A&A 477, 967-977 (2008)
S. Bukvic1 - Dj. Spasojevic1 - V. Zigman2
1 - Faculty of Physics, University of Belgrade, Studentski Trg 12-16, Belgrade, Serbia
2 - University of Nova Gorica, Vipavska cesta 13, Nova Gorica, Slovenia
Received 4 July 2006 / Accepted 23 September 2007
Aims. The purpose of this paper is to introduce a robust method of data fitting convenient for dealing with astrophysical spectra contaminated by a large fraction of outliers.
Methods. We base our approach on the suitable defined measure: the density of the least squares (DLS) that characterizes subsets of the whole data set. The best-fit parameters are obtained by the least-square method on a subset having the maximum value of DLS or, less formally, on the largest subset free of outliers.
Results. We give the FORTRAN90 source code of the subroutine that implements the DLS method. The efficiency of the DLS method is demonstrated on a few examples: estimation of continuum in the presence of spectral lines, estimation of spectral line parameters in the presence of outliers, and estimation of the thermodynamic temperature from the spectrum that is rich in spectral lines.
Conclusions. Comparison of the present results with the ones obtained with the widely used comprehensive multi-component fit yields agreement within error margins. Due to simplicity and robustness, the proposed approach could be the method of choice whenever outliers are present, or whenever unwelcome features of the spectrum are to be considered as formal outliers (e.g. spectral lines while estimating continuum).
Key words: methods: data analysis - methods: numerical - techniques: spectroscopic - line: profiles
The task often met in data processing is to find the best parameters of some model function that would represent a particular set of data in the best-fit sense. For this purpose it is a common practice to use the least squares as a merit function for various data fittings. Unfortunately, measurements are not free of errors, which occasionally cause some experimental points to simply be way off. Also, data of interest are frequently spoiled by the irregular deviation of a certain number of data points due to some undesirable process that cannot be avoided. This can easily turn the results obtained with the least-square method into nonsense.
The reliability of the best-fit could be significantly enhanced if outliers could be successfully recognized and subsequently removed. A variety of statistical tests, reviewed in (Barnett & Lewis 1994), have been developed to decide whether a data point is an outlier or not. If there is a sufficient number of replicate points at each value of x, these tests could be useful for determining whether a single value is a significant outlier. Unfortunately, each point is measured only once in typical experiments. In such a case, one can choose to perform an outlier test on the entire set of residuals of the least-square fit. The problem arises if outliers are allocated asymmetrically. In this case the best-fit curve tends to be inclined towards the outliers, so that outlying points might not be recognized.
In order to minimize the impact of outliers, a multitude of robust approaches has been proposed (Hoaglin et al. 1983; Hampel et al. 1986). While efficient in the case of symmetrically allocated outliers, most of the robust fit approaches are scale-dependent and ineffective if a large fraction of outlying points lie on the same side of the best-fit curve. Some relevant ideas for how to deal with outliers in case of a linear, as well as nonlinear, fit are presented in (Marx 1996; Motulsky & Brown 2006; Liu et al. 2005).
Well-known tasks to which the least-square method is not directly applicable include determining the spectrum base line and spectrum continuum. Both issues often arise in spectroscopy, since for many applications in astrophysics (Peterson et al. 2002; Freudling et al. 2003; Gilbert & Peterson 2003; Contini & Viegas 2000; Rossa et al. 2000; Gabel et al. 2005) and in physics (Meisel & Cote 1992; Spasojevic et al. 1996), accurate knowledge of both the local continuum value and signal base line is essential. To overcome the drawbacks of the least squares, a few specific, non robust approaches have been proposed. One of them is to estimate the continuum and power-law exponents from the spectral regions uncontaminated by emission or absorption lines (Liu & Yan 2005; Wang et al. 2005; Gabel et al. 2005). Within this approach, the spectral lines are unwelcome features - formal outliers, which can produce spurious results. Another approach is to apply a complex multi-component fit (Dietrich et al. 2002; Freudling et al. 2003; Dietrich et al. 2005; Gabel et al. 2005), which takes all known features of the spectrum into account. However, some unknown, non-implied features, or true outlying points can still induce inaccurate estimates of the fit coefficients.
In a recent paper (Bukvic & Spasojevic 2005), a robust method for estimating of the spectrum base line and continuum from large data sets has been proposed. The method was based on the close points concept (CPC), capable to distinguish outliers from close points, with the resulting best-fit parameters obtained only on the set of close points.
In this paper we develop this technique further by applying the least-square approach in accordance with CPC. The proposed method is insensitive to the presence of a large number of outlying points, which makes it appropriate for various purposes. For example, in spectrum analysis, it can be used to fit the full spectrum (base line/continuum + all spectral lines), to fit only selected spectrum characteristics (base line/continuum + selected spectral lines, while all other spectral lines are treated as outliers), or to fit the base line/continuum only (with all spectral lines being treated as outliers).
