Skip to content

Givens rotations and Householder reflectors: a history

Two orthogonal transformations do most of the work in dense numerical linear algebra. A Givens rotation zeroes one entry by rotating two rows. A Householder reflector zeroes a whole column below the diagonal by reflecting it through a hyperplane. QR, least squares, Hessenberg and tridiagonal reduction, the QR algorithm for eigenvalues, the SVD and square-root Kalman filters are all built from one or the other.

This page traces both from their origins to the modern safe-scaling algorithms, then surveys what is known about them in arithmetic other than IEEE binary32 and binary64. It is the background for the epic #666 and its studies in mp-numerics. The last section says where MTL5’s own implementations stand. The geometry is animated in givens-householder-pictures.md.

Every reference below was checked against Crossref, OpenAlex, arXiv or the publisher, and its DOI is given where one exists. A few details could not be confirmed. They are kept, because they are part of the story as it is usually told, but each is marked (unconfirmed) where it appears and is collected in Unconfirmed details at the end.

1. Plane rotations before computers: Jacobi, 1846

Section titled “1. Plane rotations before computers: Jacobi, 1846”

C. G. J. Jacobi diagonalized a symmetric matrix by a sequence of plane rotations, each chosen to annihilate one off-diagonal pair:

G(i,j,θ)TA G(i,j,θ),G=(cs−sc) in rows and columns i,j,c2+s2=1.G(i,j,\theta)^T A\, G(i,j,\theta), \qquad G = \begin{pmatrix} c & s \\ -s & c \end{pmatrix} \text{ in rows and columns } i, j,\quad c^2 + s^2 = 1 .

A later rotation can refill an entry an earlier one zeroed, so the method iterates, but the off-diagonal sum of squares falls monotonically and the iteration converges. Jacobi used it by hand to compute secular perturbations in astronomy.

  • C. G. J. Jacobi, “Über ein leichtes Verfahren, die in der Theorie der Säcularstörungen vorkommenden Gleichungen numerisch aufzulösen”, J. reine angew. Math. 30 (1846) 51–94. doi:10.1515/crll.1846.30.51

2. Givens: rotations as a direct method, 1954–1958

Section titled “2. Givens: rotations as a direct method, 1954–1958”

Wallace Givens, at Oak Ridge, turned Jacobi’s iteration into a finite algorithm. Instead of chasing every off-diagonal entry, Givens chose an order of rotations under which no zero is refilled. A symmetric matrix then reaches tridiagonal form in a fixed number of steps, after which its eigenvalues are found by Sturm-sequence bisection. The 1958 paper applied the same idea to a general matrix, reducing it to triangular form: QR by rotations.

  • W. Givens, Numerical Computation of the Characteristic Values of a Real Symmetric Matrix, Oak Ridge National Laboratory report ORNL-1574, 3 March 1954. Available from the UNT Digital Library; no DOI.
  • W. Givens, “Computation of plane unitary rotations transforming a general matrix to triangular form”, J. SIAM 6(1) (1958) 26–50. doi:10.1137/0106004. Crossref records the title with the misspelling “Plain”.

Two properties made rotations last. Each one touches only two rows, so a rotation can zero an entry while leaving a sparse or banded structure intact. And because the transformation is orthogonal, the rounding errors it commits are not amplified by the conditioning of the matrix being reduced.

The formula for the rotation also contains its weakness. Computing

r=a2+b2,c=a/r,s=b/rr = \sqrt{a^2 + b^2}, \qquad c = a/r, \qquad s = b/r

directly squares aa and bb. The squares overflow or underflow long before rr itself would, so a careful implementation scales the inputs first. That problem took until 2017 to settle (section 6).

Alston Householder observed that a single elementary reflector

H=I−τ vvT,τ=2vTv,H = I - \tau\, v v^T, \qquad \tau = \frac{2}{v^T v},

maps a vector xx onto a multiple of e1e_1 when v=x∓∥x∥e1v = x \mp \lVert x \rVert e_1. One reflector does what a column of rotations does, and for a dense QR it does it with about two thirds of the arithmetic. A sequence of reflectors triangularizes a matrix: the QR factorization in its modern form.

