Section: New Results

Floating-point and validated numerics

Error analysis of some operations involved in the Cooley-Tukey Fast Fourier Transform

We are interested in [4] in obtaining error bounds for the classical Cooley-Tukey FFT algorithm in floating-point arithmetic, for the 2-norm as well as for the infinity norm. For that purpose we also give some results on the relative error of the complex multiplication by a root of unity, and on the largest value that can take the real or imaginary part of one term of the FFT of a vector $x$, assuming that all terms of $x$ have real and imaginary parts less than some value $b$.

Algorithms for triple-word arithmetic

Triple-word arithmetic consists in representing high-precision numbers as the unevaluated sum of three floating-point numbers (with “nonoverlapping” constraints that are explicited in the paper). We introduce and analyze in [7] various algorithms for manipulating triple-word numbers: rounding a triple-word number to a floating-point number, adding, multiplying, dividing, and computing square-roots of triple-word numbers, etc. We compare our algorithms, implemented in the Campary library, with other solutions of comparable accuracy. It turns out that our new algorithms are significantly faster than what one would obtain by just using the usual floating-point expansion algorithms in the special case of expansions of length 3.

Accurate Complex Multiplication in Floating-Point Arithmetic

We deal in [24] with accurate complex multiplication in binary floating-point arithmetic, with an emphasis on the case where one of the operands in a “double-word” number. We provide an algorithm that returns a complex product with normwise relative error bound close to the best possible one, i.e., the rounding unit $u$.

Semi-automatic implementation of the complementary error function

The normal and complementary error functions are ubiquitous special functions for any mathematical library. They have a wide range of applications. Practical applications call for customized implementations that have strict accuracy requirements. Accurate numerical implementation of these functions is, however, non-trivial. In particular, the complementary error function erfc for large positive arguments heavily suffers from cancellation, which is largely due to its asymptotic behavior. We provide a semi-automatic code generator for the erfc function which is parameterized by the user-given bound on the relative error. Our solution, presented in [31], exploits the asymptotic expression of erfc and leverages the automatic code generator Metalibm that provides accurate polynomial approximations. A fine-grained a priori error analysis provides a libm developer with the required accuracy for each step of the evaluation. In critical parts, we exploit double-word arithmetic to achieve implementations that are fast, yet accurate up to 50 bits, even for large input arguments. We demonstrate that for high required accuracies the automatically generated code has performance comparable to that of the standard libm and for lower ones our code demonstrated roughly $25\%$ speedup.

Posits: the good, the bad and the ugly

Many properties of the IEEE-754 floating-point number system are taken for granted in modern computers and are deeply embedded in compilers and low-level softare routines such as elementary functions or BLAS. In [32] we review such properties on the recently proposed Posit number system. Some are still true. Some are no longer true, but sensible work-arounds are possible, and even represent exciting challenge for the community. Some, in particular the loss of scale invariance for accuracy, are extremely dangerous if Posits are to replace floating point completely. This study helps framing where Posits are better than floating-point, where they are worse, and what tools are missing in the Posit landscape. For general-purpose computing, using Posits as a storage format only could be a way to reap their benefits without loosing those of classical floating-point. The hardware cost of this alternative is studied.

The relative accuracy of $(x+y)*(x-y)$

We consider in [8] the relative accuracy of evaluating $(x+y)(x-y)$ in IEEE floating-point arithmetic, when $x$ and $y$ are two floating-point numbers and rounding is to nearest. This expression can be used for example as an efficient cancellation-free alternative to $x^2-y^2$ and is well known to have low relative error, namely, at most about $3u$ with $u$ denoting the unit roundoff. In this paper we complement this traditional analysis with a finer-grained one, aimed at improving and assessing the quality of that bound. Specifically, we show that if the tie-breaking rule is to away then the bound $3u$ is asymptotically optimal. In contrast, if the tie-breaking rule is to even, we show that asymptotically optimal bounds are now $2.25u$ for base two and $2u$ for larger bases, such as base ten. In each case, asymptotic optimality is obtained by the explicit construction of a certificate, that is, some floating-point input $(x,y)$ parametrized by $u$ and such that the error of the associated result is equivalent to the error bound as $u\to 0$. We conclude with comments on how $(x+y)(x-y)$ compares with $x^2$ in the presence of floating-point arithmetic, in particular showing cases where the computed value of $(x+y)(x-y)$ exceeds that of $x^2$.

The MPFI Library: Towards IEEE 1788-2015 Compliance

The IEEE 1788-2015 has standardized interval arithmetic. However, few libraries for interval arithmetic are compliant with this standard. In the first part of [30], the main features of the IEEE 1788-2015 standard are detailed. These features were not present in the libraries developed prior to the elaboration of the standard. MPFI is such a library: it is a C library, based on MPFR, for arbitrary precision interval arithmetic. MPFI is not (yet) compliant with the IEEE 1788-2015 standard for interval arithmetic: the planned modifications are presented.