Consider a data set consisting of n0 experimental points and a family of model functions depending on x and on the parameters , specifying a model function from the family. In the least-square approach, the best model function is usually found by applying the method to the whole data set s0. However, the set s0 may be contaminated by various types of unwelcome data. Therefore, it could happen that the model function that optimally reflects the features of interest actually corresponds not to the initial data set s0, but rather to some of its subsets, sb - the "best'' one - not known in advance.
In order to identify it, we introduce the density of least
squares - the quantity suitable for characterizing the initial
data set s0 and any particular subset s, in the following
The subset sb, singled out by having maximum density of the
least squares, is taken as the best subset; its OLS best-fit
a(sb) are considered to yield the most
reliable estimate of the features of interest, hence, are accepted
as the final outcome of the proposed method. The estimation of
of the OLS best-fit
a(sb) is explained in
Sect. 4. To conform to standard
terminology, we introduce the merit function
We call the foregoing approach the density of least-square (DLS) method. We call the points from subset sb close points, while all other points we call distant points or outliers with respect to the model function f(xi;a(s)). In the same sense, the width of the best subset can be conveniently used to set a discrimination level between distant and close points.
So far, we have considered data sets with unspecified errors of
If individual errors are, on the other hand, assigned to corresponding yi then,
instead of geometrical distances di, relative distances
should be used. Accordingly, the density
of weighted least squares is defined by an appropriate
extension of Eq. (1):
As shown in the Appendix, the best-fit parameters a(sb) found by the DLS method are, actually, the maximum likelihood parameters, obtained by a specific Local maximum-likelihood estimator. This evidence, together with the calibration procedure put forward in Sect. 3.1, forms the formal basis of the DLS method.
Now we proceed to intuitive considerations, illustrating how DLS works in the case k=2. Let s0 be a data set containing one outlying point with respect to a model function, and let s be its subset that contains all points from s0, except the outlying point. Although both the numerator and denominator in Eq. (1) are smaller on s than on s0, the effect of the decrease in in is lessened by the contribution of other points, so that the denominator decrease becomes dominant. It follows that the density Dk(s) of the subset s will typically be higher than Dk(s0). More generally, if the data set s0contains outlying points, then the density of least squares typically attains its maximum value on some subset that is free of outliers, but still contains a large percentage of data.
With respect to definition (2), we would like to emphasize that we do not simply seek for parameters that minimize the value of merit function on the initial data set. Instead, we are presently concerned with finding the best subset sb for which the density of least squares exhibits its maximum. This task cannot be reduced to the standard minimization problem.
Indeed, the standard minimization problem takes place in a multidimensional space of model function parameters. Provided the initial guess is given, the standard minimization algorithms follow the gradient direction, assuring rapid approach to some local minimum. For functions that, like merit function (2), have data subset as an argument some appropriate substitute for the gradient minimization algorithm is to be sought; see the discussion for details. To this end we establish an ordered collection of data subsets in agreement with our main intention of removing outlying points, so as to obtain the subset with the maximum density of least squares.
We define the ordered collection of data subsets by iterating the following two-stage procedure. We first calculate the OLS best-fit model function for the given subset s and then remove from sthe point with the greatest distance with respect to the best-fit curve. In the second stage, we apply OLS only to the retained data points and recalculate distances di of the remaining data points from the new best-fit model function. If there are points with distances larger than the starting , we remove them as well. We repeat these steps until all points with distances are removed. The smaller subset of retained points, with new , is now ready for the next iteration cycle. Points eliminated from s by this procedure, see Fig. 1, constitute a layer of outer points for subset s or, simply, layer.
|Figure 1: Flowchart of the procedure used to obtain the ordered collection . Index j=n-l denotes the number of points in the subset s(j). Yellow and blue colors indicate algorithm steps of the first and of the second stages, respectively. The algorithm step painted in both colors is invoked only once within the first stage, at the very beginning; all its subsequent calls are within the second stage.|
By using the foregoing procedure, we iteratively remove layer-by-layer from the initial data set s0, and we eventually obtain an ordered collection of data subsets where index pcounts the subsets in the collection. After each iteration, we calculate Dk(sp). Although one can fit the subset of n data points to a m-parameter model function technically whenever , we terminate the build-up of collection whenever a more conservative condition n< m+3 is met, or after Dk(sp) becomes indefinite - see Sect. 3.2. As the best subset sb, we accept the subset from collection having a maximum value of DLS. For the users' convenience, a FORTRAN code of the DLS data-fitting method is given in Sect. 8.2.
Let the data set contain n0normally distributed points around the value y=0. We assume that the number n0 is large enough for the statistical parameters related to s0 to be indistinguishable from their corresponding limits, when .
