1: \documentclass[12pt,preprint]{aastex}
2:
3: \usepackage{natbib}
4: %\usepackage{lscape}
5:
6: \def\SNR{S/N}
7: \newcommand{\lta}{\lesssim}
8: \newcommand{\gta}{\gtrsim}
9: \def\arcdeg{\hbox{$^\circ$}}
10: \def\arcmin{\hbox{$^\prime$}}
11: \def\arcsec{\hbox{$^{\prime\prime}$}}
12: \def\fd{\hbox{$.\!\!^{\rm d}$}}
13: \def\fh{\hbox{$.\!\!^{\rm h}$}}
14: \def\fm{\hbox{$.\!\!^{\rm m}$}}
15: \def\fs{\hbox{$.\!\!^{\rm s}$}}
16: \def\h{\hbox{$^{\rm h}$}}
17: \def\m{\hbox{$^{\rm m}$}}
18: \def\s{\hbox{$^{\rm s}$}}
19: \def\fdg{\hbox{$.\!\!^\circ$}}
20: \def\farcm{\hbox{$.\mkern-4mu^\prime$}}
21: \def\farcs{\hbox{$.\!\!^{\prime\prime}$}}
22:
23: \shorttitle{Method for Extracting Light Echo Fluxes}
24: \shortauthors{Newman and Rest}
25:
26: \begin{document}
27:
28: \title{A Method for Extracting Light Echo Fluxes Using
29: the NN2 Difference Imaging Technique}
30:
31: \author{A.~B.~Newman}
32: \affil{Department of Physics, Washington University}
33: \affil{1 Brookings Drive, St. Louis, MO 63130}
34: \email{abnewman@wustl.edu}
35:
36: \and \author{A.~Rest\altaffilmark{1}}
37: \affil{Cerro Tololo Inter-American Observatory}
38: \affil{National Optical Astronomy Observatory}
39: \affil{950 North Cherry Avenue, Tucson, AZ 85719}
40: \email{arest@ctio.noao.edu}
41: \altaffiltext{1}{Goldberg Fellow.}
42:
43: %%%%%%%%%%%%%%%%%%%%%%%
44: %%%%%%%%%%%% ABSTRACT
45: %%%%%%%%%%%%%%%%%%%%%%%
46:
47: \begin{abstract}
48: Light echoes are interesting because of the wealth of information they
49: offer about their progenitors and the reflecting dust. Due to their
50: faint surface brightnesses, difference imaging is
51: necessary to separate most light echoes from the sky background. However, difference
52: images reveal only the relative fluxes between two
53: epochs. Obtaining absolute fluxes for individual epochs has traditionally
54: relied on a single template image that is free of light echoes.
55: Since such an observation is normally unavailable, a light echo-free template
56: must be constructed by a complicated and usually subjective process.
57: Here we present an application of the NN2 method of
58: Barris et al.~to extract the relative fluxes of light echoes
59: across a range of epochs directly from a series of difference images. This
60: method requires no privileged image and makes maximal use of the
61: observational data. Statistical methods to estimate the zero-flux
62: level and thus the absolute flux are also presented. The efficacy of
63: the technique is demonstrated by an application to the light echoes
64: around SN 1987A. The resulting images reveal new detail and faint
65: light echo structures. This method can be adapted and applied to other
66: extended variable light sources, such as stellar outflows and supernova remnants.
67: \end{abstract}
68:
69: \keywords{methods: data analysis --- supernovae: general ---
70: techniques: image processing --- techniques: photometric}
71:
72: %%%%%%%%%%%%%%%%%%%%%%%
73: %%%%%%%%%%%% TEXT
74: %%%%%%%%%%%%%%%%%%%%%%%
75:
76: %%%%%%%%%%%%%%%%%%%%%%
77: %%%%%%%%%%%% INTRO
78: %%%%%%%%%%%%%%%%%%%%%%
79:
80: \section{Introduction}
81: \label{sec:intro}
82: Light echoes arise when light from a stellar event is reflected toward
83: the observer by intervening dust. The arrival of the light is delayed
84: by the extra distance it traverses and so may be observed long after
85: the source event. Light echoes provide a powerful tool for studying
86: interesting stellar phenomena and sensitively probing dust structures.
87: For many dust geometries of interest, the light echoes
88: appear as rings or arclets which usually expand superluminally \citep{Couderc39}.
89: %\citet{Tylenda04} provides a useful review of light echo geometry.
90: Light echoes from variable stars have been observed, including Nova
91: Persei 1901 \citep{Ritchey1901}, the source of the first discovered light echo,
92: and the mysterious eruptive variable V838 Monocerotis \citep[e.g.,][]{Munari02,Bond03}.
93: However, only supernovae (SNe) have sufficient luminosity to produce observable
94: scattering from interstellar dust \citep{Sugerman03}. Most light echo
95: discoveries, therefore, have been associated with SNe,
96: including SN 1991T \citep{Schmidt94}, SN 1993J \citep{Sugerman02},
97: SN 1998bu \citep{Cappellaro01}, SN 2003gd \citep{Sugerman03:discov}, and
98: %in the Large Magellanic Cloud \citep{Rest05:le}. The most famous light echoes are
99: %those surrounding SN 1987A, which
100: most famously SN 1987A, whose light echoes have been used to investigate in
101: detail the geometry, density, and composition of the circumstellar and interstellar
102: media \citep{Crotts91, Xu94, Xu95, Crotts95, Xu99, Sugerman05:II}.
103: Contributions from light echoes may
104: significantly affect the observed light curves of Type II SNe
105: \citep[e.g.,][]{Chugai92,DiCarlo02}. Polarization measurements of light echoes
106: can provide an independent distance measurement when scattering is nearly
107: $90^{\circ}$ \citep{Sparks94, Sparks96}. \citet{Rest05:le} have analyzed the
108: spectra of light echoes from ancient SNe to determine the types of their
109: progenitors. Although the wealth of information they contain has inspired
110: searches \citep[e.g.,][]{Boffi99}, most light echoes are difficult to image
111: due to their low surface brightnesses.