  • A. S. Householder, “Unitary Triangularization of a Nonsymmetric Matrix”, J. ACM 5(4) (1958) 339–342. doi:10.1145/320941.320947

The sign in v=x∓∥x∥e1v = x \mp \lVert x \rVert e_1 is chosen to avoid cancellation, so that the first entry of vv is a sum of two quantities of the same sign. LAPACK’s dlarfg makes this choice, mapping xx onto −sign⁡(x0)∥x∥e1-\operatorname{sign}(x_0)\lVert x \rVert e_1. B. N. Parlett’s alternative keeps the other sign, mapping a vector with x0>0x_0 > 0 onto +∥x∥e1+\lVert x \rVert e_1, and avoids the cancellation by rewriting x0−∥x∥=−σ/(x0+∥x∥)x_0 - \lVert x \rVert = -\sigma/(x_0 + \lVert x \rVert) with σ=∑i≥1xi2\sigma = \sum_{i \ge 1} x_i^2. MTL5 uses Parlett’s form; how it rounds in narrow formats is taken up in Where MTL5 sits.

4. Wilkinson and Golub: the methods become standard, 1960–1965

Section titled “4. Wilkinson and Golub: the methods become standard, 1960–1965”

James Wilkinson applied Householder’s reflectors to the symmetric eigenvalue problem, reducing a symmetric matrix to tridiagonal form with about half the work of Givens’ rotations. Wilkinson’s 1965 book analysed both transformations and established that they are backward stable: the computed result is exactly orthogonally similar to a matrix close to the original. It made backward error analysis the standard way to judge such algorithms.

Gene Golub used Householder QR to solve linear least-squares problems without forming the normal equations ATAA^TA, whose condition number is the square of AA‘s.

  • J. H. Wilkinson, “Householder’s Method for the Solution of the Algebraic Eigenproblem”, Computer J. 3(1) (1960) 23–27. doi:10.1093/comjnl/3.1.23
  • J. H. Wilkinson, The Algebraic Eigenvalue Problem, Clarendon Press, Oxford, 1965.
  • G. H. Golub, “Numerical methods for solving linear least squares problems”, Numer. Math. 7(3) (1965) 206–216. doi:10.1007/BF01436075

The modern statement of the error analysis is chapter 19 of Higham. For Householder QR of an m×nm \times n matrix, the computed R^\hat R is the exact triangular factor of A+ΔAA + \Delta A, with each column of ΔA\Delta A bounded by a small multiple of mn umn\,u times the corresponding column of AA. The bound depends only on the unit roundoff uu. Whether it carries over to a format without a fixed uu, such as a posit, is one of the open questions of section 10.

  • N. J. Higham, Accuracy and Stability of Numerical Algorithms, 2nd ed., SIAM, 2002, ch. 19 “QR Factorization” (pp. 353–380) and ch. 20 “The Least Squares Problem” (pp. 381–406). doi:10.1137/1.9780898718027

5. Square-root-free rotations: Gentleman and Hammarling, 1973–1974

Section titled “5. Square-root-free rotations: Gentleman and Hammarling, 1973–1974”

On the machines of the early 1970s a square root cost many multiplies, and a Givens QR needs one per rotation. W. Morven Gentleman showed how to avoid them. The rows are kept in a factored form D1/2RˉD^{1/2} \bar R with a separate diagonal scaling DD, and each rotation updates Rˉ\bar R and DD using multiplies and divides only. Sven Hammarling refined the formulation the following year. These are the “fast Givens” rotations of the textbooks.

  • W. M. Gentleman, “Least Squares Computations by Givens Transformations Without Square Roots”, IMA J. Appl. Math. 12(3) (1973) 329–336. doi:10.1093/imamat/12.3.329
  • S. Hammarling, “A Note on Modifications to the Givens Plane Rotation”, IMA J. Appl. Math. 13(2) (1974) 215–218. doi:10.1093/imamat/13.2.215

The price is range. The entries of DD grow or shrink geometrically with the number of rotations applied, so fast Givens must monitor them and rescale before they overflow or underflow. In binary64 that is an occasional bookkeeping step. In a format with an 8-bit or 5-bit exponent it happens almost immediately, and in a format that saturates instead of overflowing the monitor never sees an infinity to react to.