Then, the number dn of yi's in the interval
On the other hand, one can show that inequality holds for any subset s with , implying that the maximum of is equal to the DLS maximum. The same maximum is found in the ordered collection , since consists of d-stripes. In the top panel of Fig. 2 we present plots of scaled by for three characteristic values of parameter k; each curve manifests distinct maximum.
|Figure 2: Top panel: versus relative width of d-stripes for Gaussian distribution. Presented curves correspond to three characteristic values of k, as given in the legend, together with abscissae of the corresponding maxima; is the percentage of points contained in the corresponding db-stripes. Bottom panel: exponent k versus , according to Eq. (6) for Gaussian distribution - bold curve. As indicated, analogous curves for Lorentzian and for double-exponential distribution are also given. In DLS method we only use the values .|
The condition for the maximum of (5), i.e.
leads to a scale-independent
The bottom panel of Fig. 2 presents k versus , for a Gaussian distribution (bold line), according to (6). For comparison, analogous curves are shown for Lorentzian and for double-exponential distribution.
It should be noted that exponent k is not completely free of constraints. The upper limit is strict and amounts to k<3. Namely, from Eq. (6), it follows that , where . For , we also have , so that . Therefore, the upper limit is evaluated as .
The lower limit is not as strict and cannot be deduced from Eq. (6), which holds in the case of pure Gaussian scatter around a y=0 value. Apart from defining the width of the best subset, exponent k has a significant impact on the ability of the DLS method to resolve outliers. When outliers actually do occur, details of bulk data distribution become of secondary importance. The following example illustrates the effect of the exponent k on the order-of-magnitude scale. Consider a set s0containing an odd number n0=2m+1 of points, such that xi=i, (i=1, n0), , , and y2m+1=1. For small , the last point is a typical outlier. Let us choose the horizontal lines y(x)=a as model functions. Considering that the best horizontal line level ab is equal to the mean value of , we find for the whole data set, and for the set s, obtained by removing the outlying point from s0. For sufficiently small and k<2, it follows that Dk(s0)>Dk(s), which implies that the outlier could not be removed if k<2. Therefore, we accept the constraint as the lower limit for the exponent k, with the remark that in the case of well-behaved data distributions, with soft outliers, DLS will work even if k<2, though its ability to resolve outliers will be reduced.
By removing layer-by-layer from the initial data set, the successive data subsets contain a smaller and smaller number of points. It may happen, eventually, that we have to deal with a subset s, such that each of its n data points lies on the corresponding OLS best-fit curve y=ys(x). Then, both and , so that the density of least squares apparently becomes indefinite.
This difficulty can be resolved by taking into account that the
measured data are subject to a limited resolution of measurement.
Therefore, although all data points lie on the best-fit curve at
the current resolution, it is more likely that their actual
values are scattered around it according to some "unknown''
sub-distribution that is beyond measurement resolution. In other
words, a quantity
assigned to this sub-distribution
exists, though not accessible by the measurement technique,
characterizes the data set. That being the case, one can rewrite
the expression for Dk(s) in the form
One can try to elaborate the choice of q by arguing that any
distribution, taken in narrow limits of stochastic variable,
should be nearly uniform. Consequently, in the narrow limits set
by the resolution of measurement, the presently assumed
sub-distribution should be uniform as well. Therefore, the
expected value of quantity q, given by Eq. (9), is
If the data set is accompanied by measurement errors ,
then the density of weighted least squares is given by Eq. (3). Technically, the foregoing analysis holds,
but when applied to relative distances
rather than to geometrical distances di. The density of
weighted least squares in the indefinite case is given analogously
to Eq. (11) by
We propose a procedure for rough estimation of the uncertainties of DLS best-fit parameters based on parameter db - the width of DLS best subset. The benefit is most evident when the data set is not accompanied by measurement errors. For the sake of simplicity, we assume that the DLS exponent k is set to 2, therefore .
(i) Measurement errors are not available in the original data set. After running DLS, we assign to all close points and use these estimated errors in a WLS fitting algorithm. The reported uncertainties we accept as the errors of the DLS best-fit parameters. For data sets containing a small number of points, it may happen that all close points lie on the OLS best-fit curve; i.e. we have to deal with the indefinite case. Since experimental errors are not given, there is no way to estimate uncertainties of the best-fit parameters even for k=2, unless the resolution of measurement dr is known. In this case, one can accept as a surrogate for missing experimental variances and proceed with WLS to estimate uncertainties of the best-fit parameters.
(ii) Measurement errors are available in the original data set. Instead of db, DLS will give which is related to the standard deviation of the relative residuals - therefore: . Note that, if the individual errors are correctly estimated, then or . If so, after running the WLS algorithm on the set of close points (with original ), the returned uncertainties are accepted. However, if is notably different from 1, it signifies that original measurement errors were badly estimated. In this case, the best one can do is to go back to the experiment, if possible. If not, it is safer to correct original errors than to proceed with wrong ones, especially if . The correction is quite simple: . If we run the WLS algorithm on the set of close points with corrected errors , the reported uncertainties will suit the data scatter. Finally, if the indefinite case occurs while processing a data set with known , it is enough to apply WLS fit on the subset of close points retaining original errors.
For the sake of simplicity, in all our examples we have taken k=2 and assumed that no experimental errors have been assigned to data points. The uncertainties are estimated from the data scattering, as discussed above.