112:
113: %By observing light echoes, it is possible to
114: %infer information about the geometry and composition of circumstellar
115: %matter and the interstellar medium \citep[e.g.][]{Xu95,
116: %Crotts95,Sugerman05}. Light
117: %echoes provide a laboratory for studying important stages in stellar
118: %evolution, particularly supernovae, even at much later epochs
119: %\citep{Rest05:le}. Although much can be gleaned from the dynamical
120: %evolution of a light echo, absolute surface brightness measurements
121: %provide additional information, including distance
122: %constraints \citep[e.g.][]{Tylenda04} and dust density estimates
123: %\citep{Chevalier86, Sugerman03}. These
124: %data, however, have been difficult to obtain, principally due to the
125: %low surface brightness of the echoes.
126:
127: Sophisticated difference imaging techniques
128: \citep[e.g.,][]{Tomaney96,Alard98,Alard00,Sugerman05:II,Rest05:le} allow light echoes
129: to be cleanly separated from the sky background, but difference imaging alone
130: can reveal only the relative flux between two epochs. Each difference image
131: shows the flux difference between an image epoch and a template epoch.
132: If a light echo-free image is available, it can serve as a template
133: that can be subtracted from every epoch to isolate the light echoes.
134: Unfortunately, such an observation usually does not exist.
135: Previous studies \citep[e.g.,][]{Xu95, Sugerman05:II} have therefore focused on
136: constructing a light echo-free image. The construction involves
137: a rather complicated iterative process, requires some subjective decisions,
138: and tends to leak faint light echoes into the echo-free template.
139: Another possibility is to create difference images from epochs that both
140: contain light echo flux. In this case,
141: the sky background is eliminated in the difference image, but both positive and negative
142: light echo fluxes remain. Figure~\ref{fig:diffim_example}
143: shows a sample difference image in which these positive and negative
144: fluxes are represented by white and black pixels, respectively.
145: Separating these frequently overlapping signals is a serious challenge.
146:
147: In this paper, we present an application of the NN2 algorithm of
148: \citet{Barris05} to calculate relative light echo fluxes across a
149: range of epochs directly from a series of difference images.
150: This technique requires no light echo-free template
151: and treats all observations equally. It thus
152: makes maximal use of all the observational data.
153: We also present a statistical method for estimating the zero-flux level
154: of each pixel from the relative light curve, thus determining the absolute flux.
155: The entire procedure is tested on the light echoes around SN 1987A,
156: and the resulting images reveal intricate detail and faint structures.
157:
158: %%%%%%%%%%%%%%%%%%%%%%
159: %%%%%%%%%%%% OBSERVATIONS
160: %%%%%%%%%%%%%%%%%%%%%%
161: \section{Observations and Reduction}
162: \label{sec:observations}
163: To test the ability of our technique to accurately recover light echo
164: fluxes, the light echoes around SN 1987A were chosen for
165: analysis. Observations were performed by the SuperMACHO survey on the Victor
166: M.~Blanco 4m Telescope at Cerro Tololo Inter-American Observatory
167: (CTIO). Imaging of the Large Magellanic Cloud (LMC) was performed
168: every other night in dark time during the months of October, November,
169: December, and January, with occasional images from September and
170: February, during five seasons from 2001 through 2005. These are the
171: months in which the LMC is most visible from Cerro Tololo. We use the
172: $8K \times 8K$ Mosaic II CCD imager with a field of view (FOV) of 0.33 deg${}^2$,
173: along with a custom broadband \emph{VR} filter from 500nm to 750nm. For
174: this study, 52 epochs were chosen between 2001 November and 2005 December.
175: These were selected based on transparency, seeing conditions,
176: eccentricity, and their temporal spacing. Images from these $N = 52$ epochs were
177: first reduced by the SuperMACHO pipeline. The reduction
178: includes bias correction, cross-talk elimination, flat-fielding, image
179: deprojection, and DoPHOT photometry \citep{Schechter93} to compute
180: the zero-point magnitude.
181:
182: To isolate transient and variable objects such as light echoes, the SuperMACHO
183: pipeline includes difference image analysis. Point spread function (PSF) matched
184: subtraction is performed by the HOTPANTS software \citep{Becker04}. Since
185: the PSF can vary over the FOV due to optical distortions and
186: imperfect focus, a spatially varying kernel is employed \citep{Alard98,Alard00}.
187: For more information about the observation program, reduction, and difference
188: imaging, see \citet{Rest05:discrim}.
189:
190: \section{The NN2 Difference Imaging Technique}
191: \label{sec:analysis}
192: The principal feature of the NN2 method is that it
193: considers difference images created from all $N(N-1)/2$ possible pairs of
194: observations at $N$ epochs. For this study, the SuperMACHO pipeline was used to produce the 1326
195: difference images that served as the input to the NN2 algorithm.
196: From the beginning, therefore,
197: we consider images that ideally contain only light echo signal, and the
198: goal is to seperate the positive and negative light echo fluxes contained in
199: these images.
200: The NN2 algorithm is a single-pixel approach. For each pixel, a fit in
201: $N - 1$ free parameters is performed to determine the relative fluxes
202: across the $N$ input epochs that best reproduce the flux differences found
203: in the full set of difference images. We thus isolate the light echo signal in each
204: epoch, without the need for an echo-free template.
205: After obtaining the relative fluxes, a separate
206: method is used to establish the absolute flux level for each pixel.
207: Finally, spatial and temporal binning are performed to increase image
208: quality and depth.