6. Computing a rotation safely: Bindel et al. 2002, Anderson 2017

Section titled “6. Computing a rotation safely: Bindel et al. 2002, Anderson 2017”

Computing cc, ss and rr from aa and bb looks trivial and is not. A careful routine must avoid overflow and underflow in a2+b2a^2 + b^2, return consistent results when aa or bb is zero, infinite or NaN, choose signs so that cc and ss vary continuously with the inputs, and handle complex arguments. LAPACK’s original slartg and clartg fell short of these requirements in several ways.

David Bindel, James Demmel, William Kahan and Osni Marques set out the requirements and gave algorithms that meet them, scaling the inputs by powers of the radix so that no rounding is introduced. Edward Anderson then derived scaling constants, in the line of J. L. Blue’s 1978 algorithm for the Euclidean norm, that make rotations and norms safe with fewer branches and less scaling. The resulting routine replaced dlartg in Reference LAPACK as dlartg.f90, which cites the 2017 paper. Which LAPACK release first shipped it, possibly 3.10, is (unconfirmed), as is whether the Bindel et al. algorithms were themselves adopted into a LAPACK release before Anderson’s rework.

  • D. Bindel, J. Demmel, W. Kahan, O. Marques, “On computing Givens rotations reliably and efficiently”, ACM TOMS 28(2) (2002) 206–238. doi:10.1145/567806.567809. First circulated as LAPACK Working Note 148 (2000).
  • E. Anderson, “Algorithm 978: Safe Scaling in the Level 1 BLAS”, ACM TOMS 44(1) (2017) 1–28. doi:10.1145/3061665

Both papers derive their thresholds from the exponent range and precision of IEEE binary32 and binary64. A posit has no fixed precision: it is most precise near 1 and loses fraction bits toward the ends of its range. A saturating format has no overflow to guard against but silently clamps instead. Nobody appears to have re-derived the safe-scaling constants for either (section 10).

Blocked reflectors: the WY representations, 1987–1989

Section titled “Blocked reflectors: the WY representations, 1987–1989”

A single reflector is a rank-1 update, a Level 2 BLAS operation that runs at memory speed. Christian Bischof and Charles Van Loan showed that a product of kk reflectors can be written I+WYTI + W Y^T, so that applying all kk at once is a matrix multiply. Robert Schreiber and Van Loan gave the storage-efficient compact form I−VTVTI - V T V^T with TT upper triangular, which LAPACK’s dlarft and dlarfb use. Blocking changes the order in which roundings happen, so a blocked and an unblocked QR give different bits.

  • C. Bischof, C. Van Loan, “The WY Representation for Products of Householder Matrices”, SIAM J. Sci. Stat. Comput. 8(1) (1987) s2–s13. doi:10.1137/0908009
  • R. Schreiber, C. Van Loan, “A Storage-Efficient WY Representation for Products of Householder Transformations”, SIAM J. Sci. Stat. Comput. 10(1) (1989) 53–57. doi:10.1137/0910005

Rotations in hardware: CORDIC and systolic arrays, 1959–1992

Section titled “Rotations in hardware: CORDIC and systolic arrays, 1959–1992”

Jack Volder’s CORDIC computes a rotation by a sequence of shift-and-add micro-rotations, with no multiplier, and remains the standard way to build rotations in fixed-point hardware. Gentleman and H. T. Kung mapped Givens QR onto a triangular systolic array, in which the boundary cells compute rotations and the internal cells apply them. John McWhirter used that array for recursive least squares (QRD-RLS), the basis of adaptive beamformers and equalizers. Yu Hen Hu analysed the quantization error CORDIC commits. The fixed-point behaviour of rotations is therefore well studied; section 9 lists the newer formats for which it is not.

