The Quantum Singular Value Transformation (QSVT) is an algorithmic framework which encompasses a wide range of problems, including Hamiltonian dynamics, ground state preparation, Gibbs state preparation, amplitude amplification, and matrix inversion. For many areas of scientific interest such as quantum chemistry and condensed matter, the circuit depth of the block-encoding subroutine in QSVT can be formidable, limiting the scope of problems to which QSVT can be applied on early fault-tolerant hardware. We split linear operators A into a core operator A0 and a tail operator Aε, block-encoding the normalized sum of A0 and the average of nT randomly sampled terms from Aε. We show that using randomized block-encodings of this form allows one to lower the circuit depth of QSVT while incurring a diamond norm error that scales as O(CPδ2/nT), where CP is a prefactor which depends on the QSVT polynomial P and δ is the ratio of the ℓ1 norms of the coefficient vectors of Aε and A. We then go on to show that if one samples QSVT circuits from a joint probability distribution of both random block-encodings and random polynomials (Double-Randomized QSVT), then the circuit depth savings of these two randomization techniques combine multiplicatively, with only additive error contributions. We numerically calculate error bounds for long-range Heisenberg and iron-sulfur complex Hamiltonians, showing that for a diamond norm error bound of 10−3, RBE-QSVT can reduce the number of Pauli terms needed for the block-encoding by an order of magnitude or more in several cases. This paves the way for demonstrations of QSVT on early fault-tolerant hardware.
Introduction
The improvement of quantum computing hardware has accelerated in recent years, with below-threshold error correction [1], magic state distillation [2], and error-corrected/detected quantum algorithms being demonstrated on hardware within the past few years [3–5]. Despite these impressive advances, we are still seemingly quite far away from having computers that can run the computations involving upwards of millions of qubits of quantum operations on hundreds of logical qubits that would be needed to solve problems that classical computers cannot [6, 7]. These resource estimations typically assume the use of fault-tolerant algorithms with deep circuit depths. So-called "NISQ" algorithms such as the variational quantum eigensolver (VQE) [9, 10] [8] and sample-based diagonalization (SQD) [11, 12, 13, 14] have been investigated as a means of solving problems with shallow circuits. However, these heuristic methods often lack rigorous performance guarantees, making their near-term utility uncertain.
So-called "early fault-tolerant" algorithms [15, 16, 17] are typically constructed in such a way that rigorous runtime and error bounds can be proven in a similar fashion as their fully fault-tolerant counterparts while also reducing circuit depth. The trade-off typically made here is shallower circuits at the expense of worse asymptotic runtime scaling due to a larger sampling requirement.
Randomized compiling [18, 19, 20] and randomized algorithms [21, 22, 23, 24, 25, 26, 27, 28, 29] often naturally exhibit these trade-offs in similar spirit to Monte Carlo algorithms [30, 31], making them compelling candidates for the early fault-tolerant era. Randomized compiling has, for example, been well-established as a means of error-mitigation. [28, 32, 33] Indeed, it is common for randomized versions of algorithms to take approximation errors from O(ϵ) to O(ϵ2)[19, 29]. Quantum Signal Processing (QSP) and Quantum Singular Value Transformations (QSVT) [34, 35, 36, 37, 38, 39, 40, 41, 42, 43] are highly versatile fault-tolerant algorithmic frameworks which encompass a wide range of important scientific problems such as Hamiltonian dynamics [35, 36, 37, 39, 40, 44] and ground state preparation of quantum many-body system. [36, 45, 46, 47]. The main barrier to implementing QSVT on near-term hardware is the cost of the block-encoding itself. For example, writing A as a linear combination of N unitaries, one can block-encode A using a method such as linear combination of unitaries (LCU). [48] For problems such as quantum chemistry or other quantum many-body physics problems, N can scale polynomially with the number of qubits n, bringing the total circuit depth to O(poly(n)d).
