix containing the derivatives of P evaluated at the eigenvalue lambda i, which handles non-diagonalizable matrices correctly.

  3. Efficiency for defectiveness: The method achieves a query complexity of O log 1/epsilon log log 1/epsilon for arbitrary square matrices, even when dealing with defective generators (high condition number kappa(S)), unlike previous methods like Chebyshev and Faber history states whose complexity degrades with kappa(S) to infinity.

  4. Short-time simulation accuracy: For short-time simulations (t = O 1), the system can implement matrix exponentials or second-order wave dynamics with a success amplitude of Theta 1, providing better query complexity than methods like LCHS for simulating time independent homogeneous linear differential equations.

  5. Domain flexibility via conformal maps: The system can handle functions whose eigenvalues lie in a proper subset of the unit disk by using a conformal map M(z) to transform the region into the entire unit disk, allowing analytic GQSP to be applied effectively.

Abstract

Quantum Signal Processing (QSP) and Quantum Singular Value Transformation (QSVT) provide an efficient framework for implementing polynomials of block-encoded matrices, and thus offer a systematic approach to quantum algorithm design. However, despite a number of recent advances, important limitations remain. In particular, QSP can only transform unitary matrices, by applying a polynomial to their eigenvalues, while QSVT is a singular-value transformation and thus one can only obtain the polynomial of Hermitian matrices. As a consequence, these techniques do not directly apply to an arbitrary non-Hermitian matrix that is not diagonalizable. In this work, we propose a simple yet powerful method to extend these ideas to arbitrary square matrices by acting on their eigenvalues. To this end, we introduce the notion of an n-regular block encoding, namely, a block encoding whose k-th power reproduces the k-th power of the encoded matrix for every 0 < k < n. We show that applying QSP to any unitary with this property is equivalent to applying a polynomial of degree at most n to the block-encoded matrix, independently of its internal structure. Moreover, we provide a simple construction that transforms any block encoding into an n-regular one using only O(n) ancillary qubits and operations. Finally, we show that this construction induces the desired transformation on the eigenvalues associated with the Jordan normal form of the matrix; for non-diagonalizable matrices this yields the full matrix function P(A), whose action on each Jordan block involves the derivatives of P, and not merely the map λ P(λ).

Sources

Related papers