209:
210: \subsection{The NN2 Algorithm}
211: \label{sec:nn2}
212: The NN2 algorithm was created by \citet{Barris05} to construct optimal
213: light curves for point-source variable objects from a series of
214: observations. To apply this technique to light echoes, we first
215: consider a single pixel. Let $A$ be the antisymmetric matrix for
216: which $A_{ij}$ is flux in the difference image between epochs $j$ and
217: $i$. This flux is normalized to zero-point magnitude 30, based on the
218: zero-point magnitudes of the images at epochs $j$ and $i$.
219: The matrix $A$, along with an error matrix $E$ whose entries contain
220: the uncertainties in the corresponding entries of $A$, represents the
221: input to the NN2 algorithm from the difference imaging pipeline. We
222: fit $A$ to the form $A_{ij} = V_j - V_i$ by minimizing
223: \begin{equation}
224: \label{tomin}
225: \chi^2 = \sum_{i < j} \frac{(-A_{ij} + V_j - V_i)^2}{E_{ij}^2} + \frac{\sum V_i}{\langle{E}\rangle^2},
226: \end{equation}
227: where $\langle{E}\rangle$ represents the average error, given by
228: \begin{equation}
229: \frac{1}{\langle{E}\rangle^2} = \frac{2}{N(N-1)} \sum_{i < j} \frac{1}{E_{ij}^2}.
230: \end{equation}
231: The vector $V$ then contains the relative fluxes across the $N$
232: epochs. The second term in Equation~(\ref{tomin}) is necessary to
233: obtain a unique solution for $V$, which is otherwise insensitive to
234: the addition of a constant vector. This term implies $\sum V_i = 0$
235: and requires us to determine the zero-flux level separately for each
236: pixel. As derived in \citet{Barris05}, $\chi^2$ is minimized by
237: solving a system of $N$ linear equations using matrix inversion.
238: %\begin{equation}
239: %\sum_{i; i \neq k} \frac{A_{ik}}{E_{ik}^2} = \sum_{i} C_{ik} V_{i},
240: %\end{equation}
241: %where
242: %\begin{equation}
243: %C_{ik} = -\frac{1}{E_{ik}^2} + \sum_{j} \frac{1}{E_{kj}^2} \delta_{ik} +\
244: %\frac{1}{\langle{E}\rangle^2}.
245: %\end{equation}
246:
247: There are two sources of uncertainty in $V$. The ``internal error''
248: is due to the fact that $A$ cannot in general be perfectly represented
249: as $A_{ij} = V_{j} - V_{i}$ for any $V$. An estimate for this error arises
250: naturally in the calculation of $V$, as described in \cite{Barris05}.
251: In addition, there are uncertainties in $V$
252: due to errors $E$ in the difference images, which we call the ``external
253: error.'' To estimate the external error in $V$, we assume that the uncertainty
254: $E_{ij}$ in the difference image arises from uncertainties $\sigma_i$ and
255: $\sigma_j$ in the original images at epochs $i$ and $j$, i.e., that
256: there exists a vector $\sigma$ such that
257: \begin{equation}
258: E_{ij}^2 = \sigma_i^2 + \sigma_j^2.
259: \end{equation}
260: To calculate the best-fit $\sigma$, we minimize
261: \begin{equation}
262: \label{chiemin}
263: \chi_e^2 = \sum_{i,j; i < j} \left(-1 + \frac{\sigma_i^2}{E_{ij}^2} + \frac{\sigma_j^2}{E_{ij}^2}\right)^2.
264: \end{equation}
265: The minimization condition
266: \begin{equation}
267: \frac{\partial \chi_e^2}{\partial \sigma_k^2} = 0
268: \end{equation}
269: results in a system of $N$ equations, which can be written as
270: \begin{equation}
271: \sum_{i; i \neq k} \frac{1}{E_{ik}^2} = \sum_{i} D_{ik} \sigma_i^2,
272: \end{equation}
273: where
274: \begin{eqnarray}
275: \displaystyle
276: D_{ik} &=& \frac{1}{E_{ik}^4}, \quad i \neq k \nonumber \\
277: D_{kk} &=& \sum_{i; i \neq k} \frac{1}{E_{ik}^4},
278: \end{eqnarray}
279: and solved by inverting $D$.
280:
281: In rare cases, this method yields negative solutions for $\sigma_i^2$,
282: indicating that the original assumption that $E_{ij}^2$ can be
283: well-represented by $\sigma_i^2 + \sigma_j^2$ is not accurate. In
284: practice, this occurs in of order 1 in $10^5$ pixels, which are
285: therefore simply masked. The $\chi_e^2$ in equation~(\ref{chiemin}) is that
286: adopted in the NN2 code of M.~W.~Wood-Vasey, J.~Tonry, and
287: M.~C.~Novicki.\footnote{\texttt{See \url{http://ctiokw.ctio.noao.edu:8080/Plone/essence/Software/NN2}.}}
288: It differs slightly from the $\chi_e^2$ given in \cite{Barris05}.
289: %\begin{equation}
290: %\label{chiemin_paper}
291: %\chi_e^2 = \sum_{i, j; i < j} (-E_{ij}^2 + \sigma_i^2 + \sigma_j^2)^2.
292: %\end{equation}
293: %The difference is that Equation~\ref{chiemin} measures
294: %the fractional error in the representation of each $E_{ij}$ by
295: %$\sigma$, whereas Equation~\ref{chiemin_paper} measures the
296: %absolute error.
297: In practice, the errors obtained are
298: similar when both methods are available, but minimizing the
299: $\chi_e^2$ of \citet{Barris05} fails far more often (i.e., yields negative
300: values for $\sigma_i^2$).
301:
302: To compute the total uncertainty, the contributions of both the internal and
303: external errors must be considered. To combine these errors, we
304: use the weighted method presented in \citet{Barris05}, in which the
305: internal error is weighted by the normalized $\chi^2$ of equation~(\ref{tomin}).
