🧬 biology

Quantum oblique eigenprojection

This paper presents a quantum algorithm that efficiently performs oblique eigenprojections using two-sided block preconditioning and Fourier-transformed resolvents, enabling significant speedups in preparing eigenstates with complex spectra and solving algebraic Riccati and Sylvester equations.

Original authors: Yuan Su, Alexander Dalzell

Published 2026-10-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuan Su, Alexander Dalzell

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Quantum computers promise to solve problems that would take classical machines an impractical amount of time, particularly those involving the behavior of complex systems like molecules or materials. To do this, these machines often rely on a process called projection, which is essentially a way of filtering a chaotic mixture of possibilities down to a single, specific state of interest. Imagine a system with many different energy levels; the goal is often to isolate the lowest one, known as the ground state, to understand how a material behaves. For decades, researchers have developed highly efficient methods to perform this filtering when the system behaves in a predictable, symmetric way. However, many real-world systems, especially those involving chemical reactions or irreversible processes, do not follow these simple rules. Their internal structures are skewed and asymmetric, making the standard filtering tools ineffective or incredibly expensive to use.

A team of researchers at the AWS Center for Quantum Computing has now demonstrated a way to overcome this limitation. They have shown that a quantum computer can successfully perform this filtering operation on these complex, asymmetric systems without the massive cost penalty that previously made the task impractical. The researchers developed a new technique that allows the computer to isolate specific parts of a system's behavior, even when those parts are not neatly separated or aligned. This achievement is significant because it extends the power of quantum algorithms to a much wider range of scientific problems, including those found in quantum chemistry and control theory, where the systems being studied are inherently messy and non-symmetric.

The core challenge the team addressed involves how quantum computers handle matrices, which are grids of numbers used to describe the state of a system. In many cases, these grids can be broken down into different sections, or subspaces, each associated with a specific set of values. When these sections are perfectly aligned, standard quantum tools can easily isolate one section from the others. But in the complex systems the researchers were targeting, these sections are tilted or "oblique" relative to each other. Previous attempts to isolate these tilted sections relied on a method that worked, but it introduced a massive inefficiency. The process required a normalization factor—a scaling number that ensures the calculation remains stable—that grew drastically as the system became more difficult to separate. This scaling factor acted like a heavy tax on the computation, forcing the computer to repeat the process many more times than necessary, effectively canceling out the speed advantage quantum computers are supposed to offer.

The researchers solved this by changing how they looked at the data inside the quantum computer. Instead of treating the calculation as a single, monolithic block, they realized that the internal structure of the calculation contained a hidden pattern that could be exploited. By applying a specific mathematical transformation known as a discrete Fourier transform, they were able to rearrange the data into a new format where the different parts of the calculation were clearly visible and organized. In this new arrangement, the researchers could see that the unwanted parts of the calculation were not just random noise, but followed a predictable, repeating structure. They then applied a two-sided preconditioning technique, which is essentially a way of adjusting the weights of different parts of the calculation to balance them out. This adjustment suppressed the massive scaling factor that had plagued previous methods, bringing it down to a level that is nearly as small as theoretically possible.

The result is a method that is nearly as efficient as the best existing techniques for simple, symmetric systems, but now works for the complex, asymmetric ones that are common in nature. The researchers proved that their approach requires a number of computational steps that grows almost linearly with the difficulty of the separation, rather than the quadratic or cubic growth seen in older methods. This improvement is not just a theoretical curiosity; it directly translates to faster and more practical algorithms for real-world applications. For instance, the team showed how their method could be used to prepare specific quantum states for systems with complex energy levels, a task that was previously much slower. They also demonstrated that it could solve certain types of complex equations used in control theory and quantum chemistry significantly faster than before, reducing the computational cost by factors of two or even three in some cases.

This work suggests a promising new path for applying quantum computers to non-analytic functions, which are mathematical operations that cannot be smoothly approximated by simple polynomials. By successfully isolating these oblique projections, the researchers have removed a major barrier that prevented quantum algorithms from tackling a broad class of difficult problems. Their findings indicate that with the right structural insights, quantum computers can handle the messy, real-world complexity of non-symmetric systems with an efficiency that was previously thought to be out of reach. The ability to perform these projections with minimal overhead means that future quantum algorithms can be designed to be more robust and practical, potentially accelerating discoveries in fields ranging from drug design to materials science.