To-date, works on randomized variants of QSVT have been scarce compared to e.g., randomized Trotterization methods [13, 25, 26, 49]. Ref. [22] established a single-ancilla randomized QSVT method which involves O(1/ϵ2) calls to circuits with depth O(λ2d2) for operators of 1-norm λ. Ref. [29] showed that one can use an effective QSVT degree deff≈d/2+O(1) without sacrificing accu racy simply by sampling from an ensemble of polynomial. Ref. [23] takes a conceptually similar approach as Ref. [29], but samples from individual Chebyshev polynomials according to their coefficients. Ref. [50] showed that by sparsifying a block-encoded Hamiltonian H using SparSto [49], one can lower the block-encoding cost of QSVT while incurring operator norm error O(dVarcoeff[H]), where Varcoeff[H] is a measure of the variance of the coefficient vector of the randomly sampled operators. Ref. [28] showed that averaging over ensembles of noisy QSP circuits can mitigate coherent over or under-rotation errors. Ref. [51] studied the robustness of QSP with respect to errors in the block-encoded operator and formulated algorithms which tailor the choice of polynomial to a given error.
In this work, we present a randomized QSVT method based on randomized block-encodings of random perturbations of a linear operator A0 approximating the target operator A. That is, we express A as a linear combination of operators with unit norm, partition it into a "core" A0 and a "tail" Aϵ comprised of L0 and Lϵ, respectively and coefficient vector norms λ0 and λϵ, respectively. Each QSVT circuit run performs a polynomial transformation of an operator comprised of the sum of A0 and an operator proportional to the average of nT randomly sampled terms from Aϵ. We show that one can bound the diamond norm error of this randomized QSVT method as
nTCP(λλϵ)2
where λ is the norm of the coefficient vector of A and CP is a prefactor which depends on the choice of polynomial. We then show by combining this randomized block-encoding method with the random polynomial sampling method of Ref. [29] in a joint distribution, one can inherit the circuit reduction benefits of both randomization techniques with error bounded by the sum of that of each method individually. We dub this method "double-randomized QSVT" (DR-QSVT). Importantly, the circuit depth of both methods scales as O((L0+nT)d), avoiding the O(d2) scaling of single-ancilla randomized QSVT [22] while still achieving a lower circuit depth than conventional QSVT.
We note that the method of randomization used in this work is conceptually similar to that of Ref. [50]. We note the following important key differences of our work:
Whereas Ref. [50] uses an ensemble of QSVT circuits such that each circuit layer uses an independently-sampled block-encoded operator, we take the block-encoding to be that of the same randomly-sampled operator at each layer within a given circuit run. This lets us establish (λλϵ)2 error scaling using existing lemmas as opposed to triangle inequality arguments. Furthermore, we can characterize the degree to which randomization is beneficial on a polynomial by polynomial basis through the prefactor CP.
We do not use SparSto-style randomization [49], which allows us to make the number of Pauli terms in each block-encoding a constant as opposed to a random variable.
We derive error bounds for the case where we perform joint sampling of both random polynomials and random block-encoded operators (double-randomized QSVT).
The rest of this work is organized as follows. In Sec. II we briefly summarize randomized compiling and QSVT. In Sec. III we present the randomized QSVT method of this work, including error bounds for QSVT with randomized block-encodings (RBE-QSVT) and double-randomized QSVT (DR-QSVT). Numerical examples are given in Sec. V for spin lattice models and quantum chemistry problems with representative applications. In Sec. VI we conclude the work.
Background
Here we briefly summarize key prerequisite information for the rest of the work.
QSVT
On a high-level QSVT [36, 37, 38] is a method which takes as an input block-encodings of operators A∈Cm×n such that ∥A∥≤1:
Be[A]=[A∗∗∗](1)
and outputs block-encodings of polynomial transformations P(SV)(A) of A:
P(SV)(A)={∑k=1min(m,n)P(σk)∣ΛkL⟩⟨ΛkR∣∑k=1nP(σk)∣ΛkR⟩⟨ΛkR∣if P is oddif P is even(3)
where σk are the singular values of A, and ∣ΛkL⟩ and ∣ΛkR⟩ are the left and right singular vectors of A, and P is our target polynomial. The class of preparable polynomials P(x) must obey some mild constraints. Most importantly, they must be of definite parity (even or odd) and satisfy ∣P(x)∣≤1 on the interval [−1,1]. Block-encoding a degree-d polynomial using QSVT requires d calls to Be[A] interleaved between projector-controlled rotation gates. The circuit is of the form:
for odd d[36]. Here Πϕj and Πϕj are shorthand for the operations eiϕj(2Π−I) and eiϕj(2Π−I), where Π and Π are projectors onto the left and right singular vector spaces of A, respectively. The phase factors {ϕk} can be efficiently computed classically as a pre-processing step [52, 53, 54]. It is often the case that what we want is the real part of the polynomial and we consider the imaginary part to be extraneous. In this case we can perform a linear combination of two QSVT circuits with one additional ancilla qubit as: [37]
UReϕ=(H⊗I)(∣0⟩⟨0∣⊗Uϕ+∣1⟩⟨1∣⊗U−ϕ)(H⊗I)(6)
where H is the Hadamard gate in this context. We would then obtain the block-encoded polynomial by postselecting on the ∣0⟩ state of the ancilla as well as the ancillas already present in the QSVT circuits.