306: %The internal error is scaled as follows:
307: %\begin{equation}
308: %\tau_k = \left(C_{kk}^{-1} \frac{\chi^2}{N_{\textrm{\scriptsize{dof}}}}\right)^{1/2},
309: %\end{equation}
310: %where
311: %\begin{equation}
312: %N_{\textrm{\scriptsize{dof}}} = \frac{N(N-1)}{2} - (N-1)
313: %\end{equation}
314: %is the number of degrees of freedom. The total error is the quadrature
315: %sum of $\tau$ and $\sigma$. When $A$ is well-represented as a vector
316: %difference, $\chi^2/N_{\textrm{\scriptsize{dof}}} \ll 1$, and the
317: %total error is due mainly to the external error. If, on the other
318: %hand, $\chi^2/N_{\textrm{\scriptsize{dof}}} \gg 1$, this factor will
319: %correct $\tau$ and provide a reasonable total error.%, even if the
320: %external error is very small.
321:
322: For the heart of the NN2 calculation, the \texttt{antivec} function,
323: we adapted the code of W.~M.~Wood-Vasey, J.~Tonry, and M.~C.~Novicki. A region
324: of specified size is read into memory from each of the $N(N-1)/2$
325: difference images. For each pixel, masked data are discarded as
326: follows. The number of images associated with each epoch (either as
327: the image or template date) in which the pixel is masked is
328: calculated. All images associated with the epoch having the highest
329: count are discarded, and the process is repeated until no masked data
330: remains. We consider input pixels as masked if they are marked by the
331: SuperMACHO pipeline as saturated or bad pixels.
332: The NN2 algorithm is run with the
333: remaining data, and a crude zero-flux level is set by shifting the
334: relative fluxes so that the minimum flux for each pixel is zero. Image files containing
335: the flux, internal error, external error, and total error are
336: created. In addition, mask files are created for each epoch showing
337: which pixels were discarded. Pixels that encountered an error in
338: calculating the external error are also flagged. From the mask files, an
339: image is created, indicating the total number of epochs included in
340: the calculation for each pixel. Pixels calculated with fewer than 30
341: epochs were masked in our study of SN 1987A.
342:
343: \subsection{Zero-Flux--Level Determination}
344: \label{sec:zeroflux}
345: Since the NN2 algorithm provides only relative fluxes, the light curves
346: obtained for each pixel contain arbitrary offsets with respect to both
347: the true zero-flux level and the neighboring pixels. It is therefore
348: necessary to estimate the true zero-flux level for each pixel.
349: Figure~\ref{fig:lightcurve_weakecho}\emph{a} shows the relative
350: light curve of an example pixel as determined by the NN2 algorithm. As
351: a first crude estimate, the zero-flux level is taken to be the minimum
352: flux over all epochs. This value is subtracted from each epoch to
353: produce Figure~\ref{fig:lightcurve_weakecho}\emph{a}. Since our observations were taken in annual
354: spans of approximately 3 months for 5 years, we group the data
355: into five seasons. The example in Figure~\ref{fig:lightcurve_weakecho}
356: shows a very faint light echo in the third season. In order to
357: estimate the zero-flux level, we must identify the time periods (in
358: our case, seasons) that contain no light echo flux. For faint light
359: echoes, as in this example, this is very difficult, because the
360: individual pixel values are sometimes barely 1 $\sigma$ above the sky
361: background. To assist in
362: identifying the zero-flux epochs, we use the following properties of
363: light echoes:
364: \begin {itemize}
365: \item Light echoes are spatially extended, so nearby pixels are very likely
366: to share the same epochs with zero light echo flux. Thus, we can
367: spatially bin the fluxes by replacing each pixel with the weighted
368: mean of its unmasked neighbors within a $3 \times 3$ box, excluding
369: the highest and lowest fluxes. (Throughout this paper, extreme fluxes
370: are excluded from means only when at least six data points are available.)
371: The resulting spatially-smoothed light curve is shown in
372: Figure~\ref{fig:lightcurve_weakecho}\emph{b}.
373: \item In most cases, light echoes do not change significantly over a time scale
374: of months. Thus we can temporally bin the fluxes as well. In our case,
375: the five seasons form the natural binning periods. Within each season,
376: the epochs of highest and lowest flux are excluded. These are denoted by
377: crosses in Figure~\ref{fig:lightcurve_weakecho}\emph{b}.
378: Using the remaining data, the weighted mean flux of each season is calculated.
379: These are shown as squares in Figure~\ref{fig:lightcurve_weakecho}.
380: In this example,
381: the mean of the spatially-binned third season is $6.4$ $\sigma$ above the mean
382: of the other four seasons. Without spatial binning, this is reduced to
383: $2.6$ $\sigma$.
384: \end{itemize}
385:
386: Using the spatially- and temporally-binned light curve, we use the
387: following algorithm to determine the seasons that contain zero light
388: echo flux:
389: \begin{itemize}
390: \item Seasons whose mean is more than 3 $\sigma$ above the minimum mean among
391: all seasons are considered to contain light echoes and are thus
392: excluded. In practice, most seasons that will be cut are excluded by
393: this sigma cut.
394: \item Seasons that exhibit a strong gradient are excluded. The presence
395: of a gradient is detected by measuring how well the best-fit line fits
396: the data compared to a flat line at the mean. In particular, the ratio
397: of the $\chi^2$ statistic for the mean to
398: the $\chi^2$ statistic for the best-fit line, as
399: determined by the least-squares method,
400: is calculated. Seasons for which this ratio exceeds 10 are cut.
401: \item Seasons with a normalized $\chi^2$ greater than 3.5 are excluded.
