Computation of best \(L^{\infty }\) exponential sums for 1 / x by Remez’ algorithm
 213 Downloads
Abstract
The approximation of the function 1 / x by exponential sums has several interesting applications. It is well known that best approximations with respect to the maximum norm exist. Moreover, the error estimates exhibit exponential decay as the number of terms increases. Here we focus on the computation of the best approximations. In principle, the problem can be solved by the Remez algorithm, however, because of the very sensitive behaviour of the problem the standard approach fails for a larger number of terms. The remedy described in the paper is the use of other independent variables of the exponential sum. We discuss the approximation error of the computed exponential sums up to 63 terms and hint to a webpage containing the corresponding coefficients.
1 Introduction
This may be true for small values of k, but for \(k\rightarrow \infty \) the exponential decay of the approximation error \(e_{k}(x)=1/xE_{k}(x)\) has a negative effect. The Remez algorithm requires exponential sums \(E_{k}\) which interpolate the function 1 / x at 2k points (in this case we say that \(E_{k}\) is feasible). Since \(e_{k}\) is rather small, tiny perturbations of \(E_{k}\) can lead to a difference \(e_{k}\) with less than 2k zeros. Note that the computation of \(e_{k}\) leads to a cancellation of the leading digits. The resulting relative error is \(\eta :=\mathrm {eps/}\left\ e_{k} \right\ \) (\(\mathrm {eps}\): machine precision), provided that 1 / x and \(E_{k}(x)\) are exact machine numbers. Let \(\tilde{e}_{k}\) be the floatingpoint result. Taking into account that the computation of \(E_{k}\) by the standard Remez algorithm is illconditioned, the final relative error of \(\tilde{e}_{k}\) may become larger than one so that \(\tilde{e}_{k}\) has the wrong sign.
Another difficulty is the fact that the error \(e_{k}\) may be much smaller than \(\sqrt{\mathrm {eps}}\). This fact prevents quadratic convergence.
Because of these difficulties^{1} the best \(L^{\infty }\) approximation is often replaced by least squares approximations (cf. [7, 13]). Concerning illposedness of this least squares problem we refer to [17]. Furthermore, Sect. 2.3 will show that there are applications requiring the maximum norm, while the \(L^{2}\) norm is insufficient.
In Sect. 3 we recall the facts about the best approximation by exponential sums. To apply the Remez algorithm we introduce a crucial ‘trick’ in Sect. 4.3. We use other variables than the coefficients in (1.1). As a consequence, the represented exponential sums are always feasible. The drawback is an increased computational cost, since we apply an outer Newton method involving several inner Newton iterations per outer iteration step. The best approximation refers to an underlying interval [1, R]. In Sect. 4.6 we describe how to proceed from one R to another \(R^{\prime }\) so that always good initial values are available. Another ‘continuation’ is the step from k to \(k+1.\)
The tables in this contribution show the approximation error in dependence of the parameters k and R. They are part of a large collection of best approximations as described in Sect. 2.4. The obtained approximation errors are compared with the theoretical bounds.
2 Definition, properties, and applications
2.1 Definition
\(e_{k,[1,R]}(k)\) for \(k=1,\ldots ,7\) and various R
R  \(k=1\)  2  3  4  5  6  7 

2E00  2.128E\(\)02  2.080E\(\)04  1.834E\(\)06  1.542E\(\)08  1.261E\(\)10  1.012E\(\)12  8.020E\(\)15 
3E00  4.358E\(\)02  1.035E\(\)03  2.223E\(\)05  4.556E\(\)07  9.088E\(\)09  1.780E\(\)10  3.444E\(\)12 
4E00  5.960E\(\)02  2.191E\(\)03  7.279E\(\)05  2.311E\(\)06  7.139E\(\)08  2.167E\(\)09  6.498E\(\)11 
5E00  7.075E\(\)02  3.437E\(\)03  1.500E\(\)04  6.258E\(\)06  2.543E\(\)07  1.016E\(\)08  4.007E\(\)10 
6E00  7.825E\(\)02  4.659E\(\)03  2.463E\(\)04  1.246E\(\)05  6.143E\(\)07  2.976E\(\)08  1.424E\(\)09 
