First-hand information for everyone
This paper presents two new quantum algorithms for solving linear systems that avoid the traditional dependence on the spectral condition number. The condition number (denoted κ) is a common worst-case measure of how hard a matrix inversion is. The authors produce the normalized quantum solution |x⟩ of Ax = |b⟩ to accuracy ε with runtimes that do not scale directly with κ. Instead the costs depend on more refined, instance-specific quantities that can be much smaller than κ for typical inputs.
The work assumes the standard quantum input model: the matrix A is accessed through a block-encoding oracle (a common way to feed a matrix into a quantum computer) and the right-hand side |b⟩ can be prepared by a unitary. The first algorithm is truncation-based. It replaces the full matrix inverse by a truncated version that ignores very small singular values. This leads to an “effective condition number” κ_eff which governs the cost. The authors prove a family of upper bounds on κ_eff. Concretely, for different integer choices t the bound trades off sensitivity to small singular values against the inverse error term (1/ε)^{1/t}. For t = 1 one gets the simple bound κ_eff ≤ ||A^{-1†}|x⟩|| / (||x|| ε), where ||·|| denotes the usual vector norm and A^{-1†} is a matrix-related vector that appears in the analysis. The truncation solver uses an optimal number of queries to prepare |b⟩ and roughly O(κ_eff · polylog(κ_eff/ε)) queries to the block-encoding of A.
The second algorithm is filtering-based and is especially simple. When the norm of the solution ||x|| is known up to a constant, its leading-order query cost to both the matrix oracle and the state-preparation oracle is about 6 · (||A^{-1†}|x⟩|| / ||x||) · (1/ε) · ln(1/ε). The paper also gives a comparably simple estimator for the solution norm with the same asymptotic cost up to logarithmic factors, so the method can be used even when ||x|| is not known in advance. Compared with a prior beyond-κ algorithm by Li, which had cost scaling like 1/ε^2 in important regimes, these new solvers lower the dependence on the inverse accuracy and improve the constant factors.
Why this matters: the spectral condition number κ often overestimates how hard a particular instance is. By replacing κ with an effective condition number or a vector-dependent norm, the new methods can solve many ill-conditioned systems faster in practice. This is useful because quantum linear-system solvers are building blocks for other quantum algorithms in physics and numerical analysis. Making ill-conditioned problems more feasible on quantum hardware brings a wider class of problems closer to possible quantum advantage.
Important caveats and uncertainties: the improved cost depends on instance-specific vector norms such as ||A^{-1†}|x⟩||/||x||, which can themselves be large. In the worst-case these quantities approach values tied to κ, and the algorithms’ costs then degrade toward the usual κ-dependent scaling. The truncation approach also involves a tradeoff parameter t that moves dependence between the effective condition number and the accuracy term (1/ε)^{1/t}. The results are stated in the block-encoding and a new affine-dilation input model; practical speedups therefore depend on whether those input assumptions fit a given application. Finally, the excerpt provided here is a portion of the full paper; the full text contains detailed proofs, parameter choices, and further technical comparisons that determine exact costs in specific settings.