402: \end{itemize}
403: The remaining seasons are deemed to contain zero light echo flux. In
404: the example of Figure~\ref{fig:lightcurve_weakecho}, these are seasons
405: 1, 2, 4, and 5. (Season 3 is excluded by the sigma cut.) Since at least one
406: season must be used to calculate the zero-flux level, none
407: of the above steps is allowed
408: to exclude all the remaining seasons. If one initially does so, the
409: parameter being measured is examined, and the season that deviates
410: least from a perfect zero-flux season is retained. The critical values of
411: the parameters in the second and third steps were chosen by sampling pixels
412: from regions containing passing light echoes and choosing values to best
413: reproduce the zero-flux level seen in the light curve. They can be adjusted
414: for different data sets; however, since the second and third steps are included
415: as safeguards after the dominant sigma cut, the exact critical values are not
416: likely to significantly affect the output.
417:
418: Having determined the zero-flux seasons, we use this information to
419: correct the original, unbinned light curve, shown in
420: Figure~\ref{fig:lightcurve_weakecho}\emph{a}. For each zero-flux season,
421: the highest and lowest fluxes are cut to reduce the influence of outlier data.
422: (We deemed this cut sufficient for our data. Those planning
423: future applications may want to consider an iterative sigma clip instead.)
424: %These are shown by square symbols
425: %in Figure~\ref{fig:lightcurve_weakecho}(c).
426: The weighted mean of all zero-flux seasons is taken to be the
427: zero-flux level, shown by the dotted line in Figure~
428: \ref{fig:lightcurve_weakecho}\emph{a}. This value is subtracted from the flux at every
429: epoch, and its uncertainty is combined quadratically with errors in
430: the fluxes, to produce the zero-flux corrected light curve, as shown in
431: Figure~\ref{fig:lightcurve_weakecho}\emph{c}.
432:
433: The results of this method, applied not only to a single pixel but to
434: an entire region, are shown in Figure~\ref{fig:zerofluxlevel}. Panel \emph{a}
435: shows a difference image, panel \emph{b} shows the NN2
436: output for the image epoch of the difference image, and panel \emph{c}
437: shows the zero-flux--corrected image for the same epoch. Note
438: how much more clearly the light echoes appear above the sky background
439: in panel \emph{c}, especially the fainter features.
440:
441: The procedure described here is dependent on the availability of true
442: zero-flux epochs in the neighborhood of every pixel. For some pixels,
443: only one season is identified as containing zero light echo flux. For
444: them, it is possible that all seasons actually contain light echoes,
445: and that the identified season really contains the faintest light
446: echo region, rather than zero flux. In this case, the calculated
447: zero-flux level is an overestimate of the true value, and the fluxes
448: obtained are lower limits on the true fluxes. In our example of SN
449: 1987A, the rings move quickly enough and are thin enough that few
450: pixels should contain light echoes for 5 consecutive
451: years. However, there are some regions that contain several closely
452: spaced light echo arclets. An extreme case is shown in
453: Figure~\ref{fig:onebin}, in which the final frame indicates those
454: pixels with only one zero-flux season identified. The red lines are constant
455: and indicate regions that are dense in such pixels. It is possible that
456: some of these pixels are illuminated by light echoes throughout the
457: observation period. The season during which the faintest echoes pass
458: is used to calculate the zero-flux level, and consequently the light
459: echoes disappear from the zero-flux--corrected images for that
460: season. This effect must be accounted for, especially if the technique
461: presented here is applied to light echoes with more complicated
462: structure.
463:
464: In our images of SN 1987A, 7\% of all usable pixels have only one
465: zero-flux season identified. In all cases we have examined, this
466: season does seem to represent the true zero-flux level. Sometimes
467: this can be verified from the
468: light curve. Figure~\ref{fig:lightcurve_strongecho} shows the zero-flux--corrected
469: light curve for a pixel containing a very bright light echo.
470: Despite being the only zero-flux season identified, the first season
471: does appear to contain zero light echo
472: flux. Pixels with only one zero-flux season can also be inspected by
473: following the motion of light echo arcs. If the zero-flux season falls
474: in a gap between arcs
475: that appears in images from the other seasons, we can be
476: confident in its veracity. If, on the other hand, a conspicuous hole
477: were to appear in the zero-flux season, altering the shape of the
478: light echo, the calculated fluxes would be interpreted as lower limits.
479:
480: \subsection{Stacking and Smoothing}
481: \label{sec:stacksmooth}
482: To assist in separating very faint light echo pixels from the sky background, the
483: NN2 images from each season were mean-combined (stacked), excluding masked
484: epochs as well as those containing the highest and lowest fluxes,
485: resulting in five stacked images. Corresponding error and mask images
486: were also created. Since the flux can change by a factor of 2 or more
487: within a single season in a region with a moving light echo, this
488: binning reduces the detail contained in the flux information. However, it
489: does increase the signal-to-noise ratio and improve the visibility
490: of faint rings for detection. The number of bad pixels with extreme
491: values is also reduced. These can arise, for example, from difference
492: imaging artifacts in the original difference images. To further improve
493: visibility, the five stacked images were smoothed using the same spatial
494: binning technique described in section
495: \ref{sec:zeroflux}. The results are shown in
496: Figure~\ref{fig:weakringseries}. Panels \emph{a} and \emph{b} show a sample difference
497: image and the zero-flux--corrected output for its template epoch,
498: respectively. Panel \emph{c} shows the season mean-combined
499: image, formed by stacking many images like the one in \emph{b},
500: and \emph{d} shows the same image after spatial smoothing.
501: The stacking and smoothing procedure
502: typically increased the signal-to-noise ratio by a factor of 10 in our
503: images of SN 1987A. This could be anticipated, since stacking should
504: reduce the noise by a factor of approximately $(N/n)^{1/2}$, where $n$
505: is the number of seasons, and smoothing should further reduce the
506: noise by a factor of $\sqrt{s}$, where $s$ is the number of pixels
507: used in the smoothing kernel. In this case, $N = 52$, $n = 5$, and $s
508: = 9$, so $(Ns/n)^{1/2} \approx 10$.