  • J. E. Volder, “The CORDIC Trigonometric Computing Technique”, IRE Trans. Electronic Computers EC-8(3) (1959) 330–334. doi:10.1109/TEC.1959.5222693
  • W. M. Gentleman, H. T. Kung, “Matrix Triangularization by Systolic Arrays”, Proc. SPIE 298 (1981) 19–26. doi:10.1117/12.932507
  • J. G. McWhirter, “Recursive Least-Squares Minimization Using a Systolic Array”, Proc. SPIE 431 (1983) 105–113. doi:10.1117/12.936448
  • Y. H. Hu, “The quantization effects of the CORDIC algorithm”, IEEE Trans. Signal Process. 40(4) (1992) 834–844. doi:10.1109/78.127956

The Kalman filter’s covariance update loses symmetry and definiteness in finite precision (gemmt-origins.md traces the Joseph form, one remedy). The other remedy is to propagate a factor of the covariance instead of the covariance itself: a square root SS with P=SSTP = SS^T, or a UDUTUDU^T factorization. A factor squares back to a positive semi-definite matrix whatever rounding it carries, and its condition number is the square root of PP‘s, so it needs about half the digits.

James Potter’s square-root measurement update is usually traced to an MIT Instrumentation Laboratory memo written for Apollo navigation, and to Richard Battin’s 1964 account of it; the memo, its date and Battin’s account are (unconfirmed). The confirmed published version is Potter and Stern’s 1963 AIAA paper. Kaminski, Bryson and Schmidt surveyed the square-root filters in 1971. Gerald Bierman’s 1977 book gave the UDUTUDU^T measurement update, and C. L. Thornton the matching time update by modified weighted Gram–Schmidt, in a JPL technical memorandum in 1976 and with Bierman in 1977. Whether Thornton’s 1976 work was also a UCLA doctoral dissertation is (unconfirmed). Michel Verhaegen and Paul Van Dooren compared the implementations in 1986 and showed why the factored forms are more robust.

Rotations and reflectors enter these filters directly. The square-root time and measurement updates are QR factorizations of a stacked array, the “array algorithms”. For the time update P−=FPFT+QP^- = F P F^T + Q with P=SSTP = SS^T and Q=SQSQTQ = S_Q S_Q^T, triangularize [STFTSQT]\begin{bmatrix} S^T F^T \\ S_Q^T \end{bmatrix} by Householder or Givens: the triangular factor RR satisfies RTR=FSSTFT+SQSQT=P−R^T R = F S S^T F^T + S_Q S_Q^T = P^-, so RTR^T is the new square root. Their behaviour in narrow arithmetic is the subject of mp-numerics #15.

  • J. E. Potter, R. G. Stern, “Statistical Filtering of Space Navigation Measurements”, AIAA Guidance and Control Conference, 1963. doi:10.2514/6.1963-333
  • P. Kaminski, A. Bryson, S. Schmidt, “Discrete square root filtering: A survey of current techniques”, IEEE Trans. Autom. Control 16(6) (1971) 727–736. doi:10.1109/TAC.1971.1099816
  • G. J. Bierman, Factorization Methods for Discrete Sequential Estimation, Academic Press, 1977.
  • C. L. Thornton, Triangular Covariance Factorizations for Kalman Filtering, JPL Technical Memorandum 33-798, 1976 (NASA NTRS).
  • C. L. Thornton, G. J. Bierman, “Gram-Schmidt algorithms for covariance propagation”, Int. J. Control 25(2) (1977) 243–260. doi:10.1080/00207177708922225
  • M. Verhaegen, P. Van Dooren, “Numerical aspects of different Kalman filter implementations”, IEEE Trans. Autom. Control 31(10) (1986) 907–917. doi:10.1109/TAC.1986.1104128

The two transformations do the same thing, introduce zeros orthogonally, at different granularities, and the granularity decides where each is used.