In Fig. 3, we present a data set consisting of 11 moderately scattered points. By applying the DLS method to the family of straight-line model functions, we found 3 outlying points, while 8 points were recognized as close points. In order to find the best-fit parameters and the corresponding uncertainties, one can simply run WLS for the close points only. In the inset, we present the density of least squares versus width of the consecutive data subsets. The rightmost point represents the density of least squares for the whole data set. The next point represents the density of least squares after the first layer is removed, and so on. The density of least squares attains its maximum after removal of three layers, each containing one point. Therefore, the best subset contains 8 points, recognized as close points.
|Figure 3: Eleven moderately scattered points and the best-fit straight line according to the DLS method. Three points are recognized as distant points, i.e. outliers. The remaining eight points are close points. The best-fit parameters are only obtained on the subset of close points only. The inset displays the density of least squares, DLS, versus subset width. The maximum of DLS is clearly noticeable. The density of least squares for the whole data set is shown by the rightmost point. The density of least squares of the subset obtained after the first layer has been removed is presented by the next point and so on. We removed 6 layers, each containing 1 point. The width of the last set analyzed, containing n=5 points, is 0.03106.|
An artificially generated spectrum with three spectral lines is shown in Fig. 4. The magnitude of applied Gaussian noise is . To demonstrate stability and robustness of the DLS method, we fitted the horizontal base line with the fourth-order polynomials. Note the pronounced stability of DLS best-fit: though determined by five coefficients, the best-fit polynomial has the appearance of a horizontal line. For the sake of comparison, we also present the best-fit polynomial according to the OLS method and the best-fit polynomial obtained by the robust approach with a Lorentzian probability distribution. The difference is self-evident. The width of the best subset db=0.038 is quite close to the magnitude of the applied Gaussian noise, and it can serve as a distinguishing parameter: points that satisfy di>db belong to spectral lines and are treated as formal outliers, while data points satisfying di<dbare associated to the feature of interest, the base line in our example. The scattering of the data points di<db around the best-fit curve can be attributed to noise.
|Figure 4: Artificially generated spectrum with three spectral lines superimposed on line y(x)=0. The magnitude of the applied Gaussian noise is . Fourth-order polynomials are employed as base-line model functions. The solid gray line represents the DLS best-fit polynomial, with respective parameters given in the legend. The dashed line represents the OLS best-fit polynomial, while the solid black line represents the best-fit polynomial obtained by the robust approach with a Lorentzian probability distribution. The inset shows the density of least squares, DLS, versus subset width. The width of the best subset is db=0.038, which is almost the same as the magnitude of the applied noise. In DLS method, db determines the discrimination level; accordingly, spectral lines are recognized as (formal) outliers.|
In the next example we compare two methods of data fitting: our DLS method and the multi-component fit (MCF). Here we estimate only the parameters of the He II 468.6 nm Pashen alpha line ( ), recorded in a high-voltage pulsed discharge (Djenize et al. 2004) - see Fig. 5. In the captured spectrum, the central broad line is the line, while the two unambiguous peaks, left and right of the central line, correspond to the O II 467.66 nm and the O II 469.93 nm lines, respectively. Several low-magnitude peak-like crests visible on both wings are treated as stochastic fluctuations. Note also the presence of two typical outlying points, encircled in the lower panel, probably caused by occasional misfire of the high-voltage trigger. In this particular experiment, the high electron density of the discharge, induces a notable Stark effect. Therefore, the adequate model function for the spectral line is considered to be of the Lorentz profile type , depending on wavelength , and on the line parameters: height h, center c, and half-width w. Since the wavelength interval of the spectrum is rather narrow, it is enough to take the base line in the form of a straight line: .
|Figure 5: Comparison of DLS and multi-component fit (MCF) of a spectrum from a high-voltage pulsed discharge. The central line is He II 468.6 nm Pashen alpha line ( ). The two lateral peaks are the O II 467.66 nm and O II 469.93 nm lines. In DLS fit, the model function has the form , with a straight base line of slope b and y-intercept a and a superimposed Lorentzian profile of height h, center c, and half-width w, describing the He II line. In MCF, we have also superimposed two Lorentzians, corresponding to the O II lines. The obtained DLS and MCF best-fit parameters coincide within the limits of their errors.|
In this problem, the proper DLS model function only consists of the base line and the superimposed He II Lorentzian, i.e. . In contrast, if we apply the multi-component fit, two additional Lorentzians, corresponding to the lateral O II lines, have to be included as well; hence, the proper MCF model function is: . One can notice that sets of -line parameters, provided by the DLS fit and by MCF, coincide within the limits of their uncertainties. Note also that, though the base line was an auxiliary feature (not shown in Fig. 5), both methods give a fair agreement between the baseline parameters: the y-intercept a and slope b.
The specifics of both methods are clearly outlined by the performed parallel fittings. The multi-component fit necessitates a comprehensive form of model function with all (known) features taken into account. In contrast, only the features of interest are incorporated into the DLS model function, as the DLS merit function enables removal of distant points. Some of these points belong to the spectral lines, some are typical outliers, while the rest of them cannot be easy classified. However, when a large number of undocumented features and true outliers are expected, instead of the foregoing straightforward MCF, the use of a multi-component model function within the DLS approach is recommended to reduce the risk of inaccurate parameter estimation.