509:
510: \section{Discussion}
511: \label{sec:discussion}
512:
513: Figure~\ref{fig:lightechotimeseries} shows the evolution of one light
514: echo region over our five seasons of observation. The left column shows difference
515: images with the same template epoch in 2001 November. The image epochs are
516: separated by 1 year and fall in December 2001, 2002, 2003, 2004 and
517: 2005. The right column shows the zero-flux--corrected NN2 images for the
518: same five epochs. Note that in the first difference image, the
519: template and image dates are separated by less than 1 month. The
520: light echoes move very little during this short interval, so the
521: difference image shows almost complete cancellation between the light
522: echoes at the two epochs. However, the accompanying zero-flux--corrected NN2
523: image shows the full extent of the light echo. The second
524: and subsequent difference images show some regions that appear to
525: contain flux only from the image epoch, and others that appear to
526: contain flux only from the template epoch. However, it is difficult
527: to make these judgments with certainty, and there are always
528: significant regions of overlap, in which the difference image reveals
529: only the change in flux. The NN2 algorithm allows us to disentangle
530: this information and extract the relative light echo fluxes across a
531: range of epochs rather than only two. It is important to note that
532: the images in the right column of Figure~\ref{fig:lightechotimeseries}
533: are not obtained directly from the accompanying difference images, but
534: rather depend on all $N(N-1)/2$ difference images obtained from the
535: $N$ observations. This figure clearly shows the limitations of the
536: single-template difference imaging method and the advantages offered
537: by the NN2 algorithm.
538:
539: In addition to uncovering the spatial extent of light echoes, our technique
540: can reveal faint light echo features not visible in the difference
541: images. Figure~\ref{fig:weakringseries} shows stages of analysis for a
542: region containing a faint light echo ring in the left part of the
543: frame. Note that this ring is barely tracable in the
544: sample difference image. In our full mean-combined images, light echo
545: features with peak surface brightnesses of 25 mag arcsec${}^{-2}$
546: are clearly visible, whereas the light echo detection limit in the
547: difference images is about 24 mag arcsec${}^{-2}$.
548: %The brightest
549: %knots have a surface brightness of 19.6 mag arcsec${}^{-2}$, in excellent agreement with
550: %what we expect based on difference images.
551:
552: Our code, along with images from an example region around SN 1987A,
553: are available online.\footnote{See \texttt{\url{http://www.ctio.noao.edu/supermacho/NN2}}.}
554: The NN2 calculations involve several large matrix inversions and are CPU-intensive.
555: Our Python code required roughly 30 hours on 20 CPUs to process 52 images of
556: dimensions $4160 \times 1100$. Although vast speed improvements can likely be
557: expected from C code, the processing demands should be borne in mind when
558: considering larger scale applications.
559: %We developed scripts to split the calculations among several
560: %processes which could be run on a 20 CPU cluster. The output
561: %from these processes was then reassembled into the final images. For
562: %52 images of dimensions $4160 \times 1100$, this procedure still
563: %required about 30 hours using our Python code.
564: %Speed improvements by a factor of 10-100 can likely be expected from C code.
565:
566: \section{Conclusions}
567: \label{sec:conclusions}
568: Difference imaging is a standard tool in the analysis of light
569: echoes. Without a light echo-free template, however, difference images
570: contain overlapping light echo fluxes from two epochs. We present here a new method that
571: produces images containing absolute light echo fluxes from one epoch
572: only. The NN2 algorithm is used to calculate a relative light curve
573: across a range of $N$ epochs by considering difference images
574: constructed from every pair of epochs. By applying this method over a
575: region, we obtain light curves for each pixel. However, these lightcurves
576: have arbitrary offsets with respect to each other and to the
577: true zero-flux level. Hence, we apply a statistical method to estimate
578: the true zero-flux level for each pixel and shift the light curves
579: accordingly. This method is optimized for detecting faint light echoes
580: with peak surface brightnesses as faint as 25 mag arcsec${}^{-2}$.
581: By spatially and temporally binning the images, we greatly increase
582: the signal-to-noise ratio. Our technique is capable of revealing fine
583: structure and faint light echo regions. It requires no privileged or
584: light echo-free images and can be applied completely automatically to
585: a wide variety of light echo situations. In addition, variations on this
586: technique could permit analysis of other extended variable light sources,
587: such as stellar outflows and supernova remnants.
588: Future work will include applications to recently discovered light echoes from
589: ancient supernovae in the Large Magellanic Cloud \citep{Rest05:le}.
590:
591: \section{Acknowledgments}
592:
593: The SuperMACHO survey is being undertaken under the auspices of the
594: NOAO Survey Program. We are very grateful for the support provided to
595: the Survey program from the NOAO and the National Science
596: Foundation. A.~Rest thanks NOAO for the Goldberg
597: Fellowship. SuperMACHO is supported by the STScI grant GO-10583 and
598: GO-10903. Discussions with Michael Wood-Vasey were very
599: valuable. A.~Newman thanks the National Science Foundation for their
600: support via the Research Experience for Undergraduates (REU) program.
601: We thank the referee B.~Sugerman for very useful comments and suggestions
602: which helped improve the quality of this paper.
603:
604: \bibliographystyle{apj}
605: \bibliography{ms}
606: \clearpage
607: %\begin{thebibliography}{}
608: %
609: %\bibitem[\protect\citeauthoryear{{Alard}}{{Alard}}{2000}]{Alard00}
610: %{Alard}, C. 2000, \aaps, 144, 363
611: %
612: %\bibitem[\protect\citeauthoryear{{Barris} et~al.}{{Barris}
613: % et~al.}{2005}]{Barris05}
614: %{Barris}, B.~J., {Tonry}, J.~L., {Novicki}, M.~C., \& {Wood-Vasey}, W.~M.