  • Dense QR and least squares: Householder. One reflector clears a whole column, at about two thirds of the cost of the rotations that would clear it. Through the compact WY form a block of reflectors is applied as matrix multiplies, so the factorization runs at Level 3 BLAS speed. LAPACK’s dgeqrf works this way. MTL5’s qr_factor hands float and double matrices to geqrf when LAPACK dispatch is enabled; its generic loop, which every other scalar type runs, applies one reflector at a time.
  • Sparse, banded, structured and updating problems: Givens. A rotation touches two rows and zeroes one chosen entry, so it creates no fill outside those rows. Adding or deleting a row of a least-squares problem, updating a QR factorization after a rank-one change, the Hessenberg least-squares problem inside GMRES and the tridiagonal one inside MINRES are all a short sequence of rotations. So is each cell of a systolic QR array.
  • Eigenvalues and the SVD: both. The reduction to Hessenberg, tridiagonal or bidiagonal form is done by reflectors. The QR iterations that follow chase a bulge down the band: the symmetric tridiagonal QR step and the Golub–Kahan SVD step use a rotation per position, and the Francis double-shift step on a Hessenberg matrix uses a 3×3 reflector per position.

Both are orthogonal, so both preserve the 2-norm of what they act on. A rounding error committed at one step is carried forward at its own size rather than multiplied by the conditioning of the matrix, which is what makes the backward-stability results of section 4 possible. Gaussian elimination has no such property: its multipliers can grow, and its stability depends on pivoting.

9. Narrow and non-IEEE arithmetic: what has been studied

Section titled “9. Narrow and non-IEEE arithmetic: what has been studied”

Half precision and bfloat16. Householder QR in fp16 and bf16 is well covered. Yang, Fox and Sanders analysed rounding-error bounds for QR in low precision, Higham and Mary surveyed mixed-precision linear algebra, and Connolly and Higham studied probabilistic rounding error analysis applied to Householder QR. Zhang and Wu built QR on fp16 tensor cores. Givens rotations in these formats have had less attention.

  • L. M. Yang, A. Fox, G. Sanders, “Rounding Error Analysis of Mixed Precision Block Householder QR Algorithms”, SIAM J. Sci. Comput. 43(3) (2021) A1723–A1753. doi:10.1137/19M1296367
  • N. J. Higham, T. Mary, “Mixed precision algorithms in numerical linear algebra”, Acta Numerica 31 (2022) 347–414. doi:10.1017/S0962492922000022
  • M. P. Connolly, N. J. Higham, “Probabilistic Rounding Error Analysis of Householder QR Factorization”, SIAM J. Matrix Anal. Appl. 44(3) (2023) 1146–1163. doi:10.1137/22M1514817
  • S. Zhang, P. Wu, fp16 tensor-core QR, arXiv:1912.05508.

Posits and takums. The only work found that runs Householder QR in posit or takum arithmetic is Hunhold and Quinlan (ARITH 2025). It compares posit, takum, bfloat16, float16 and OFP8 E4M3 in sparse LU, sparse Householder QR, GMRES with ILU, and mixed-precision iterative refinement. For the QR, SuiteSparseQR runs once in float64 to choose the row and column permutations; the factorization itself is the authors’ own unblocked Householder QR, one reflector per column, in the target format. At equal width the posits and takums are generally as accurate as the IEEE-style formats or more, with exceptions the paper reports: posit16 is less accurate in LU for the lower quartile of matrices, it occasionally lags in GMRES, and takums are occasionally marginally less accurate than posits. The paper reports solution accuracy, not the loss of orthogonality of QQ, and has no Givens rotations. Hunhold, Quinlan and Wesner’s Arnoldi study (SC25) uses the same formats, but its dense Schur and QR core runs in IEEE arithmetic through LAPACK. The ARITH paper does not say whether its OFP8 conversions saturate or overflow; it records any overflow as a failed run. What the SC25 study does is (unconfirmed). Other posit studies cover Cholesky, LU, conjugate gradients and iterative refinement only.