In Fig. 6 we present an astrophysical spectrum - Far Ultraviolet Spectroscopic Explorer (FUSE) composite spectrum, originally reported in (Scott et al. 2004). The spectrum continuum is supposed to be of a power-law type . The best-fit power-law exponent quoted in (Scott et al. 2004) is with standard deviation 0.11. We emphasize that the reported estimate is obtained in a fit performed on selected wavelength regions, free of emission lines: 630-750, 800-820, 850-900, 1095-1100, and 1135-1150 Å. The solid black line in Fig. 6 shows this fit in wavelength scale, with the power law and exponent .
Contrary to the procedure adopted in Scott et al. (2004), the DLS fit is performed on the whole spectrum range (630-1155 Å) of the rest-frame wavelengths; the resulting best-fit curve is represented by the solid gray line in Fig. 6. The DLS best-fit exponent is , while the standard deviation, relying on the inherent noise, is estimated to 0.034. Without any prior knowledge taken into account, the DLS best-fit value is slightly less than the one quoted originally. It is indicative that the two tiny wavelength intervals 1095 Å-1100 Å and 1135 Å-1150 Å, (data in black line on the very right of Fig. 6), recognized in (Scott et al. 2004) as regions of spectrum free of emissions lines, have a significant impact on the best-fit value of the power-law exponent. Inclusion of these intervals into the continuum can be justified only if data pertaining to those intervals are reliable and based on a priori knowledge, which obviously has priority over any general procedure.
However, if we run DLS fit exactly on the same intervals as Scott et al., the trough at about 720 Å is attributed again to distant points signaling that this feature should be considered as a candidate for rejection, even if the corresponding wavelength interval is free of contamination. As mentioned in the preceding example, by applying the DLS method independently, an unbiased review and further insight into the considered problem could be gained.
|Figure 6: FUSE composite spectrum originally reported in (Scott et al. 2004). Smooth solid lines represent best-fits of the spectrum continuum to the power-law form . Scott et al. have applied the fit only to selected wavelength intervals (marked in black) that are free of spectral lines. Their best fit (black line) corresponds to . The best power law according to DLS (gray line) is obtained on the whole interval of wavelengths, and it corresponds to . Data for DLS fit are recovered from the original graph with limited accuracy, allowing only a rough comparison of the two approaches.|
|Figure 7: Time-integrated spectrum of an argon high-voltage pulsed discharge recorded by a pocket-sized Ocean Optics spectrograph in our plasma spectroscopy laboratory. The most intensive spectral lines are clipped, in order to provide a noticeable continuum. By DLS fitting of the spectrum continuum with Planck functions (15), we determined the time-averaged thermodynamic temperature K. Its small uncertainty reflects a low scattering of close points.|
In Fig. 7 we present a time-averaged spectrum recorded
by a miniature Ocean Optics spectrograph during a single pulse of
an argon high-voltage pulsed discharge, operating in the Plasma
Spectroscopy Laboratory of the Faculty of Physics, University of
Belgrade. The spectrum covers the wavelength range from m
(near UV) to m (near IR). The created argon plasma is in
a state close to thermodynamic equilibrium (Djenize et al. 2006),
generating a spectrum with a large number of spectral lines and a
sizeable continuum. Consequently, the continuum profile defines
the base line of the spectrum. The simplest model function follows
from the Planck law of black body radiation, which relates the
intensity of radiation I at wavelength
thermodynamic temperature T. Within this example, we employ
Planck formula in the form
In Sect. 2 we defined the merit function (2) on subsets s of a given data set s0 as arguments. Since the number of such subsets is typically huge (2n0, n0 is the number of points in s0), it is evident that the search for the merit function minimum by direct inspection would be extremely inefficient. Therefore, some substitute for the standard gradient methods, used in minimization of ordinary functions (e.g. the steepest descent method), has to be proposed.
Suppose that a data set s0, comprised of n0 points, is given. In order to mimic a single step of a gradient method, we remove point p from s0 so that the subset s0(p), obtained by its removal, has the minimum value of merit function among all n0 subsets of s0 containing n0-1 points. Next, we repeat the foregoing procedure on the subset s0(p)instead of on s0, and so on. Though it is the closest equivalent to the standard gradient minimization method, this simple iterative procedure is still inefficient because, without some additional criteria, the total number of subsets to be scanned (n02/2) is still large.
However, the structure of merit function (2) implies that the outer points are the best candidates for removal. Indeed, the greater the distance |yp-f(xp,a(s))| of a discarded point from the OLS best-fit curve y=f(x,a(s)), the greater the local chances of proceeding to the merit function minimum. Accordingly, if we continue to remove the outermost points, the overall chances of achieving the minimum of merit function (2) are greatest. Any other choice would reduce the prospects of reaching the minimum. Therefore, the ordered collection , introduced in Sect. 2, is an algorithmically optimal assortment of subsets, which provides an efficient way to find the minimum of merit function (2). As in the case of standard minimization, there is no full guarantee that a minimum found within the ordered collection of subsets is the absolute one.