Randomized Compiling
Naively, one might be tempted to think that the concept of randomizing quantum circuit does not make much sense. After all, we already spend a lot of resources suppressing other sources of randomization such as gate errors. Why purposefully inject randomization? Would we not make the circuit less accurate?
This intuition turns out to be incorrect in general. This (hypothetical) misconception stems from the unrealistic assumption that we had access to the exact unitary U that we wish to run in the first place. In reality, we almost always have to resort to ϵ-approximations Uϵ of U. That is, if we have two quantum channels U(ρ)=UρU† and Uϵ(ρ)=UϵρUϵ†, then:
∥U−Uϵ∥⋄≤ϵ.(7)
The key insight of randomized compilation (and by extension many randomized algorithms) is that instead of implementing Uϵ, we introduce classical mixing, randomly sampling unitaries Uj with probability pj from some predetermined ensemble. If we choose this ensemble judiciously, then we can bound the diamond norm error of the resulting channel [18, 19]:
M(ρ)=j∑pjUjρUj†(8)
by
∥M−U∥⋄=O(ϵ2).(9)
The diamond norm of a quantum channel Φ acting on a Hilbert space H is defined as:
∥Φ∥⋄=ρsup∥ρ∥1∥(Φ⊗In)(ρ)∥1(10)
where n=dim(H) and ∥⋅∥1 is the trace norm defined as tr(X†X) for operator X. This quadratic suppression in error is the main finding of Hastings [18] and Campbell [19]. Typically, {Uj} are chosen to be easier to prepare than U.
Methods
QSVT With Randomized Block-Encodings
In this section we consider a type of randomization where we have some linear operator A0 which approximates a full operator A∈Cm×n. The full operator is a sum of many terms with norm 1:
References
[9]J. Tilly, H. Chen, S. Cao, D. Picozzi, K. Setia, Y. Li, E. Grant, L. Wossnig, I. Rungger, G. H. Booth, and J. Tennyson, The Variational Quantum Eigensolver: A review of methods and best practices, Physics Reports 986, 1 (2022).DOI
[10]H. R. Grimsley, S. E. Economou, E. Barnes, and N. J. Mayhall, An adaptive variational algorithm for exact molecular simulations on a quantum computer, Nature Communications 10, 3007 (2019).DOI
[11]J. Robledo-Moreno, M. Motta, H. Haas, A. Javadi-Abhari, P. Jurcevic, W. Kirby, S. Martiel, K. Sharma, S. Sharma, T. Shirakawa, I. Sitdikov, R.-Y. Sun, K. J. Sung, M. Takita, M. C. Tran, S. Yunoki, and A. Mezzacapo, Chemistry beyond the scale of exact diagonalization on a quantum-centric supercomputer, Science Advances 11, eadu9991 (2025), publisher: American Association for the Advancement of Science.DOI
[12]J. Yu, J. R. Moreno, J. T. Josue, L. Bertels, D. Claudino, B. Fuller, P. Groszkowski, T. S. Humble, P. Jurcevic, W. Kirby, T. A. Maier, M. Motta, B. Pokharel, A. Seif, A. Shehatа, K. J. Sung, M. C. Tran, V. Trapali, A. Mezzacapo, and K. Sharma, Quantum-Centric Algorithm for Sample-Based Krylov Diagonalization (2025), arXiv:2501.09702 [quant-ph].arxiv.org/abs/2501.09702
[13]S. Piccinelli, A. Baiardi, S. Barison, M. Rossmannek, A. C. Vazquez, F. Tacchino, S. Mensa, E. Altamura, A. Alavi, M. Motta, J. Robledo-Moreno, W. Kirby, K. Sharma, A. Mezzacapo, and I. Tavernelli, Quantum chemistry with provable convergence via randomized sample-based Krylov quantum diagonalization (2025).arxiv.org/abs/2508.02578