7E00  8.288E\(\)02  5.811E\(\)03  3.553E\(\)04  2.079E\(\)05  1.185E\(\)06  6.643E\(\)08  3.677E\(\)09 
8E00  8.516E\(\)02  6.878E\(\)03  4.718E\(\)04  3.098E\(\)05  1.982E\(\)06  1.246E\(\)07  7.741E\(\)09 
9E00  8.556E\(\)02  7.857E\(\)03  5.924E\(\)04  4.273E\(\)05  3.004E\(\)06  2.076E\(\)07  1.417E\(\)08 
1E01  8.752E\(\)03  7.145E\(\)04  5.577E\(\)05  4.243E\(\)06  3.173E\(\)07  2.344E\(\)08  
2E01  1.448E\(\)02  1.819E\(\)03  2.169E\(\)04  2.521E\(\)05  2.880E\(\)06  3.252E\(\)07  
3E01  1.699E\(\)02  2.627E\(\)03  3.795E\(\)04  5.336E\(\)05  7.379E\(\)06  1.008E\(\)06  
4E01  1.784E\(\)02  3.215E\(\)03  5.230E\(\)04  8.266E\(\)05  1.285E\(\)05  1.973E\(\)06  
5E01  3.659E\(\)03  6.469E\(\)04  1.110E\(\)04  1.872E\(\)05  3.121E\(\)06  
6E01  4.001E\(\)03  7.541E\(\)04  1.377E\(\)04  2.471E\(\)05  4.382E\(\)06  
7E01  4.271E\(\)03  8.474E\(\)04  1.626E\(\)04  3.066E\(\)05  5.711E\(\)06  
8E01  4.485E\(\)03  9.293E\(\)04  1.858E\(\)04  3.648E\(\)05  7.077E\(\)06  
9E01  4.655E\(\)03  1.002E\(\)03  2.074E\(\)04  4.213E\(\)05  8.458E\(\)06  
1E02  4.789E\(\)03  1.066E\(\)03  2.274E\(\)04  4.760E\(\)05  9.841E\(\)06  
2E02  1.456E\(\)03  3.707E\(\)04  9.217E\(\)05  2.261E\(\)05  
3E02  1.628E\(\)03  4.554E\(\)04  1.235E\(\)04  3.297E\(\)05  
4E02  1.695E\(\)03  5.117E\(\)04  1.467E\(\)04  4.141E\(\)05  
5E02  5.517E\(\)04  1.649E\(\)04  4.842E\(\)05  
6E02  5.811E\(\)04  1.795E\(\)04  5.438E\(\)05  
7E02  6.031E\(\)04  1.915E\(\)04  5.952E\(\)05  
8E02  6.193E\(\)04  2.016E\(\)04  6.401E\(\)05  
9E02  6.309E\(\)04  2.102E\(\)04  6.799E\(\)05  
1E03  6.385E\(\)04  2.177E\(\)04  7.153E\(\)05  
2E03  2.570E\(\)04  9.365E\(\)05  
3E03  1.047E\(\)04  
4E03  1.110E\(\)04  
5E03  1.146E\(\)04  
6E03  1.162E\(\)04  
\(\infty \)  8.556E\(\)02  1.785E\(\)02  5.052E\(\)03  1.700E\(\)03  6.428E\(\)04  2.646E\(\)04  1.163E\(\)04 
\(R_{k}^{*}\)  8.667  41.54  146.8  436.1  1154  2807  6373 
\(k=1\)  2  3  4  5  6  7 
\(e_{k,[1,R]}(k)\) for \(k=8,\ldots ,14\) and various R
R  \(k=8\)  9  10  11  12  13  14 

1E01  1.716E\(\)09  1.248E\(\)10  9.021E\(\)12  6.492E\(\)13  4.654E\(\)14  3.326E\(\)15  2.371E\(\)16 
1E02  2.016E\(\)06  4.103E\(\)07  8.303E\(\)08  1.673E\(\)08  3.357E\(\)09  6.716E\(\)10  1.340E\(\)10 
1E03  2.321E\(\)05  7.468E\(\)06  2.389E\(\)06  7.605E\(\)07  2.412E\(\)07  7.623E\(\)08  2.403E\(\)08 
1E04  5.271E\(\)05  2.232E\(\)05  9.296E\(\)06  3.844E\(\)06  1.582E\(\)06  6.481E\(\)07  2.648E\(\)07 
1E05  6.795E\(\)06  3.379E\(\)06  1.646E\(\)06  7.973E\(\)07  
\(\infty \)  5.392E\(\)05  2.611E\(\)05  1.312E\(\)05  6.807E\(\)06  3.630E\(\)06  1.984E\(\)06  1.108E\(\)06 
\(R_{k}^{*}\)  13749  28387  56502  1.089E+5  2.042E+6  3.737E+5  6.691E+5 
2.2 Properties
The optimal exponential sums allow a stable evaluation since the coefficients \(a_{\nu }\), \(b_{\nu }\) are positive.
\(e_{k,[1,R]}(k)\) for \(k=15,\ldots ,21\) and various R
R  \(k=15\)  16  17  18  19  20  21 

1E01  1.708E\(\)17  
1E02  2.667E\(\)11  5.298E\(\)12  1.050E\(\)12  2.079E\(\)13  4.110E\(\)14  8.114E\(\)15  1.600E\(\)15 
1E03  7.555E\(\)09  2.371E\(\)09  7.426E\(\)10  2.322E\(\)10  7.251E\(\)11  2.261E\(\)11  7.044E\(\)12 
1E04  1.079E\(\)07  4.388E\(\)08  1.780E\(\)08  7.213E\(\)09  2.918E\(\)09  1.179E\(\)09  4.755E\(\)10 
1E05  3.847E\(\)07  1.850E\(\)07  8.877E\(\)08  4.251E\(\)08  2.032E\(\)08  9.700E\(\)09  4.624E\(\)09 
1E06  6.280E\(\)07  3.445E\(\)07  1.867E\(\)07  1.007E\(\)07  5.421E\(\)08  2.911E\(\)08  1.560E\(\)08 
1E07  4.679E\(\)08  2.752E\(\)08  
\(\infty \)  6.311E\(\)07  3.659E\(\)07  2.155E\(\)07  1.289E\(\)07  7.811E\(\)08  4.794E\(\)08  2.976E\(\)08 
\(R_{k}^{*}\)  1.175E\(+\)6  2.027E\(+\)6  3.440E\(+\)6  5.753E\(+\)6  9.491E\(+\)6  1.546E\(+\)7  2.491E\(+\)7 
2.3 Applications
The typical property \(\exp (x+y)=\exp (x)\exp (y)\) of the exponential function shows that \(E_{k}(\sum _{\mu }x_{\mu })=\sum _{\nu =1}^{k}a_{\nu }\prod _{\mu } \exp (b_{\nu }x_{\mu })\) is a sum of separable terms.