615: % 2005, \aj, 130, 2272
616: %
617: %\bibitem[\protect\citeauthoryear{{Chevalier}}{{Chevalier}}{1986}]{Chevalier86}
618: %{Chevalier}, R.~A. 1986, \apj, 308, 225
619: %
620: %\bibitem[\protect\citeauthoryear{{Couderc}}{{Couderc}}{1939}]{Couderc39}
621: %{Couderc}, P. 1939, Annales d'Astrophysique, 2, 271
622: %
623: %\bibitem[\protect\citeauthoryear{{Crotts}, {Kunkel}, \& {Heathcote}}{{Crotts}
624: % et~al.}{1995}]{Crotts95}
625: %{Crotts}, A.~P.~S., {Kunkel}, W.~E., \& {Heathcote}, S.~R. 1995, \apj, 438,
626: % 724
627: %
628: %\bibitem[\protect\citeauthoryear{{Rest} et~al.}{{Rest}
629: % et~al.}{2005a}]{Rest05:discrim}
630: %{Rest}, A., et~al. 2005a, \apj, 634, 1103
631: %
632: %\bibitem[\protect\citeauthoryear{{Rest} et~al.}{{Rest}
633: % et~al.}{2005b}]{Rest05:le}
634: %{Rest}, A., et~al. 2005b, \nat, 438, 1132
635: %
636: %\bibitem[\protect\citeauthoryear{{Sugerman}}{{Sugerman}}{2003}]{Sugerman03}
637: %{Sugerman}, B.~E.~K. 2003, \aj, 126, 1939
638: %
639: %\bibitem[\protect\citeauthoryear{{Sugerman} et~al.}{{Sugerman}
640: % et~al.}{2005}]{Sugerman05}
641: %{Sugerman}, B.~E.~K., {Crotts}, A.~P.~S., {Kunkel}, W.~E., {Heathcote}, S.~R.,
642: % \& {Lawrence}, S.~S. 2005, \apj, 627, 888
643: %
644: %\bibitem[\protect\citeauthoryear{{Tylenda}}{{Tylenda}}{2004}]{Tylenda04}
645: %{Tylenda}, R. 2004, \aap, 414, 223
646: %
647: %\bibitem[\protect\citeauthoryear{{Xu}, {Crotts}, \& {Kunkel}}{{Xu}
648: % et~al.}{1995}]{Xu95}
649: %{Xu}, J., {Crotts}, A.~P.~S., \& {Kunkel}, W.~E. 1995, \apj, 451, 806
650: %
651: %\end{thebibliography}
652:
653: \clearpage
654:
655: \begin{figure}
656: \begin{center}
657: \includegraphics[scale=0.75]{f1}
658: \end{center}
659: \caption{Sample difference image showing 2005 December 31 (image epoch)
660: minus 2005 January 7 (template epoch). White pixels are positive and thus
661: enriched in the image epoch, while black pixels are negative
662: and enriched in the template epoch. Note that this image contains light
663: echo fluxes from both epochs which frequently overlap and exhibit
664: partial cancellation. The FOV is $3.1^{\prime} \times 2.3^{\prime}$.
665: \label{fig:diffim_example}}
666: \end{figure}
667:
668: \clearpage
669:
670: \begin{figure}
671: \begin{center}
672: \begin{tabular}{cc}
673: \includegraphics[scale=0.65]{f2a.eps} \\
674: \includegraphics[scale=0.65]{f2b.eps} \\
675: \includegraphics[scale=0.65]{f2c.eps}
676: \end{tabular}
677: \end{center}
678: \caption{Light curves for a pixel containing a faint light echo in season 3.
679: (\emph{a}) Relative fluxes obtained from the NN2 algorithm. (\emph{b}) Smoothed fluxes
680: obtained by spatially binning pixels in a $3 \times 3$ box. The crosses
681: denote the highest and lowest flux values from each season,
682: which are excluded in calculating the season means, denoted by squares.
683: A light echo is identified in season 3 by the sigma cut, and the remaining
684: four seasons are deemed to represent the zero-flux level. The dotted
685: line shows their mean.
686: (\emph{c}) Zero-flux--corrected light curve. The mean of the unbinned fluxes in the
687: zero-flux seasons is set as the zero-flux
688: level, the dotted line in (\emph{a}). This is subtracted from each epoch
689: to produce the corrected light curve here. Squares and crosses have the
690: same meaning as above.
691: %The surface brightness in season 3 is $24.9 \pm 2.1$ mag arcsec${}^{-2}$.
692: Error bars show the 1 $\sigma$ uncertainties.
693: \label{fig:lightcurve_weakecho}}
694: \end{figure}
695:
696: \clearpage
697:
698: \begin{figure}
699: \centering
700: \begin{tabular}{c}
701: \includegraphics[scale=0.75]{f3a.eps} \\
702: \includegraphics[scale=0.75]{f3b.eps} \\
703: \includegraphics[scale=0.75]{f3c.eps}
704: \end{tabular}
705: % ds9 workspace/sm041213_sm011211/15/sm77.041213_0622.106_15.diff.fits nn2_output/nn2.sn87a4/sum/sn87a4.041213_0622.15.nn2.fits nn2_output/nn2.sn87a4/sum/zpt_corrected/sn87a4.041213_0622.15.nn2z.fits
706: % size 656 x 453 = 3.0' x 2.0'
707: \caption{(\emph{a}) Difference image with image and template epochs
708: of 2004 December 13 and 2001 December 11, respectively.