[14]J. Bierman and Y. Liu, Towards utility-scale electronic structure with sample-based quantum bootstrap embedding, Digital Discovery 5, 945 (2026), https://pubs.rsc.org/d/articles-pdf/d5dd00416k.pdf.DOI
[15]A. Katabarwa, K. Gratsea, A. Caesura, and P. D. Johnson, Early fault-tolerant quantum computing, PRX Quantum 5, 020101 (2024).DOI
[16]Y. Dong, L. Lin, and Y. Tong, Ground-State Preparation and Energy Estimation on Early Fault-Tolerant Quantum Computers via Quantum Eigenvalue Transformation of Unitary Matrices, PRX Quantum 3, 040305 (2022), publisher: American Physical Society.DOI
[17]D. Wang, O. Higgott, and S. Brierley, Accelerated Variational Quantum Eigensolver, Physical Review Letters 122, 140504 (2019), publisher: American Physical Society.DOI
[18]M. B. Hastings, Turning gate synthesis errors into incoherent errors (2016), arXiv:1612.01011 [quant-ph].arxiv.org/abs/1612.01011
[19]E. Campbell, Shorter gate sequences for quantum computing by mixing unitaries, Physical Review A 95, 042306 (2017), publisher: American Physical Society.DOI
[20]O. Maupin, A. D. Burch, C. G. Yale, M. N. H. Chow, T. Colvin, B. Ruzic, M. C. Revelle, B. K. McFarland, E. Ibarra-García-Padilla, A. Rascon, A. J. Landahl, S. M. Clark, and P. J. Love, Solovay-kitaev algorithm and randomized compilation, Phys. Rev. A 113, 062418 (2026).DOI
[21]K. Wan, M. Berta, and E. T. Campbell, Randomized Quantum Algorithm for Statistical Phase Estimation, Physical Review Letters 129, 050305 (2022).DOI
[22]X. Wang, Y. Zhang, S. Hazra, T. Li, C. Shao, and S. Chakraborty, Randomized Quantum Singular Value Transformation (2025), arXiv:2510.06851 [quant-ph].arxiv.org/abs/2510.06851
[23]A. Tosta, T. de Lima Silva, G. Camilo, and L. Aolita, Randomized semi-quantum matrix processing, npj Quantum Information 10, 93 (2024), publisher: Nature Publishing Group.DOI
[24]I. D. Kivlichan, C. E. Granade, and N. Wiebe, Phase estimation with randomized Hamiltonians (2019), arXiv:1907.10070 [quant-ph].arxiv.org/abs/1907.10070
[25]K. Nakaji, M. Bagherimehrab, and A. Aspuru-Guzik, High-order randomized compiler for hamiltonian simulation, PRX Quantum 5, 020330 (2024).DOI
[26]E. Campbell, Random Compiler for Fast Hamiltonian Simulation, Physical Review Letters 123, 070503 (2019), publisher: American Physical Society.DOI
[27]S. Wang, S. McArdle, and M. Berta, Qubit-efficient randomized quantum algorithms for linear algebra, PRX Quantum 5, 020324 (2024).
[28]S. Liu, Y. Dong, D. An, and M. Y. Niu, Ensemble-Based Quantum Signal Processing for Error Mitigation (2026), arXiv:2601.20073 [quant-ph] version: 1.arxiv.org/abs/2601.20073
[29]J. M. Martyn and P. Rall, Taming the cost of quantum algorithms with randomization, npj Quantum Information 11, 47 (2025), publisher: Nature Publishing Group.DOI
[30]M. H. Kalos and P. A. Whitlock, Monte carlo methods (John Wiley & Sons, 2008).
[31]J. Gubernatis, N. Kawashima, and P. Werner, Quantum Monte Carlo Methods (Cambridge University Press, 2016).