\(e_{k,[1,R]}(k)\) for \(k=22,\ldots ,28\) and various R
R  \(k=22\)  23  24  25  26  27  28 

1E02  3.153E\(\)16  6.218E\(\)17  
1E03  2.192E\(\)12  6.813E\(\)13  2.116E\(\)13  6.566E\(\)14  2.036E\(\)14  6.309E\(\)15  1.954E\(\)15 
1E04  1.916E\(\)10  7.715E\(\)11  3.103E\(\)11  1.247E\(\)11  5.009E\(\)12  2.010E\(\)12  8.061E\(\)13 
1E05  2.201E\(\)09  1.047E\(\)09  4.975E\(\)10  2.362E\(\)10  1.120E\(\)10  5.310E\(\)11  2.515E\(\)11 
1E06  8.351E\(\)09  4.464E\(\)09  2.384E\(\)09  1.272E\(\)09  6.777E\(\)10  3.609E\(\)10  1.920E\(\)10 
1E07  1.611E\(\)08  9.404E\(\)09  5.481E\(\)09  3.190E\(\)09  1.854E\(\)09  1.076E\(\)09  6.244E\(\)10 
1E08  4.802E\(\)09  2.999E\(\)09  1.866E\(\)09  1.159E\(\)09  
\(\infty \)  1.868E\(\)08  1.185E\(\)08  7.583E\(\)09  4.898E\(\)09  3.190E\(\)09  2.094E\(\)09  1.385E\(\)09 
\(R_{k}^{*}\)  3.969E\(+\)7  6.258E\(+\)7  9.776E\(+\)7  1.513E\(+\)8  2.325E\(+\)8  3.540E\(+\)8  5.353E\(+\)8 
2.4 Available data
This paper contains tables of the approximation errors \(\varepsilon _{[1,R]}(k)\) for \(1\le k\le 56\) and various R. The data for \(1\le k\le 7\) are shown in Table 1. The parameter R takes all values \(R=n\cdot 10^{m}\) (\(n\in \mathbb {N},\)\(m\in \mathbb {N}_{0}:=\mathbb {N}\cup \{0\}\)), subject to the following bounds. The largest value is \(R=R_{k}^{*},\) where \(R_{k}^{*}\) is explained in Sect. 3.3. Larger values of R are uninteresting since \(E_{k,[1,R]}^{*}=E_{k,[1,R_{k}^{*}]}^{*}\) for \(R\ge R_{k}^{*}.\) Besides \(R\ge 2\) the lower bound is implicitly given by \(\varepsilon _{[1,R]}(k)\approx 10^{16}.\) Such values are available for all \(1\le k\le 56\) (see web page https://www.mis.mpg.de/scicomp/EXP_SUM/1_x/). The Tables 2, 3, 4, 5, 6, 7 and 8 given here contain only results for R being powers of 10. The additional Table 9 is restricted to \(R=\infty \) and \(57\le k\le 63\).
The web page https://www.mis.mpg.de/scicomp/EXP_SUM/1_x/ contains a complete table (see file ‘tabelle’). For each pair \(\left( k,R\right) \) contained in the table there is a file^{3} with the coefficients^{4}\(a_{\nu },\)\(b_{\nu }\) (\(1\le \nu \le k\)). The file contains additional data which are important as input for the computer program. In particular, the points \(\xi _{i}\) (\(1\le i\le 2k\)) with \(1/\xi _{i}=E_{k}^{*}(\xi _{i})\) are given (cf. Sect. 4.3).
\(e_{k,[1,R]}(k)\) for \(k=29,\ldots ,35\) and various R
R  \(k=29\)  30  31  32  33  34  35 

1E03  6.048E\(\)16  1.872E\(\)16  6.218E\(\)17  
1E04  3.231E\(\)13  1.294E\(\)13  5.181E\(\)14  2.073E\(\)14  8.292E\(\)15  3.315E\(\)15  1.325E\(\)15 
1E05  1.191E\(\)11  5.633E\(\)12  2.663E\(\)12  1.259E\(\)12  5.945E\(\)13  2.807E\(\)13  1.325E\(\)13 
1E06  1.021E\(\)10  5.426E\(\)11  2.882E\(\)11  1.530E\(\)11  8.114E\(\)12  4.303E\(\)12  2.281E\(\)12 
1E07  3.619E\(\)10  2.096E\(\)10  1.213E\(\)10  7.018E\(\)11  4.057E\(\)11  2.344E\(\)11  1.354E\(\)11 
1E08  7.188E\(\)10  4.452E\(\)10  2.755E\(\)10  1.704E\(\)10  1.053E\(\)10  6.499E\(\)11  4.011E\(\)11 
1E09  6.162E\(\)10  4.053E\(\)10  2.651E\(\)10  1.730E\(\)10  1.128E\(\)10  7.343E\(\)11  
\(\infty \)  9.227E\(\)10  6.188E\(\)10  4.177E\(\)10  2.837E\(\)10  1.938E\(\)10  1.331E\(\)10  9.194E\(\)11 
\(R_{k}^{*}\)  8.036E\(+\)8  1.198E\(+\)9  1.775E\(+\)9  2.614E\(+\)9  3.826E\(+\)9  5.569E\(+\)9  8.063E\(+\)9 
\(e_{k,[1,R]}(k)\) for \(k=36,\ldots ,42\) and various R
R  \(k=36\)  37  38  39  40  41  42 

1E04  5.294E\(\)16  2.117E\(\)16  8.451E\(\)17  
1E05  6.249E\(\)14  2.947E\(\)14  1.389E\(\)14  6.546E\(\)15  3.084E\(\)15  1.453E\(\)15  6.839E\(\)16 
1E06  1.208E\(\)12  6.399E\(\)13  3.388E\(\)13  1.793E\(\)13  9.484E\(\)14  5.016E\(\)14  2.652E\(\)14 
1E07  7.814E\(\)12  4.509E\(\)12  2.600E\(\)12  1.499E\(\)12  8.641E\(\)13  4.978E\(\)13  2.867E\(\)13 