The rough estimate of errors in the best-fit parameters, introduced in Sect. 4, is based on the assumption that both - ordinary minimization algorithms employed by DLS, as well the DLS itself - are exactly correct. The full confidence limits on DLS best-fit parameters can be determined by the Monte-Carlo method (Press et al. 2001) applied to the set of close points, using estimated according to Sect. 4.
An important feature of the DLS method is its use of some OLS algorithm as a basic engine. Consequently, for a very broad class of model functions depending linearly on their parameters (e.g. polynomials), the solution is achieved without initial guess. This is a serious advantage with respect to the original CPC method (Bukvic & Spasojevic 2005), which requires the explicit use of a minimization algorithm for all model functions, except for the family of horizontal lines . Nevertheless, if the model function is a nonlinear function of its parameters, then the explicit use of a minimization algorithm, like the Levenberg-Marquardt algorithm (Press et al. 2001), is necessary.
In this paper, we have proposed a new form of a merit function, which measures the density of least squares (DLS) for subsets of a given data set. The best-fit parameters are obtained by the ordinary least-square method on a subset having a maximum density of least squares. For a rapid approach to DLS maximum a simple algorithm is put forward. Owing to its extreme robustness to the presence of outlying points, DLS method enables flexible data processing in which only features of interest are modeled. The consequence is a notable reduction of fitting parameters and an increased numerical stability, which are important in analyses of complex real-world spectra.
The efficiency of DLS method has been illustrated by several examples. Particularly persuasive is the low uncertainty estimate of thermodynamic temperature from the continuum radiation of a pulsed discharge.
The DLS method of data fitting allows plenty of extensions, which we intend to study in the future.
This work was supported by the Serbian Ministry of Science under projects 141010 and 141014 and by the Slovenian Research Agency, under program P2-0056.
From the statistical point of view, many of the data-fitting methods are local-M estimators. The basic assumption of the local-M approach is that the probability density P'(ym,x), of obtaining a measured value ym at given x only depends on the relative deviation from the model function f(x;a), so that P'(ym,x)=P'(z(a)); here, a stands for the true model function parameters, whereas is the uncertainty of ym.
Assuming further that yi's are independent stochastic variables, the overall probability density (of obtaining a data set in a measurement) is .
In practise, the true model function parameters are not known in advance. In the maximum likelihood estimation, the probability density P(a) is taken as a likelihood of parameters, so that the best-fit parameters are those that maximize P(a). Equivalently, one can minimize the merit function , commonly written as , where .
The best known local-M method is the OLS method. Here , so that the corresponding probability distribution of relative deviations is the Gaussian distribution, . The consequence is that the OLS method provides the best-fit parameters assuming a Gaussian noise. If, however, the noise conforms to some other distribution, then the OLS method will not be the most suitable; for example, in the case of a double-exponential noise, , it is most appropriate to use a local-M estimator with , i.e. to minimize the sum of absolute deviations. When applied to data contaminated by outliers, this least-absolute-deviation method is more robust than the OLS method because it gives less weight to large deviations.
By assigning less and less weight to suspiciously large
deviations, Lorentzian and Tukey's biweight methods
provide increasingly robust estimates. Eventually, one can decide
to discard suspicious points altogether, like in the
least-trimmed-square (LTS) method - see
(Rousseeuw 1984). The LTS method is based on the metric
trimming local-M estimator
In what follows, we relate the proposed DLS to the LTS method. To
this end, we first demonstrate some of LTS basic features on a
pure Gaussian noise (4) with a
known standard deviation
lines y(x)=a taken as model functions. It follows that the
merit function is a sum of two terms,
However, the close-points contribution exhibits a less simple behavior. Thus, for large c's, has two equal and symmetrically positioned maxima. Between the maxima, there is a single local minimum, located at a=0; cf. Fig. 8. For decreasing c, i.e. for narrowing the boundary, the minimum shallows and the adjacent maxima flatten approaching each other, and eventually merge at a=0 as cattains its critical value cc. At the value cc, which coincides with the solution of Eq. (6) for k=2, the local minimum vanishes.
In the region , the function exhibits a single absolute maximum at a=0; cf. Fig. 8. Hence, one can find the best-fit parameters by maximizing the close-points contribution to the merit function, instead of minimizing the merit function itself. We will use this fact as a link between the LTS and DLS methods.
|Figure 8: LTS method: close-points contribution to merit function, versus horizontal line level a for Gaussian distribution of deviations. The five presented curves in descending order correspond to the boundary c values 1.7, 1.5, 1.36876, 1.3, and 1.1, respectively. The critical value cc=1.36876, separates the two modes of behavior. For c>cc, the function exhibits two maxima, separated by a local minimum at a=0; when approaching the critical value cc from above, the two maxima merge with the local minimum. For , the function has an absolute and single maximum at a=0.|
Now, let us take , where db is the width of the best subset found by the DLS method for a given k - see Sect. 3.1. It can be recognized that the best subset sb, having maximum density of least squares D(s), is actually the set of c-close points. Furthermore, the LTS best-fit parameters are the OLS best-fit parameters, obtained on the set of c-close points. They coincide with DLS best-fit parameters, obtained by the same method (OLS), and on the same subset ( ). This indicates that DLS best-fit parameters are in fact the maximum likelihood parameters for a suitably chosen boundary value c separating close and distant points in the LTS method.