[32]K. Temme, S. Bravyi, and J. M. Gambetta, Error Mitigation for Short-Depth Quantum Circuits, Physical Review Letters 119, 180509 (2017), publisher: American Physical Society.DOI
[33]Y. Suzuki, S. Endo, K. Fujii, and Y. Tokunaga, Quantum Error Mitigation as a Universal Error Reduction Technique: Applications from the NISQ to the Fault-Tolerant Quantum Computing Era, PRX Quantum 3, 010345 (2022), publisher: American Physical Society.DOI
[34]G. H. Low, T. J. Yoder, and I. L. Chuang, Methodology of resonant equiangular composite quantum gates, Physical Review X 6, 041067 (2016).DOI
[35]G. H. Low and I. L. Chuang, Optimal Hamiltonian Simulation by Quantum Signal Processing, Physical Review Letters 118, 010501 (2017), publisher: American Physical Society.DOI
[36]J. M. Martyn, Z. M. Rossi, A. K. Tan, and I. L. Chuang, Grand unification of quantum algorithms, PRX Quantum 2, 040203 (2021).DOI
[37]A. Gilyén, Y. Su, G. H. Low, and N. Wiebe, Quantum singular value transformation and beyond: exponential improvements for quantum matrix arithmetics, in Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing (2019) pp. 193–204, arXiv:1806.01838v1 [quant-ph].DOI
[38]K. J. Joven, E. R. Das, J. Bierman, A. Majumdar, M. H. Heris, and Y. Liu, Scalable quantum computational science: A perspective from block-encodings and polynomial transformations, APL Computational Physics 2, 010901 (2026).arxiv.org/abs/2511.16738
[39]D. Motlagh and N. Wiebe, Generalized Quantum Signal Processing, PRX Quantum 5, 020368 (2024), publisher: American Physical Society.arxiv.org/abs/2308.01501
[40]D. W. Berry, D. Motlagh, G. Pantaleoni, and N. Wiebe, Doubling the efficiency of hamiltonian simulation via generalized quantum signal processing, Phys. Rev. A 110, 012612 (2024).DOI
[41]Z. M. Rossi and I. L. Chuang, Multivariable quantum signal processing (M-QSP): prophecies of the two-headed oracle, Quantum 6, 811 (2022), publisher: Verein zur Förderung des Open Access Publizierens in den Quantenwissenschaften.DOI
[42]Y. Dong, L. Lin, H. Ni, and J. Wang, Infinite quantum signal processing, Quantum 8, 1558 (2024), publisher: Verein zur Förderung des Open Access Publizierens in den Quantenwissenschaften.
[43]X. Lu, Y. Liu, and H. Lin, Quantum Signal Processing and Quantum Singular Value Transformation on \(U(N)\), Quantum 10, 2048 (2026), publisher: Verein zur Förderung des Open Access Publizierens in den Quantenwissenschaften.arxiv.org/abs/2408.01439
[44]J. M. Martyn, Y. Liu, Z. E. Chin, and I. L. Chuang, Efficient fully-coherent quantum signal processing algorithms for real-time dynamics simulation, The Journal of Chemical Physics 158, 024106 (2023).DOI
[45]L. Lin and Y. Tong, Optimal polynomial based quantum eigenstate filtering with application to solving quantum linear systems, Quantum 4, 361 (2020), publisher: Verein zur Förderung des Open Access Publizierens in den Quantenwissenschaften.DOI
[46]L. Lin and Y. Tong, Near-optimal ground state preparation, Quantum 4, 372 (2020).
[47]Y. Dong and L. Lin, Multi-level quantum signal processing with applications to ground state preparation using fast-forwarded hamiltonian evolution, arXiv:2406.02086 (2024).arxiv.org/abs/2406.02086
[48]A. M. Childs and N. Wiebe, Hamiltonian simulation using linear combinations of unitary operations, arXiv:1202.5822 (2012).DOI
[49]Y. Ouyang, D. R. White, and E. T. Campbell, Compilation by stochastic Hamiltonian sparsification, Quantum 4, 235 (2020), arXiv:1910.06255 [quant-ph].DOI
[50]F. Paganelli, M. Grossi, A. Giachero, T. E. O’Brien, and O. Kiss, When randomization advantageous in quantum simulation? (2026), arXiv:2604.07448 [quant-ph].arxiv.org/abs/2604.07448
[51]T. O. Maciel, G. Camilo, A. Tosta, D. Stilck-França, L. Aolita, and T. d. L. Silva, Robustness of Quantum Signal Processing and Resource-aware Polynomial Design (2026), arXiv:2601.03885 [quant-ph].