1E08  2.474E\(\)11  1.525E\(\)11  9.396E\(\)12  5.787E\(\)12  3.563E\(\)12  2.193E\(\)12  1.349E\(\)12 
1E09  4.777E\(\)11  3.105E\(\)11  2.018E\(\)11  1.310E\(\)11  8.501E\(\)12  5.515E\(\)12  3.576E\(\)12 
1E10  6.365E\(\)11  4.353E\(\)11  2.962E\(\)11  2.011E\(\)11  1.364E\(\)11  9.246E\(\)12  6.262E\(\)12 
\(\infty \)  6.382E\(\)11  4.452E\(\)11  3.121E\(\)11  2.197E\(\)11  1.554E\(\)11  1.104E\(\)11  7.869E\(\)12 
\(R_{k}^{*}\)  1.162E\(+\)10  1.665E\(+\)10  2.375E\(+\)10  3.374E\(+\)10  4.772E\(+\)10  6.719E\(+\)10  9.424E\(+\)10 
\(e_{k,[1,R]}(k)\) for \(k=43,\ldots ,49\) and various R
R  \(k=43\)  44  45  46  47  48  49 

1E05  3.220E\(\)16  1.516E\(\)16  
1E06  1.402E\(\)14  7.407E\(\)15  3.913E\(\)15  2.067E\(\)15  1.091E\(\)15  5.763E\(\)16  3.042E\(\)16 
1E07  1.651E\(\)13  9.503E\(\)14  5.469E\(\)14  3.146E\(\)14  1.810E\(\)14  1.041E\(\)14  5.983E\(\)15 
1E08  8.296E\(\)13  5.100E\(\)13  3.135E\(\)13  1.926E\(\)13  1.183E\(\)13  7.266E\(\)14  4.461E\(\)14 
1E09  2.318E\(\)12  1.502E\(\)12  9.725E\(\)13  6.297E\(\)13  4.076E\(\)13  2.637E\(\)13  1.706E\(\)13 
1E10  4.238E\(\)12  2.867E\(\)12  1.939E\(\)12  1.310E\(\)12  8.852E\(\)13  5.978E\(\)13  4.036E\(\)13 
1E11  5.591E\(\)12  3.938E\(\)12  2.766E\(\)12  1.940E\(\)12  1.360E\(\)12  9.524E\(\)13  6.667E\(\)13 
\(\infty \)  5.633E\(\)12  4.047E\(\)12  2.919E\(\)12  2.113E\(\)12  1.534E\(\)12  1.118E\(\)12  8.172E\(\)13 
\(R_{k}^{*}\)  1.316E\(+\)11  1.832E\(+\)11  2.540E\(+\)11  3.509E\(+\)11  4.833E\(+\)11  6.631E\(+\)11  9.074E\(+\)11 
\(e_{k,[1,R]}(k)\) for \(k=50,\ldots ,56\) and various R
R  \(k=50\)  51  52  53  54  55  56 

1E06  1.606E\(\)16  
1E07  3.439E\(\)15  1.976E\(\)15  1.136E\(\)15  6.524E\(\)16  3.749E\(\)16  2.154E\(\)16  1.239E\(\)16 
1E08  2.739E\(\)14  1.681E\(\)14  1.031E\(\)14  6.327E\(\)15  3.880E\(\)15  2.379E\(\)15  1.459E\(\)15 
1E09  1.103E\(\)13  7.135E\(\)14  4.612E\(\)14  2.981E\(\)14  1.926E\(\)14  1.244E\(\)14  8.038E\(\)15 
1E10  2.724E\(\)13  1.838E\(\)13  1.240E\(\)13  8.362E\(\)14  5.638E\(\)14  3.800E\(\)14  2.571E\(\)14 
1E11  4.664E\(\)13  3.262E\(\)13  2.281E\(\)13  1.594E\(\)13  1.113E\(\)13  7.776E\(\)14  5.429E\(\)14 
1E12  5.966E\(\)13  4.321E\(\)13  3.120E\(\)13  2.251E\(\)13  1.622E\(\)13  1.168E\(\)13  8.410E\(\)14 
\(\infty \)  5.992E\(\)13  4.407E\(\)13  3.251E\(\)13  2.405E\(\)13  1.784E\(\)13  1.327E\(\)13  9.897E\(\)14 
\(R_{k}^{*}\)  1.238E\(+\)12  1.683E\(+\)12  2.281E\(+\)12  3.083E\(+\)12  4.155E\(+\)12  5.587E\(+\)12  7.491E\(+\)12 
\(e_{k,[1,\infty )}(k)\) for \(k=57,\ldots ,63\)
R  \(k=57\)  58  59  60  61  62  63 

\(\infty \)  7.400E\(\)14  5.547E\(\)14  4.168E\(\)14  3.139E\(\)14  3.139E\(\)14  1.793E\(\)14  1.360E\(\)14 
\(R_{k}^{*}\)  1.002E\(+\)13  1.337E\(+\)13  1.779E\(+\)13  2.362E\(+\)13  3.064E\(+\)13  4.134E\(+\)13  5.453E\(+\)13 
3 Existence and error estimates
3.1 Existence and equioscillation
The approximation problem is closely related to the interpolation by exponential sums. Because of the nonlinear nature, an interpolating exponential may fail to exist (example: \(f(x)=x\) cannot by interpolated by some \(E_{1}\) at \(x=\pm 1).\) For sufficient conditions we refer to [3, §VI.3].
3.2 Error estimate for finite R
3.3 Case of \(R=\infty \)
4 Computation
4.1 Machine precision
The coefficients of \(E_{k}^{*}\) given in https://www.mis.mpg.de/scicomp/EXP_SUM/1_x/ are computed with extended precision (\(\mathrm {eps}=1_{10}{\text {}}19\)). This fact allows us to reach approximations with \(\varepsilon _{[1,R]}\approx 10^{16}\) and better.^{5} Of course, using the corresponding coefficients in double precision, the floatingpoint errors of the function evaluation may be larger than \(\varepsilon _{[1,R]}.\)
Using standard double precision, computations up to \(\varepsilon _{[1,R]}\approx 10^{13}\) can be expected.