709: (\emph{b}) NN2 output for the image epoch of the difference image. For each pixel,
710: the relative light curve obtained from the NN2 algorithm is shifted
711: so that the minimum flux over all epochs is zero. Note that light echo
712: flux from the template epoch in (\emph{a}) has been eliminated.
713: (\emph{c}) Zero-flux--corrected image for the same epoch. Each pixel level has been
714: shifted to the proper flux zero point.
715: The FOV is $3.0^{\prime} \times 2.0^{\prime}$.
716: \label{fig:zerofluxlevel}}
717: \end{figure}
718:
719: \clearpage
720:
721: \begin{figure}
722: \centering
723: \begin{tabular}{cc}
724: \includegraphics[scale=0.75]{f4a.eps} &
725: \includegraphics[scale=0.75]{f4b.eps} \\
726: \includegraphics[scale=0.75]{f4c.eps} &
727: \includegraphics[scale=0.75]{f4d.eps} \\
728: \includegraphics[scale=0.75]{f4e.eps} &
729: \includegraphics[scale=0.75]{f4f.eps}
730: \end{tabular}
731: % Dimensions 780 x 492 = 3.5' x 2.2'
732: \caption{Mean-combined images for the 2001--2005 seasons
733: showing a region with closely spaced light echo arcs. The final frame
734: shows those pixels for which only one zero-flux season is identified. The close
735: spacing of the light echo arcs in this region causes an unusually high number
736: of pixels to contain light echoes in all but one season.
737: For these pixels, the flux values obtained here may be only
738: lower limits on the true flux. The red lines are constant
739: in each image and highlight areas dense in pixels with one zero-flux season.
740: Note that every such pixel necessarily has a calculated flux of zero in
741: one season, even if light echoes were physically present throughout the
742: observation period. By following the motion of the light echoes, however,
743: we can generally exclude this possibility. The FOV is $3.5^{\prime} \times 2.2^{\prime}$.
744: \label{fig:onebin}}
745: \end{figure}
746:
747: \clearpage
748:
749: \begin{figure}
750: \centering
751: \includegraphics[scale=0.7]{f5.eps}
752: \caption{Zero-flux--corrected light curve for a pixel containing a bright light echo.
753: Seasons 2 through 5 are excluded by the sigma cut, and the zero-flux level is determined
754: by season 1 alone. Nevertheless, given the smooth nature of the curve and the flatness of the
755: season 1 data, it is very likely that season 1 represents the true zero-flux level.
756: Uncertainties are comparable to the size of the points.
757: \label{fig:lightcurve_strongecho}}
758: \end{figure}
759:
760: \clearpage
761:
762: \begin{figure}
763: \centering
764: \begin{tabular}{cc}
765: \includegraphics[scale=0.75]{f6a.eps} &
766: %\includegraphics[scale=0.25,angle=0]{fig.weakseries/diff041215_011221_-10_10_a}
767: \includegraphics[scale=0.75]{f6b.eps} \\
768: %\includegraphics[scale=0.25,angle=0]{fig.weakseries/zflc011221_-4_5_b} \\
769: \includegraphics[scale=0.75]{f6c.eps} &
770: \includegraphics[scale=0.75]{f6d.eps}
771: \end{tabular}
772: % ds9 -nomagnifier -nocolorbar ../../../workspace/sm041215_sm011221/15/sm77.041215_0438.096_15.diff.fits zpt_corrected/sn87a4.011221_0353.15.nn2z.fits zpt_corrected/stacked.15/season1.15.stacked.fits zpt_corrected/stacked.15/season1.15.stacked.smoothed.fits
773: % Size 655 x 655 = 2.9' x 2.9'
774: \caption{Stages of light echo extraction
775: showing a region containing a weak light
776: echo arc in the left part of the frame. The FOV is $2.9^{\prime} \times 2.9^{\prime}$.
777: (\emph{a}) Difference image showing 2004 December 15 minus 2001 December 21.
778: (\emph{b}) Zero-flux corrected image for 2001 December 21. This is the template epoch of the difference image, in which it appears black.
779: (\emph{c}) Mean-combined image for the 2001 season.
780: (\emph{d}) Smoothed mean-combined image for the same season.
781: \label{fig:weakringseries}}
782: \end{figure}
783:
784: \clearpage
785:
786: \begin{figure}
787: \centering
788: \begin{tabular}{cc}
789: \includegraphics[scale=0.75]{f7a.eps}
790: & \includegraphics[scale=0.75]{f7b.eps} \\
791: \includegraphics[scale=0.75]{f7c.eps}
792: & \includegraphics[scale=0.75]{f7d.eps} \\
793: \includegraphics[scale=0.75]{f7e.eps}
794: & \includegraphics[scale=0.75]{f7f.eps} \\
795: \includegraphics[scale=0.75]{f7g.eps}
796: & \includegraphics[scale=0.75]{f7h.eps} \\
797: \includegraphics[scale=0.75]{f7i.eps}
798: & \includegraphics[scale=0.75]{f7j.eps}
799: \end{tabular}
800: \caption{
801: \emph{Left column:} Difference images with template epoch 2001 November 15 and image epochs
802: 2001 December 11, 2002 December 14, 2003 December 15, 2004 December 15,
803: and 2005 December 21, respectively,
804: from top to bottom. \emph{Right column:} Zero-flux--corrected images for the
805: same five image epochs. The red arcs provide a constant point of reference.
806: In each difference image, the light echo positions
807: in the image and template epochs overlap, causing partial or complete cancellation.
808: The NN2 algorithm disentangles this information to reveal the complete extent of the light
809: echoes. Note especially in the first image the near complete cancellation evident
810: in the difference image, in which the image and template dates are separated by
811: only 1 month. Compare this to the image on the right. The FOV is
812: $3.0^{\prime} \times 2.0^{\prime}$.\label{fig:lightechotimeseries}}
813: \end{figure}
814:
815: \end{document}
816: