611 INTRODUCING ADAPTED NELDER & MEAD’S DOWNHILL SIMPLEX METHOD TO A FULLY AUTOMATED ANALYSIS OF ECLIPSING BINARIES A. Prsa,ˇ T. Zwitter University of Ljubljana, Dept. of Astrophysics, Jadranska 19, 1000 Ljubljana, Slovenia ABSTRACT of Multiple Subsets (MMS) (Wilson 1993). The DC method has already been applied successfully to auto- matic parameter extraction (see e.g., Wyithe & Wilson Eclipsing binaries are extremely attractive objects be- 2001, 2002; Prsaˇ 2003), but its original philosophy is cause absolute physical parameters (masses, luminosi- based on interactive monitoring of each convergence step. ties, radii) of both components may be determined from The algorithm is very fast and works well if the discrep- observations. Since most efforts to extract these param- ancy between the observed and computed curves is rel- eters were based on dedicated observing programmes, atively small, but it tends to diverge or give physically existing modelling code is based on interactivity. Gaia unmeaningful results if the discrepancy is large. As part will make a revolutionary advance in shear number of of an effort to create a reliable and powerful package for observed eclipsing binaries and new methods for auto- EB analysis, a complementing minimization scheme is matic handling must be introduced and thoroughly tested. proposed. This paper focuses on Nelder & Mead’s downhill simplex method applied to a synthetically created test binary as it will be observed by Gaia. 2.1. Nelder & Mead’s Downhill Simplex Key words: Binaries: eclipsing; Methods: numerical, sta- There are two main deficiences of the DC method that are tistical; Missions: Gaia. especially striking. 1) Once a DC method converges to a minimum, there is no way of telling whether that mini- mum is local or global; even if it is local, the method is 1. INTRODUCTION stuck and cannot escape. 2) The main source of diver- gence and the loss of accuracy in the DC algorithm is the computation of numerical derivatives. Many assessments have already been done for Gaia’s expected harvest of eclipsing binaries (EB) to V < 15, To circumvent these two problems, Nelder & Mead’s where both photometric and RV observations will be downhill Simplex1 method (Nelder & Mead 1965), here- available (see Munari et al. 2001, Zwitter et al. 2003, after NMS, is implemented. Since NMS doesn’t compute Marrese et al. 2004 and others). Out of 50 million ob- derivatives but relies only on function evaluations, it is served stars, roughly 100 000 will be double-lined eclips- a promising candidate for our purpose. The basic form ing binaries. However, based on experience from Hippar- of the NMS method along with WD implementation was cos, out of 1 billion stars observed to V<20, there will be first proposed by Kallrath & Linnell (1987). We take a 2 million (Eyer 2005) eclipsing binaries without spec- step further and adapt the method specifically to EBs. troscopic» observations, but with quite decent photomet- ric accuracy. This study presents a new approach devel- The NMS method acts in n-dimensional parameter hy- oped for automatic reduction of observed data along with perspace. It constructs n vectors pi from the vector of an estimate of how much we may expect to obtain from initial parameter values x and the vector of step-sizes s them. as follows: pi = (x0; x1; : : : ; xi + si; : : : ; xn) (1) 2. THE METHOD These vectors form (n + 1) vertices of an n-dimensional simplex. During each iteration step, the algorithm tries Obtaining physical parameters from observations is an in- to improve parameter vectors pi by modifying the ver- verse problem solved numerically by a modelling pro- tex with the highest function value by simple geometrical gramme. Affirmed and most widely used is the WD 1Nelder & Mead’s downhill simplex is not to be confused with linear code (Wilson & Devinney 1971), which features Differ- or non-linear programming algorithms, which are also referred to as ential Corrections method (DC) powered by the Method Simplex methods. 612 transformations: reflection, reflection followed by expan- where index i runs over M measurements within a sin- sion, contraction and multiple contraction (Galassi et al. gle data set and index k runs over N data sets (different 2003). Using these transformations, the simplex moves photometric curves). Since the variance is given by: through parameter space towards the deepest minimum, 2 1 2 where it contracts itself. sk = (xi x¹) ; (3) Nk 1 ¡ ¡ Xi PHOEBE2 is a software package built on top of the WD 2 code that extends its basic functionality to encompass, we may readily express Âk as: among other extensions summarized in Prsaˇ & Zwitter s2 (2005), the NMS method. It is especially suited for EBs: 2 k Âk = (Nk 1) 2 (4) powered by heuristic scans, parameter kicking and con- ¡ k ditional constraining, the method is able to efficiently es- and the overall Â2 value as: cape from local minima. 2 2 sk  = (Nk 1) (5) ¡ µk ¶ 2.2. Heuristic Scan Xk If are fair and all data sets contain approximately the NMS is a robust method that always converges, but it can k same number of observations, then the ratio sk/k is of converge to a local minimum, particularly since param- the order unity and Â2 of the order NM. This we use for eter hyperspace in the vicinity of the global minimum is parametrizing Â2 values: typically very flat, with lots of local minima. In addition, 2 2 2 the global minimum may be shadowed by data noise and Â0 = NM; ¸ :=  =Â0 : quantization. (6) degeneracy. ¡ ¢ Parameter kicking is a way of knocking the obtained The heuristic scan is a method by which a minimiza- parameter-set out of the minimum: using the Gaussian tion algorithm selects a set of starting points in parameter PDF, the method randomly picks an offset for each pa- hyperspace and starts the minimization from each such rameter. The strength of the kick is determined by the point. It then sorts all solutions by the cost function (the Gaussian dispersion , which depends on ¸: if the Â2, for example) and calculates parameter histograms and kick value is high, then the kick should be strong, but if it is convergence tracers for given hyperspace cross-sections low, i.e., around , then only subtle perturbations (specific examples are given in Section 4). The way of ¸ 1 should be allowed. »Experience shows that a simple ex- how the algorithm selects starting points is determined by pression such as: the user: the points may be gridded, stochastically dis- persed, distributed according to some probability distri- 0:5¸ = (7) bution function (PDF) etc. The basic idea of the heuristic kick 100 scan is to obtain decent statistics of adjusted parameter values from which a fair and realistic error estimate may works very reliably. This causes kick to assume a value be given. of 0.5 for 10 offsets and 0.005 for 1 offsets, being lin- ear in between. Note that this kick is relative, i.e., given by: abs rel (8) 2.3. Parameter Kicking kick = x kick; where x is the value of the given parameter. Another possible approach to detect and escape from lo- cal minima is to use some stochastic method like Simu- 2.4. Conditional Constraining lated Annealing (SA). However, such methods are notori- ously slow. Since the EB hyperspace is typically very flat, stochastic methods would be practical only in the vicin- Having purely photometric (LC) observations, it is im- ity of the global minimum. Thus instead of a full-featured possible to determine absolute physical parameters of the SA scan, a simple new procedure has been developed that observed binary. However, if the distance to EB is mea- achieves the same effect as stochastic methods, but in sig- sured independently or if additional assumptions about nificantly shorter time. the EB are set, even purely photometric observations can yield absolute values of physical parameters. If addi- The idea is as follows: whenever NMS reaches a min- tional constraints are imposed on the model from the out- imum within a given accuracy, the algorithm runs a side, the model is referred to as conditionally constrained globality assessment on that minimum. If we presume (CC’d). that standard deviations k of observations are estimated properly and that they apply to all data points regardless Two CCs are immediately evident: 1) astrometric mea- of phase or nightly variations, we may use them for Â2 surements on-board Gaia may be used to measure the weighting: distance with the accuracy of 11 ¹as at V = 15 to » M M 165 ¹as at V = 20 (ESA 2000); 2) since a substan- 1 tial» number of stars are main-sequence stars, M–L, L–T Â2 = w (x x¹)2 = (x x¹)2; (2) k k i 2 i and T–R relations may be adopted as constraints. See ¡ k ¡ Xi=1 Xi=1 Prsaˇ & Zwitter (2005) for details on CC implementation 2http://phoebe.fiz.uni-lj.si in PHOEBE. 613 3. SIMULATION To test the suitability of NMS for EBs, we built a partially-eclipsing synthetic main-sequence F8 V–G1 V binary using PHOEBE. Table 1 lists all of its princi- pal parameters. The simulation presented in this paper is based exclusively on photometric data: two sets of observations corresponding to Johnson B and V filters are created, each with 82 points with Poissonian scatter obs = 0.02 mag at V = 20, values typical to expect from Gaia (ESA 2000). Simulation flow is the following: all physical parameters Figure 1. ¸-histograms for 3 consecutive parameter kicks. Top-left figure demonstrates how numerous are lo- set for adjustment (a, i, T1, T2, ­1, ­2, L1 and L2) were displaced by 50%.