4.2 Remez algorithm
Usually the Remez algorithm is applied to the \(L^{\infty }\) approximation by polynomials or rational functions (cf. [18]). Approximation by nonlinear families is studied by Barrar and Loeb [2].
4.2.1 Polynomial case
The limiting polynomial \(P_{n}^{*}\) satisfies \(P_{n}^{*}(\mu _{j}^{*})+(1)^{j}E^{*}=f(\mu _{j}^{*})\) at the extremal points \(\mu _{j}^{*}\) of \(fP_{n}^{*},\) while \(E^{*}\) is the maximum norm error.
4.2.2 Case of exponential sums
Next one has to determine the extrema \(e_{k}(\mu _{i}).\) The arguments \(\mu _{i}\) belong to \(\left( \xi _{i},\xi _{i+1}\right) \) for \(1\le i\le 2k1.\) The extremum \(e_{k}(\mu _{0})\) is taken at the boundary: \(\mu _{0}=1.\) If \(R\le R_{k}^{*},\) the extremum \(e_{k}(\mu _{2k})\) belongs to \(\mu _{2k}=R.\) If \(R>R_{k}^{*},\)\(\mu _{2k}\) lies in \(\left( \xi _{2k},\infty \right) \).
Given the values \(\mu _{i},\) (4.4) represents 2k equations for the 2k coefficients. This yields a new \(E_{k}^{\mathrm {new}}.\) If \(E_{k} ^{\mathrm {new}}\) is feasible, one can determine the new \(\mu _{i}\) etc.
The main difference to the polynomial case in Sect. 4.2.1 is the nonlinear nature of the system (4.4).
4.3 Choice of variables
The cause of the difficulty is explained as follows. Due to the good approximation, \(\varepsilon _{[1,R]}(k)\) is small. Assume that there is an iterate \(E_{k}\) with \(\Vert \frac{1}{\cdot }E_{k}\Vert _{\infty }\le 10^{13}.\) Then a perturbation of \(E_{k}\) by a tiny shift of, e.g., \(2\cdot 10^{13}\) may yield an \(E_{k}^{\mathrm {new}}\) which does not interpolate \(\frac{1}{\cdot }.\) Hence there are no zeros \(\xi _{i}\) (or at least not 2k of them), \(E_{k}^{\mathrm {new}}\) is infeasible, and the algorithm cannot be continued.
4.4 Computation of \(\hat{E}_{k}(\cdot ;\varvec{\xi })\): inner iteration
For the practical computation of \(\hat{E}_{k}(\cdot ;\varvec{\xi })\), one uses the pair \((\varvec{\xi },\mathbf {a}(\varvec{\xi })).\) If one wants to determine \(\hat{E}_{k}(\cdot ;\varvec{\xi }^{\prime })\) for \(\varvec{\xi }^{\prime }\) close to \(\varvec{\xi },\) one has to solve \(E_{k}(\xi _{i}^{\prime };\mathbf {a}^{\prime })=1/\xi _{i}^{\prime }\) with respect to \(\mathbf {a}^{\prime }.\) Here we exploit the fact that the Newton method has a starting value \(\mathbf {a}=\mathbf {a}(\varvec{\xi })\) very close to \(\mathbf {a}^{\prime }=\mathbf {a}(\varvec{\xi }^{\prime })\).
The interpolation is harder to compute if \(\xi _{i}\approx \xi _{i+1}\) are very close. However, the zeros \(\xi _{i}\) of the best approximation \(E_{k}^{*}\) are well separated (cf. Remark 4.2).
Note that the use of the parameters \(\varvec{\xi }\) ensures that \(\hat{E}_{k}(\cdot ;\varvec{\xi })\) is a feasible exponential sum. The drawback of this approach is a larger computational work. Instead of the evaluation of \(E_{k}(\cdot ;\mathbf {a})\) the Newton method requires several evaluations of exponential sums and their derivatives (which are again exponential sums).
In the following, iterates \(\varvec{\xi }^{0},\varvec{\xi }^{1},\ldots \) of an outer iteration will appear. Each \(\varvec{\xi }^{\nu }\) requires an inner iteration for the computation of \(\mathbf {a} (\varvec{\xi }^{\nu })\) by the Newton method described above. A standard value of the number of inner iteration steps used in Sect. 2.4 is 4.
4.5 Computation of the extrema
In principle, given \(\varvec{\xi }\) and \(\mathbf {a}=\mathbf {a} (\varvec{\xi }),\) the location \(\mu _{i}\) of the extrema can be determined by Newton’s method. For the results given in the tables the extrema are computed differently. In each interval \([\xi _{i1},\xi _{i}]\) the values at \(x_{\nu }=\xi _{i1}+\nu \left( \xi _{i}\xi _{i1}\right) /N\) are evaluated for \(\nu =1,\ldots ,N1.\) Let \(\nu ^{*}\) be the index with \(x_{\nu ^{*} }={\text {argmax}}_{\nu }e_{k}(x_{\nu })\) and let P the cubic interpolation at \(x_{\nu ^{*}1},x_{\nu ^{*}},x_{\nu ^{*}+1}.\) The location of the maximum of P is taken as \(\mu _{i}.\)
Note that the evaluation of \(e_{k}\) around the extremal point \(\mu _{i}\) is harmless because of \(\frac{\mathrm {d}}{\mathrm {d}x}e_{k}(\mu _{i})=0.\) Already the choice of \(x_{\nu ^{*}}\) leads to an error \(e_{k}(x_{\nu ^{*} })=(1+{\mathcal {O}}(N^{2}))\max _{\xi _{i1}\le x\le \xi _{i}}e_{k} (x)+{\mathcal {O}}(\mathrm {eps}).\) Cubic interpolation leads to \(e_{k}(\mu _{i}^{\mathrm {cubic}})=(1+{\mathcal {O}}(N^{8}))\max _{\xi _{i1}\le x\le \xi _{i}}e_{k}(x)+{\mathcal {O}}(\mathrm {eps}).\) Here \({\mathcal {O}} (\mathrm {eps})\) is the unavoidable cancellation error.