  • L. Hunhold, J. Quinlan, posit, takum, bfloat16, float16 and OFP8 in sparse linear solvers, ARITH 2025, 61–68. doi:10.1109/ARITH64983.2025.00019, arXiv:2412.20268.
  • L. Hunhold, J. Quinlan, S. Wesner, Arnoldi in OFP8, bfloat16, posit and takum arithmetic, SC25, 681–694. doi:10.1145/3712285.3759837, arXiv:2504.21130.
  • L. Hunhold, “Beating Posits at Their Own Game: Takum Arithmetic”, CoNGA 2024. doi:10.1007/978-3-031-72709-2_1
  • J. L. Gustafson, I. Yonemoto, “Beating Floating Point at its Own Game: Posit Arithmetic”, Supercomputing Frontiers and Innovations 4(2) (2017). doi:10.14529/jsfi170206
  • J. Buoncristiani, S. Shah, D. Donofrio, J. Shalf, IPDPS 2020, 612–621 (posit Cholesky, CG, iterative refinement). doi:10.1109/IPDPS47924.2020.00069
  • N. Nakasato et al., HPC Asia 2024, 62–72 (posit32 Cholesky and LU on FPGA and GPU). doi:10.1145/3635035.3635046
  • J. Quinlan, E. T. L. Omtzigt, posit16 LU with iterative refinement, arXiv:2408.13400 (2024).
  • M. Klöwer, P. D. Düben, T. N. Palmer, “Posits as an alternative to floats for weather and climate models”, CoNGA 2019. doi:10.1145/3316279.3316281

A posit Kalman filter. Ilangovan, Rayan and Saxena, “Improving the Stability of Kalman Filters with Posit Arithmetic” (CoNGA 2023), is the one paper found on Kalman filtering in posits. Its conference slides show the conventional filter only, with the update P=(I−KH)PP = (I - KH)P. Joseph, Potter, Carlson and UDUTUDU^T forms appear only in an operation-count comparison, as costlier alternatives. Two small filters, a rotating shaft (2 states) and the relative motion of two shafts (4 states), run in 32-bit posits with eses from 1 to 6 against binary32, with binary64 as the reference. A posit32 with a suitable eses keeps the filter stable where binary32 diverges, but the right eses depends on the range of the data: posit<32,1> fails when the covariance starts near 10910^9, where posit<32,4> works. The authors propose choosing eses by the average precision over the range the application uses. The slides were read; the proceedings paper, which is paywalled, was not. doi:10.1007/978-3-031-32180-1_9

Logarithmic number systems. Givens rotations in LNS appear only in hardware signal-processing work: an LNS QRD-RLS processor (Sacha and Irwin, 1998), the European Logarithmic Microprocessor and its QR-RLS application (Coleman et al., 2000–2008), and an LNS complex QR decomposition for MIMO receivers (Rust, Ludwig and Paul, 2013). These measure energy, area and accuracy for one application. None analyses the rotation numerically, and no LNS Householder QR was found.

Custom-width floating point. Wang and Leeser built Givens QR on FPGAs with variable-precision floating point, and Fang and Leeser’s VFloat library generalizes the formats. These papers report throughput and resource use, not how accuracy varies with width.

The searches behind this page used Crossref, OpenAlex, arXiv and the web. Full-text coverage of IEEE Xplore and Springer is weak, so “nothing found” below means nothing indexed, not proof that nothing exists. With that caveat, these appear to be open:

TopicFoundConfidence it is open
Givens rotations in posit or takumnothinghigh
Dense Householder QR in posit or takum, with orthogonality loss measuredonly sparse QR, solution accuracy only (Hunhold and Quinlan)medium–high
Householder QR in LNSnothingmedium–high
Accuracy against width for odd-width IEEE-like formats, and the effect of subnormals on rotations and reflectorsFPGA throughput onlyhigh
Saturating floating point (clamp instead of overflow) in hypot, rotation norms and reflector scalingnothinghigh
Square-root and UDUTUDU^T Kalman filters in any non-IEEE formatnothing (the one posit Kalman study uses the conventional form)high
Dense eigenvalue and SVD kernels (Hessenberg, tridiagonal and bidiagonal reduction, QR iteration) in posit, takum or LNSArnoldi with posit and takum Krylov vectors, but its dense core in IEEE (Hunhold, Quinlan and Wesner)high
Safe-scaling rotation constants (Bindel et al., Anderson) re-derived for tapered or saturating formatsnothinghigh

The studies in mp-numerics #13 (rotations and reflectors as operators), #14 (the decompositions built on them) and #15 (ill-conditioned Kalman filters) are designed to fill these. The first two are published on the mp-numerics site: rotations and reflectors and QR and least squares.