At the end of this section, we would like to point out one major difference between two methods, which is relevant, in particular, when neither the width of noise nor the extent to which outliers have contaminated the experimental data are known. In the LTS method, the boundary c perfectly suits tuning the extent of contamination, having the role of some kind of confidence level, but requires the width of noise to be known in advance. Though less obvious, the exponent k within the DLS method serves the same purpose, with an essential advantage that in the DLS method we do not need to know the width of noise in advance. Instead, the width is estimated by the DLS method itself, together with best-fit parameters.
In this section we present a subroutine DLSFIT that implements the DLS method. The subroutine DLSFIT is not self-contained, because it needs a linear and a nonlinear weighted least-square (WLS) subroutine to be supplied as working engines. For this purpose, we have chosen well-documented (Press et al. 2001) subroutines: lfit (for general-linear WLS data fitting) and mrqmin (for nonlinear WLS data fitting). These routines, together with the subroutines: zbrent, gaussj, covsrt, and mrqcof (called by DLSFIT, lfit and mrqmin), are available on-line on the official Numerical Recipes (NR) site http://www.nr.com, governed by Cambridge University Press.
Here we present a FORTRAN code of DLSFIT subroutine. To keep it as short as possible, we use concise FORTRAN90 syntax. For the same reason we refer to the simply portable FORTRAN77 (version 2.10) code of subroutines zbrent, gaussj, covsrt, and mrqcof - see, for instance, http://www.nrbook.com/a/bookfpdf/f15-7.pdf. The program which calls DLSFIT should be compiled in FORTRAN90 compiler environment with DOUBLE PRECISION as default real kind.
SUBROUTINE DLSFIT(x,y,sig,n0,a,m,rp,& & chisq,dls,db,ns,spos,sbno,& & covar,ia,funcs,& & lf,k,res) ! Remove the next line if Compaq compiler is not used !USE numerical_libraries !Compaq compiler directive REAL,PARAMETER:: Pi=3.1415926535897932384626433832795D0, q=1./3. INTEGER ia(m),spos(n0),sbno REAL x(n0),y(n0),sig(n0),k REAL a(m),covar(m,m),dyda(m),afunc(m),da(m) EXTERNAL funcs LOGICAL lf IF(k>2.9.OR.k<2.)RETURN dtiny=(MAXVAL(y)-MINVAL(y))/10.**PRECISION(db);nt=n0;ns=1;dls=0. CALL fit(x,y,sig,nt,a,ia,m,covar,chisq,funcs,lf) DO WHILE (nt>m+3) CALL NewMeritVals;IF(dbt==0) EXIT CALL ArrangeLayer END DO CALL fit(x,y,sig,spos(sbno),a,ia,m,covar,chisq,funcs,lf) CONTAINS SUBROUTINE ArrangeLayer ! Rearranges arrays x,y,sig so that their first spos(i) ! elements belong to the i-th subset in ordered collection dr=rp*dbt;ns=ns+1 DO i=1;nto=nt DO WHILE (i<=nt) !Remove sublayer yc=fun(x(i)) IF(ABS(y(i)-yc)/sig(i)>=dr)THEN yc=fun(x(nt)) IF (ABS(y(nt)-yc)/sig(i)<dr)THEN swap=x(i);x(i)=x(nt);x(nt)=swap swap=y(i);y(i)=y(nt);y(nt)=swap swap=sig(i);sig(i)=sig(nt);sig(nt)=swap END IF nt=nt-1 ELSE i=i+1 END IF END DO IF((nt==nto).OR.(nt<=m+3))EXIT CALL fit(x,y,sig,nt,a,ia,m,covar,chisq,funcs,lf) END DO END SUBROUTINE ArrangeLayer SUBROUTINE NewMeritVals ! Calculates new DLS value spos(ns)=nt;dbt=0 DO i=1,nt d=ABS(y(i)-fun(x(i)))/sig(i); IF(dbt<d) dbt=d END DO IF(dbt>dtiny)THEN dlst=chisq/(dbt**k) ELSE os2=SUM((/(1./(sig(i)*sig(i)),i=1,nt)/))/nt dlst=os2*(1.+q*(nt-1))*res**(2-k);dbt=0 END IF IF(dlst>dls)THEN dls=dlst;sbno=ns;db=dbt END IF END SUBROUTINE NewMeritVals FUNCTION fun(x) ! For given x, calculates value of model function IF(lf)THEN CALL funcs(x,afunc,m);fun=DOT_PRODUCT(a,afunc) ELSE CALL funcs(x,a,fun,dyda,m) END IF END FUNCTION fun END SUBROUTINE DLSFIT
SUBROUTINE fit(x,y,sig,n0,a,ia,m,covar,chisq,funcs,lf) ! If lf=.TRUE. performs linear fit by calling lfit ! Otherwise, drives mrqmin which performs non-linear fit DIMENSION x(n0),y(n0),sig(n0) DIMENSION a(m),ia(m),covar(m,m),al(m,m) REAL:: lamda,lamdaO EXTERNAL funcs LOGICAL lf PARAMETER (MaxIt=100,RelErr=1.D-5) IF(lf)THEN CALL lfit(x,y,sig,n0,a,ia,m,covar,m,chisq,funcs);RETURN END IF lamda=-1;chisqO=0;lamdaO=0;it=0 DO WHILE (it.LE.MaxIt) CALL mrqmin(x,y,sig,n0,a,ia,m,covar,al,m,chisq,funcs,lamda) IF(chisq<=0)EXIT IF((ABS(chisqO-chisq)/chisq<RelErr).AND.(lamda<lamdaO))EXIT chisqO=chisq;lamdaO=lamda;it=it+1 END DO lamda=0. CALL mrqmin(x,y,sig,n0,a,ia,m,covar,al,m,chisq,funcs,lamda) END SUBROUTINE fit