Remark 4.1
Since the computational work increases with N, only the final data should be performed with a large N. Intermediate steps can be done with small N.
The values in the tables are based on \(N=1000.\)
4.6 Newton’s method for (4.4) and continuation principle
4.6.1 The outer iteration
The Newton method for solving (4.4) is formulated with respect to the parameters \(\varvec{\xi }.\) Equation (4.4) becomes \(0=e_{k} (\mu _{i1})+e_{k}(\mu _{i})=\frac{1}{\mu _{i1}}\hat{E}_{k}(\mu _{i1} ;\varvec{\xi })+\frac{1}{\mu _{i}}\hat{E}_{k}(\mu _{i};\varvec{\xi })=:\phi _{i}(\varvec{\xi }).\) The evaluation of \(\phi _{i}(\varvec{\xi })\) is explained in Sect. 4.3. The derivatives \(\frac{\partial }{\partial \xi _{j}}\phi _{i}(\varvec{\xi })\) required by the Newton method are replaced by divided differences.
We call this iteration the outer iteration since each evaluation of \(\hat{E}_{k}(\cdot ;\varvec{\xi })\) for a new \(\varvec{\xi }\) requires inner iteration steps according to Sect. 4.4.

the approximation error \(\left\ e_{k}\right\ _{\infty }\) is decreasing,

the zeros \(\xi _{i}\) of \(e_{k}\) should not come too close.
Remark 4.2
In a first approximation, the zeros behave as \(\xi _{i}\approx R^{(i/(2k))^{c}}\) with c between 1.2 and 1.3.
If it happens that \(\xi _{i}\) and \(\xi _{i+1}\) are too close, one should change the positions, e.g., by \(\xi _{i+1}:=(\xi _{i}+\xi _{i+2})/2\).
The continuation method explained in Sect. 4.6.3 ensures that the Newton method can be started with initial iterates close to the solution.
4.6.2 Start
The process is started by computations for \(k=1.\) Consider, e.g., \(R=2.\) A simple choice of interpolation points is \(\xi _{1}=4/3\) and \(\xi _{1}=5/3.\) The computation of \(\mathbf {a}(\varvec{\xi })\) is harmless. For all initial values \(0.1\le a_{1},b_{1}\le 5\), the Newton method converges (to \(a_{1}=1.831\ldots ,\ b_{1}=0.6694\ldots \)). After 6 steps of the outer Newton method for solving (4.4), one obtains the best approximation up to machine precision.
Given any best approximation \(E_{k,[1,R]}^{*}=E_{k,[1,R]}^{*} (\cdot ,\varvec{\xi }^{*})\) together with the interpolation points \(\varvec{\xi }^{*},\) one can obtain \(E_{k,[1,R^{\prime }]}^{*}\) for other \(R^{\prime }\) according to Sect. 4.6.3.
4.6.3 Change of R
In the following, k is fixed. Assume that a best approximation \(E_{k,[1,R]}^{*}\) is already available for some R.
The first task is to compute \(E_{k,[1,R^{\prime }]}^{*}\) for a larger\(R^{\prime }>R.\) The approximation error of \(E_{k,[1,R]}^{*}\) is \(\varepsilon _{[1,R]}(k).\) Take \(E_{k}:=E_{k,[1,R]}^{*}\) as initial value for the outer iteration of Sect. 4.6.1 on \([1,R^{\prime }].\) If \(R^{\prime }\le R_{k}^{*},\) the maximum in the last subinterval \([\xi _{2k},R^{\prime }]\) is taken at \(x=R^{\prime }.\) If \(R^{\prime }\) is not close enough to R, the maximum \(e_{k}(R^{\prime })\) may be much larger than \(\varepsilon _{[1,R]}(k)\) and the Newton method may fail. In that case one has apply the continuation method: compute \(E_{k,[1,R_{m}]}^{*}\) for a sequence \(R=R_{0}<R_{1}<\cdots <R_{M}=R^{\prime },\) where each \(R_{m+1}\) is sufficiently close to \(R_{m}.\)
If it happens that the last extremum is at \(\mu _{2k}<R^{\prime },\) one has obtained the best approximation in \([1,\infty ]\) and \(\mu _{2k}=R_{k}^{*}\) holds (cf. Sect. 3.3).
In the case of a smaller\(R^{\prime }<R\) the same continuation method can be used. However, the new value \(R^{\prime }\) should be larger than \(\xi _{2k}\). Otherwise, the interval \([1,R^{\prime }]\) contains less than 2k values \(\xi _{i}\) of the vector \(\varvec{\xi }\) defining the initial value \(E_{k,[1,R]}^{*}=E_{k}(\cdot ,\varvec{\xi })\) and the Newton iteration may fail.
For the intermediate computations with \(R=R_{0}<R_{1}<\cdots <R_{M}=R^{\prime }\) one may save computational work by Remark 4.1.
For k large and R small, the points \(\xi _{i}\) are rather close. In this case, the restriction \(\xi _{2k}<R^{\prime }<R\) for the new value \(R^{\prime }\) is very restrictive. Here, another strategy can be applied. The affine^{6} map \(x\mapsto 1+\frac{R^{\prime }1}{R1}\left( x1\right) \) maps [1, R] onto \([1,R^{\prime }].\) Applying this map to \(\varvec{\xi },\) one obtains a rough approximation of \(\varvec{\xi }^{\prime }\) corresponding to \(E_{k,[1,R^{\prime }]}^{*}=E_{k}(\cdot ,\varvec{\xi }^{\prime }).\) Note that \(\xi _{2k}^{\prime }<R^{\prime }\) holds. After defining \(\varvec{\xi }^{\prime },\) the coefficients \(\mathbf {a}^{\prime }=\mathbf {a}(\varvec{\xi }^{\prime })\) have to be computed by the inner iteration of Sect. 4.4. If the factor \(\frac{R^{\prime }1}{R1}\) is chosen too small, this iteration may fail to converge.