MTL5 is generic over the scalar type, so each of its rotations and reflectors is already an experiment in whatever arithmetic the caller supplies. They do not all take the same care.

  • Householder reflectors (include/mtl/operation/householder.hpp) follow LAPACK’s dlarfg by default in two respects: the column is scaled before its norm is formed, and vv is normalized to v0=1v_0 = 1. They differ in sign. MTL5 uses Parlett’s formula v0=−σ/(x0+∥x∥)v_0 = -\sigma/(x_0 + \lVert x \rVert), which maps a column with x0>0x_0 > 0 onto +∥x∥e1+\lVert x \rVert e_1, where dlarfg maps it onto −∥x∥e1-\lVert x \rVert e_1; for x=(1,1)Tx = (1, 1)^T the two give different reflectors. Both avoid cancellation, but they round differently, which matters when comparing MTL5 against LAPACK in narrow formats. A textbook form is available as an option; householder-reflector-forms.md compares the two. Since 5.21.0, householder() checks that the τ\tau and vv it computed define a reflector the type can represent, and returns the identity when they do not. Without the check, fp8 produced NaN and 16-bit fixed point a far from orthogonal QQ. The rule and its measured effect are in sylvester.md.
  • Givens rotations are computed with the unscaled formula r=a2+b2r = \sqrt{a^2 + b^2} in the GMRES Hessenberg QR (include/mtl/operation/givens.hpp), in MINRES and in the implicit QR step of the symmetric tridiagonal eigensolver (include/mtl/operation/eigenvalue_symmetric.hpp). None of them scales its inputs the way Bindel et al. and Anderson do. In binary64 the squares overflow only for inputs near 1015410^{154}; in an 8-bit or 16-bit format they overflow, saturate or underflow for ordinary inputs.
  • The nonsymmetric eigensolver (include/mtl/detail/francis_qr.hpp) uses a generic lapy2, LAPACK’s scaled x2+y2\sqrt{x^2 + y^2}, written for any scalar type because std::hypot exists only for the built-in ones.

So MTL5 has a scaled reflector with a fallback for narrow types, a scaled hypot in one eigensolver, and unscaled rotations elsewhere. Measuring what the unscaled rotations cost in each format, and what a safe-scaling rotation would have to look like for posits, takums and saturating formats, is the first study of mp-numerics #13. Its results, across 57 types, are in Givens and Householder across number systems.

These are part of the usual account but could not be checked against a primary source. They are kept above, marked (unconfirmed), and should be corrected or removed when a source is found.

  1. Potter’s MIT memo. The square-root filter is attributed to an MIT Instrumentation Laboratory memo by James Potter, around 1962–1963, written for Apollo navigation, and to its description in Richard Battin’s 1964 book Astronautical Guidance. Neither the memo’s number and date nor Battin’s account could be confirmed. The 1963 AIAA paper by Potter and Stern is confirmed.
  2. Thornton’s dissertation. Whether C. L. Thornton’s 1976 work on triangular covariance factorizations was also a UCLA PhD dissertation. JPL Technical Memorandum 33-798 (1976) is confirmed.
  3. The LAPACK release that first shipped Anderson’s dlartg.f90. The routine is in Reference LAPACK, is authored by Anderson and cites the 2017 paper; the first release to include it, possibly 3.10, is not confirmed. Nor is whether the Bindel et al. algorithms of 2002 were themselves adopted into a LAPACK release before it.
  4. OFP8 overflow semantics in the SC25 Arnoldi study. Whether its OFP8 formats saturate on overflow or produce infinity or NaN. The OCP 8-bit specification defines both a saturating and a non-saturating conversion, and how an OFP8 failure should be read depends on which one was used. The ARITH 2025 paper was checked: it does not say, and records an overflow as a failed run.