Technical Summary: Quantum Oblique Eigenprojection

Problem Statement

The paper addresses the challenge of performing oblique eigenprojections on a quantum computer. While quantum algorithms for Hermitian matrices (where eigenspaces are orthogonal) are well-established using frameworks like Quantum Singular Value Transformation (QSVT), general non-normal matrices present a significant hurdle. For non-normal matrices, the Hilbert space decomposes into generalized, non-orthogonal eigensubspaces. The task is to construct a quantum algorithm that projects an input state onto a specific invariant subspace associated with a target region of the complex spectrum (e.g., eigenvalues inside a unit disk), while suppressing components associated with the complementary region.

This problem is fundamentally obstructed when the spectrum is complex. Unlike the real spectrum case, where polynomials can approximate the indicator function of a target region, the maximum modulus principle prevents analytic functions (and thus polynomials) from approximating such an indicator in the complex plane without significant error. Consequently, standard QSVT and Quantum EigenValue Transformation (QEVT) techniques, which rely on polynomial approximations of analytic functions, cannot directly solve this problem.

Previous approaches, such as resolvent integration (using Cauchy's integral formula), can handle non-analytic functions by integrating the matrix resolvent $(\zeta I - A)^{-1}$ along a contour enclosing the target spectrum. However, these methods suffer from a critical inefficiency: they incur a normalization factor $\alpha$ that scales as $\Theta(\|\Pi\|/\delta)$ , where $\|\Pi\|$ is the norm of the projection and $\delta$ is the spectral gap. This inflated normalization factor significantly increases the query complexity of any algorithm using the projection as a subroutine, particularly because the success probability of the block encoding scales inversely with $\alpha^2$ .

Methodology

The authors propose a novel approach that improves upon the standard resolvent integration method by exploiting the full structure of the block encoding generated by the discretized resolvent integral.

Discretized Resolvent Integral: The oblique eigenprojection $\Pi$ is approximated by discretizing the contour integral of the resolvent using the trapezoidal rule. This yields a sum of shifted resolvents: $\Pi \approx \frac{1}{n} \sum_{j=0}^{n-1} (I - \omega^{-j}_n A)^{-1}$ .
Block Encoding Construction: The standard approach constructs a block-diagonal matrix $M$ containing these shifted resolvents and inverts it using QSVT. By applying a Hadamard transform to the ancilla register, the top-left block of the resulting inverse matrix approximates $\Pi$ . However, the remaining blocks are left unstructured, forcing the normalization factor to be determined by the worst-case norm of the entire matrix, which is $\Theta(1/\delta)$ .
Discrete Fourier Transform (DFT) Analysis: The key insight of this work is that the blocks of the inverted matrix $M^{-1}$ $M^{- 1}$ are not arbitrary; they possess a circulant structure. By conjugating $M^{-1}$ $M^{- 1}$ with the Quantum Fourier Transform (QFT) instead of the Hadamard transform, the authors reveal that every block $(k, l)$ $(k, l)$ corresponds to a specific discrete Fourier mode $S_m$ $S_{m}$ of the shifted resolvent.
- The zeroth mode $S_0$ approximates the eigenprojection $\Pi$ .
- Higher-order modes $S_m$ decay geometrically under spectral-set conditions (e.g., if the matrix is diagonalizable or has a bounded numerical range).
Two-Sided Block Preconditioning: The authors introduce a two-sided block preconditioning technique in the Fourier basis. They apply a scaling operator $T$ $T$ (diagonal in the computational basis) that modifies the matrix before inversion.
- As the scaling parameter $t$ decreases, the first block row and column of the preconditioned matrix grow.
- However, this preconditioning suppresses the inverse-gap growth of the normalization factor of the extracted top-left block, reducing it from $\Theta(\|\Pi\|/\delta)$ to $O(c \|\Pi\|)$ , where $c$ is a constant depending on the spectral-set condition (e.g., $c = 1+\sqrt{2}$ for numerical range, or the condition number for diagonalizable matrices).
- This normalization is nearly optimal, as $\|\Pi\|$ is the theoretical lower bound for any block encoding of $\Pi$ .