If one follows these hints, the outer Newton iteration even works without damping.
We illustrate these hints by two examples. First we start from the best approximation \(E_{k,[1,R]}^{*}\) for \(k=7,\)\(R=1000.\) This case is less sensitive, i.e., the Newton iteration behaves rather robust. For a larger \(R^{\prime }\in [R,10\cdot R]\) the outer Newton iteration converges. The choice of a smaller \(R^{\prime }\) is restricted by \(R^{\prime }>\xi _{14}=838.\) Much smaller \(R^{\prime }\) can be obtained by the second strategy using new \(\varvec{\xi }^{\prime }.\) Convergence is observed for amplification factors \(\frac{R^{\prime }1}{R1}\in [0.5,2].\)
A more sensitive example is \(E_{k,[1,R]}^{*}\) for \(k=50,\)\(R=2_{10}8\) since \(\varepsilon _{[1,2_{10}8]}(50)=4.43_{10}{\text {}}14\). An increase of R leads to convergence as long as \(R^{\prime }\le 3.5_{10}8=\frac{7}{4}R.\) A decrease of R is restricted by \(R>\xi _{100}=1.967_{10}8.\) The second strategy is more helpful. Using the factor \(\frac{R^{\prime }1}{R1}\in [0.92,1.1],\) the inner iterations converge and enable the computation for \(R^{\prime }\in [1.84_{10}8,\,2.2_{10}8]\).
4.6.4 Increasing k
The continuation principle applied to the kdependence means that we increase k only to the next integer. The step \(k\mapsto k+1\) is more delicate since two additional parameters must be created. They must be such that the exponential sum is feasible and has two additional zeros. The strategy which is successfully applied for computing the data of the tables is explained using an example with rather large k. Note that the difficulty increases with the size of k.
\(\nu \)  1  2  3  ...  56 

\(a_{\nu }\)  1.65\(_{10}\)\(\)12  1.08\(_{10}\)\(\)11  4.44\(_{10}\)\(\)10  ...  5.7 
\(b_{\nu }\)  5.16\(_{10}\)\(\)13  5.68\(_{10}\)\(\)12  3.01\(_{10}\)\(\)11  ...  14.4 
The ansatz for \(k=57\) is \(E_{57}(x):=a_{0}\exp (b_{0}x)+E_{56,[1,R]}^{*},\) i.e., we use the previous exponential sum supplemented by a further term. The index 0 emphasises that the new term is on the side of the very small coefficients. Afterwards the indices \(\{0,\dots ,56\}\) are replaced by \(\{1,\dots ,57\}\).
The additional term \(a_{0}\exp (b_{0}x)\) should be so small that the equioscillation structure of \(E_{56,[1,R]}^{*}\) is not perturbed. Since \(\varepsilon _{[1,R]}(56)\approx 1_{10}\)13, we need \(a_{0}\le 1_{10}\)13. On the other hand, \(a_{0}\) should not be too small. The reason is the Newton process of Sect. 4.4 for computing \(\mathbf {a} (\varvec{\xi }).\) Here the derivative \(\frac{\mathrm {d}}{\mathrm {d}b_{0} }a_{0}\exp (b_{0}x)=a_{0}x\exp (b_{0}x)\) should not be too close to zero.
i  1  2  3  ...  109  110  111  112 

\(\xi _{i}\)  1.003  1.026  1.074  ...  159\(_{10}\)11  3.57\(_{10}\)11  9.27\(_{10}\)11  3.26\(_{10}\)12 
The first step are many (damped) Newton iterations for computing the coefficients \(\mathbf {a}(\varvec{\xi })\) and \(E_{57}\) for the new \(\varvec{\xi }\). It turns out that the last maximum is taken at \(\mu _{115}=1.9_{10}14.\) Accordingly we choose \(R=1.9_{10}14.\)
The first outer iteration requires a damping^{7} of the Newton correction by 1 / 8. The values \(\xi _{i}\) as well as \(\mu _{115}\) have decreased and \(R=\mu _{115}=5.1_{10}13\) can be chosen. About 4 damped outer Newton steps are needed, before the Newton method works without damping.
The choice of the initial values \(a_{0},\ b_{0},\)\(\xi _{2k1},\xi _{2k}\) may be a matter of trial and error.
4.7 Modifications
4.7.1 Wavelet applications
In [9] an application in quantum mechanics is mentioned which involves exponential sums for \(1/\sqrt{x}.\) The technique explained in [9] works equally well for the function 1 / x.
Polynomial degree  \(1\)  0  1 
Approximation error  2.344\(_{10}{\text {}}8\)  6.554\(_{10}{\text {}}9\)  1.934\(_{10}{\text {}}9\) 
4.7.2 Weighted norm
4.7.3 Quadrature
Since \(\frac{1}{x}=\int _{0}^{\infty }\exp (xt)\mathrm {d}t,\) any quadrature formula \(Q(f)=\sum _{\nu =1}^{k}a_{\nu }f(b_{\nu })\) applied to the function \(f(\cdot )=\exp (x\,\cdot )\) yields an exponential sum of the form (1.1).