[52]H. Ni and L. Ying, Fast Phase Factor Finding for Quantum Signal Processing (2024), arXiv:2410.06409.arxiv.org/abs/2410.06409
[53]Y. Dong, X. Meng, K. B. Whaley, and L. Lin, Efficient phase-factor evaluation in quantum signal processing, Physical Review A 103, 042419 (2021), publisher: American Physical Society.DOI
[54]S. E. Skelton, The Hitchhiker’s Guide to QSP pre-processing (2025), arXiv:2501.05977 [quant-ph].arxiv.org/abs/2501.05977
[55]J. Watrous, The Theory of Quantum Information (Cambridge University Press, 2018).DOI
[56]M. Motta, E. Ye, J. R. McClean, Z. Li, A. J. Minnich, R. Babbush, and G. K.-L. Chan, Low rank representations for quantum simulation of electronic structure, npj Quantum Information 7, 83 (2021).DOI
[57]S. Lee, J. Lee, H. Zhai, Y. Tong, A. M. Dalzell, A. Kumar, P. Helms, J. Gray, Z.-H. Cui, W. Liu, M. Kastoryano, R. Babbush, J. Preskill, D. R. Reichman, E. T. Campbell, E. F. Valeev, L. Lin, and G. K.-L. Chan, Evaluating the evidence for exponential quantum advantage in ground-state quantum chemistry, Nature Communications 14, 1952 (2023).DOI
[58]G. H. Low, R. King, D. W. Berry, Q. Han, A. E. DePrince, A. F. White, R. Babbush, R. D. Somma, and N. C. Rubin, Fast quantum simulation of electronic structure by spectral amplification, Phys. Rev. X 15, 041016 (2025).DOI
[59]S. Kanasugi, R. Toshio, K. Maruyama, and H. Oshima, Enabling chemically accurate quantum phase estimation in the early fault-tolerant regime, arXiv preprint arXiv:2603.22778 (2026).arxiv.org/abs/2603.22778
[60]D. Wecker, B. Bauer, B. K. Clark, M. B. Hastings, and M. Troyer, Gate-count estimates for performing quantum chemistry on small quantum computers, Physical Review A 90, 022305 (2014).DOI
[61]V. Bergholm, J. Izaac, M. Schuld, C. Gogolin, S. Ahmed, V. Ajith, M. S. Alam, G. Alonso-Linaje, B. Akash-Narayanan, A. Asadi, et al., PennyLane: Automatic differentiation of hybrid quantum-classical computations, arXiv preprint arXiv:1811.04968 (2018).arxiv.org/abs/1811.04968
[62]Z. Li and G. K.-L. Chan, Spin-projected matrix product states: Versatile tool for strongly correlated systems, Journal of Chemical Theory and Computation 13, 2681 (2017).DOI
[63]E. B. Davies, Lipschitz continuity of functions of operators in the Schatten classes, Journal of the London Mathematical Society s2-37, 148 (1988).DOI
[64]A. M. Childs, Y. Su, M. C. Tran, N. Wiebe, and S. Zhu, Theory of trotter error with commutator scaling, Physical Review X 11, 011020 (2021).DOI
[65]E. Kissin and V. S. Shulman, Classes of operator-smooth functions. i. operator-lipschitz functions, Proceedings of the Edinburgh Mathematical Society 48, 151–173 (2005).
[66]A. B. Aleksandrov and V. V. Peller, Operator lipschitz functions, Russian Mathematical Surveys 71, 605 (2016).DOI
[67]N. J. Higham and S. D. Relton, Higher order fréchet derivatives of matrix functions and the level-2 condition number, SIAM Journal on Matrix Analysis and Applications 35, 1019 (2014), https://doi.org/10.1137/130945259.
[68]DLMF, NIST Digital Library of Mathematical Functions, https://dlmf.nist.gov/, Release 1.2.8 of 2026-09-15, f. W. J. Olver, A. B. Olde Daalhuis, D. W. Lozier, B. I. Schneider, R. F. Boisvert, C. W. Clark, B. R. Miller, B. V. Saunders, H. S. Cohl, and M. A. McClain, eds.
[69]F. N. Fritsch and R. E. Carlson, Monotone piecewise cubic interpolation, SIAM Journal on Numerical Analysis 17, 238 (1980), https://doi.org/10.1137/0717021.