Key Results

The paper establishes three main theorems:

Theorem 1 (DFT of Matrix Resolvent): Provides a closed-form expression for the blocks of the resolvent integral's discretization in the Fourier basis. It shows that the blocks are determined by the restrictions of the matrix to the inside and outside invariant subspaces, decaying geometrically with the power of the shift.
Theorem 2 (Unit Disk Projection): Demonstrates that for a matrix $A$ with spectrum disjoint from the unit circle, the oblique eigenprojection $\Pi$ onto the inside subspace can be block encoded with normalization factor $\alpha_\Pi = O(c \|\Pi\|)$ and query complexity $O\left(\frac{c \alpha_A \|\Pi\|}{\delta} \log \frac{c \|\Pi\|}{\varepsilon} \log \frac{1}{\varepsilon}\right)$ . This matches the query complexity of the standard approach but drastically improves the normalization factor.
Theorem 3 (General Regions): Extends the result to arbitrary target regions in the complex plane. By using an analytic separating map $f$ (e.g., a rational function mapping the target region to the unit disk), the problem is reduced to the unit disk case applied to $f(A)$ . The cost depends on the gap of $f(A)$ to the unit circle.

Applications

The authors demonstrate that this improved primitive leads to significant speedups in three specific applications:

Eigenstate Preparation for Complex Spectra: Extending the QEVT algorithm of Low and Su to non-normal matrices with complex eigenvalues. The new method prepares eigenstates with a query complexity scaling as $O(\|\Pi\|^2 / (\eta \delta))$ , a quadratic improvement over the $O(\|\Pi\|^2 / (\eta \delta^2))$ scaling of the standard resolvent integral approach. This improvement arises because the number of amplitude amplification rounds becomes independent of the gap (scaling as $O(\|\Pi\|/\eta)$ ), whereas the standard approach requires $O(\|\Pi\|/(\eta \delta))$ rounds.
Continuous-Time Algebraic Riccati Equations: Solving these equations, which arise in optimal control and quantum chemistry, reduces to projecting onto an invariant subspace of a non-normal Hamiltonian. The new method cubically improves the gap dependence compared to the prior solver by Rodenas-Ruiz, Zhao, and Lee, removing two factors of $1/\delta$ from the total complexity.
Ordinary Sylvester Equations: Solving $AX + XB = C$ by projecting onto the invariant subspace of a block matrix. The method quadratically improves the gap dependence of the direct augmented method of Wang and Liu.

Significance and Claims

The paper claims that this work provides a fundamental primitive for quantum algorithms involving non-normal matrices. By achieving a normalization factor close to the optimal $\|\Pi\|$ while maintaining near-linear query complexity in the inverse gap, the authors bridge the gap between the efficiency of QSVT for Hermitian matrices and the capabilities required for non-Hermitian problems.

The authors emphasize that their result is significant because:

It resolves the issue of "inflated normalization factors" that have plagued previous resolvent-based methods for non-analytic functions.
It demonstrates that non-analytic matrix functions (specifically oblique eigenprojections) can be implemented on quantum computers with complexity analogous to QSVT, despite the inability to use polynomial approximations.
The technique of analyzing the full block structure via DFT and applying two-sided preconditioning offers a new paradigm for optimizing block encodings, potentially applicable to other non-analytic functions.

The authors remain modest regarding the scope, noting that the preconditioning relies on the specific closed-form DFT of the resolvent and does not immediately apply to all constructions (e.g., the product-of-resolvents approach of Wang and Liu). They also note that the cost is governed by the spectral gap of the input matrix, unlike QSVT for Hermitian matrices where cost is determined solely by the target function's geometry.

A scientific accuracy reviewer checked the draft against the paper and flagged these problems:

Summary claims preconditioning amplifies the top-left block; paper states it suppresses the inverse-gap growth of the normalization factor. (the paper says: "preconditioning against it suppresses the inverse-gap growth of the normalization factor down to nearly ∥Π∥")
Summary claims scaling operator suppresses off-diagonal blocks; paper states it causes the first block row/column to grow. (the paper says: "As t decreases, the first block row and column grow")
Summary claims the method quadratically improves eigenstate preparation; paper states it is independent of gap in rounds, yielding quadratic improvement in total queries. (the paper says: "our initial state is prepared about O(∥Π∥/η) times independent of the gap")

Produce a corrected version of the draft. Fix ONLY what the reviewer flagged (verify each point against the paper) and keep everything else — the structure, the level of detail, the terminology — unchanged. Output ONLY the corrected summary.