A particular choice is the sinc quadrature. It requires an integral over \(\mathbb {R}.\) For instance, the substitution \(t=\exp (y)\) yields an integral of the desired form: \(\frac{1}{x}=\int _{\infty }^{\infty }\exp (x\exp (y))\exp (y)\mathrm {d}y.\) The sinc quadrature is defined and analysed in [11, Sect. D.4]. The drawback is that the quadrature is not adapted to the fact that the interesting parameters x belong to the interval [1, R]. The \(L^{\infty }\) error estimate is of the form \(\le c\exp (c^{\prime }\sqrt{k})\) (cf. [4, (3.6)]). Figure 3 shows the error \(e_{k}\) for \(k=45.\) The error bound is 2.63\(_{10}{\text {}}7\), while the best approximation—say in \([1,10^{6}]\)—is 3.913\(_{10}{\text {}}15\). It is oscillating, but remains in the positive part, i.e., it is infeasible. Therefore it is not possible to use the quadrature result as starting value for the Remez algorithm.
Footnotes
 1.
In [12] the author writes: In general, the problem of finding a best uniform approximation ... is quite difficult, and consequently various schemes ... have sometimes been used to produce “good” if not best approximations.
 2.
Here we only consider real, in particular positive, coefficients.
 3.
For instance, the file k12.3E5 contains the coefficients corresponding to \(k=12\) and \(R=3_{10}5.\) The file with the largest value of R (k fixed) yields the best approximation for \([1,\infty ].\)
 4.
In the file, \(a_{\nu }\) and \(b_{\nu }\) are called omega[\(\nu \)] and alpha[\(\nu \)].
 5.
In the case of \(\varepsilon _{[1,R]}\approx 10^{17}\) the extrema \(e_{k}(\mu _{i})\) start to be hard to equalise for small i where the cancellation effect is strongest. Nevertheless, the function \(E_{k}\) is still feasible. Then \(E_{k}\) is not the best approximation, but \(\Vert \frac{1}{\cdot }E_{k}\Vert _{\infty }\le \max _{i}e_{k}(\mu _{i})\) holds.
 6.
An improvement could be a mapping making use of Remark 4.2.
 7.
The damping strategy uses the criteria of Sect. 4.6.1.
Notes
Acknowledgements
Open access funding provided by Max Planck Society.
References
 1.Ayala, P.Y., Scuseria, G.E.: Linear scaling secondorder Moller–Plesset theory in the atomic orbital basis for large molecular systems. J. Chem. Phys. 110, 3660 (1999)CrossRefGoogle Scholar
 2.Barrar, R.B., Loeb, H.L.: On the Remez algorithm for nonlinear families. Numer. Math. 15, 382–391 (1970)MathSciNetCrossRefzbMATHGoogle Scholar
 3.Braess, D.: Nonlinear Approximation Theory. Springer, Berlin (1986)CrossRefzbMATHGoogle Scholar
 4.Braess, D., Hackbusch, W.: Approximation of \(1/x\) by exponential sums in \([1,\infty )\). IMA J. Numer. Anal. 25, 685–697 (2005)MathSciNetCrossRefzbMATHGoogle Scholar
 5.Braess, D., Hackbusch, W.: On the efficient computation of highdimensional integrals and the approximation by exponential sums. In: DeVore, Ronald, A., Kunoth, A. (eds.) Multiscale, Nonlinear and Adaptive Approximation, pp. 39–74. Springer, Berlin (2009)CrossRefGoogle Scholar
 6.de la Vallée Poussin, C.J.: Leçons sur l’approximation des fonctions d’une variable réelle. GauthierVillars, Paris (1919)zbMATHGoogle Scholar
 7.Evans, J.W., Gragg, W.B., LeVeque, R.J.: On least squares exponential sum approximation with positive coefficients. Math. Comput. 34, 203–211 (1980)MathSciNetCrossRefzbMATHGoogle Scholar
 8.Hackbusch, W.: Entwicklungen nach Exponentialsummen. Techn. Bericht 25, MaxPlanckInstitut für Mathematik in den Naturwissenschaften, Leipzig (2005)Google Scholar
 9.Hackbusch, W.: Approximation of \(1/\leftxy\right\) by exponentials for wavelet applications. Computing 76, 359–366 (2006)MathSciNetCrossRefzbMATHGoogle Scholar
 10.Hackbusch, W.: Tensor Spaces and Numerical Tensor Calculus, volume 42 of SSCM. Springer, Berlin (2012)CrossRefGoogle Scholar
 11.Hackbusch, W.: Hierarchical Matrices—Algorithms and Analysis, volume 49 of SSCM. Springer, Berlin (2015)Google Scholar
 12.Kammler, D.W.: Chebyshev approximation of completely monotonic functions by sums of exponentials. SIAM J. Numer. Anal. 13, 761–774 (1976)MathSciNetCrossRefzbMATHGoogle Scholar
 13.Kammler, D.W.: Least squares approximation of completely monotoic functions by sums of exponentials. SIAM J. Numer. Anal. 16, 801–818 (1979)MathSciNetCrossRefzbMATHGoogle Scholar
 14.Remez, E.J.: Sur un procédé convergent d’approximations successives pour déterminer les polynômes d’approximation. Compt. Rend. Acad. Sci. 198, 2063–2065 (1934)zbMATHGoogle Scholar
 15.Süli, E., Mayers, D.F.: An Introduction to Numerical Analysis, 4th edn. Cambridge University Press, Cambridge (2008)zbMATHGoogle Scholar
 16.Takatsuka, A., Tenno, S., Hackbusch, W.: Minimax approximation for the decomposition of energy denominators in Laplacetransformed Møller–Plesset perturbation theories. J. Chem. Phys. 129, 044112 (2008)CrossRefGoogle Scholar
 17.Varah, J.M.: On fitting exponentials by nonlinear least squares. SIAM J. Sci. Statist. Comput. 6, 30–44 (1985)MathSciNetCrossRefzbMATHGoogle Scholar
 18.Werner, H.: Vorlesung über Approximationstheorie, volume 14 of Lect. Notes Math. Springer, Berlin (1966)Google Scholar
Copyright information
OpenAccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.