The user supplies to DLSFIT subroutine a set of n0 experimental data points x(1:n0), y(1:n0), with individual standard deviations sig(1:n0) of their yi's; if sig(1:n0) are not available, then set all sig(i)=1 prior to DLSFIT call. The next argument in the DLSFIT calling list is array a(1:m) of length m. On input, the user supplies initial values of model function parameters in a(1:m) in the case when the model function is a nonlinear function of its parameters; otherwise, the input values supplied in a(1:m) are irrelevant. On output, the main result of DLS method - the best-fit parameters associated with the best subset - are returned in a(1:m).
The input parameter rp is the removal parameter, introduced in Sect. 2. In our experience, when 0.9 rp 1 the best-fit parameters returned for different values of rp coincide within the error margins.
Six parameters (chisq, dls, db, ns, spos, sbno), which follow in DLSFIT calling list, are output parameters. Thus, chisq is and dls is the density of least squares obtained on the best subset in the ordered collection (see Sect. 2), found by DLSFIT, while db is the width of the best subset. Additionally, ns is the total number of subsets in , spos(1:ns) is an integer array that contains the number of data points in each subset from , and sbno is the ordinal number of the best subset in .
Note that the arrays x(1:n0), y(1:n0) and sig(1:n0) are reordered on output so that the first spos(i) data points belong to the subset . In this way, the user can obtain relevant data for each subset by applying the OLS method on the first spos(i) elements of arrays x, y, and sig.
The next three parameters (covar, ia, funcs) in the calling list of the DLSFIT subroutine are specific to the applied lfit and mrqmin NR subroutines. Thus, covar(1:m,1:m) is a two-dimensional array, which on output contains the elements of covariance matrix. The integer array ia(1:m) provides a convenient way to keep some of the parameters frozen on their input values (ia(j)=0 for frozen a(j), otherwise ia(j) is non-zero); the values of all frozen parameters must be supplied on input in array a. Finally, funcs is the name of a user-supplied subroutine (fpoly from Press et al. 2001 in the case of polynomial fit) proceeded to DLSFIT when DLSFIT is called; when a general-linear fit is performed, then funcs(x,afunc,m) should return the m basis functions evaluated at x in the array afunc(1:m). In the opposite case, funcs is a user supplied subroutine having syntax funcs(x,a,y,dyda,m). For given x and model function parameters a, supplied in array a(1:m) of length m, it returns the calculated value f(x;a) of model function in variable y, together with values of all model function partial derivatives , returned in array dyda(1:m).
The last three parameters in DLSFIT calling list are lf, k, and res. The first of them, lf, is a LOGICAL input parameter. When lf is set to .TRUE., then DLSFIT performs general-linear WLS data fitting by calling lfit subroutine; otherwise, nonlinear fit is performed with an aid of mrqmin subroutine. The value of DLS exponent k is supplied to DLSFIT subroutine in the second parameter of the group, k. Finally, we must also supply the value of (or ) used in DLS calculations in the indefinite case - see Sect. 3.2. This is supplied through parameter res - the original resolution of measurement.
Note that the DLSFIT does not return errors in the best-fit parameters. Estimation of errors, detailed in 4, is left to the user as it depends on the specifics of actual data; nonetheless, a full-code example is given in the online material only available in electronic form at the CDS.
At the end, we explain the subroutine fit, called by DLSFIT routine. The calling parameters have the same meaning as for the DLSFIT routine. If model function depends linearly on its parameters, then the subroutine fit calls NR routine lfit, which performs the general linear WLS data fitting. Otherwise, it serves as a simple driver for the NR routine mrqmin, used for a nonlinear WLS data fitting. The subroutine fit is singled out from DLSFIT in order to provide a way for flexible estimation of the best-fit parameters on any subset from the ordered collection of subsets obtained by DLS method.