CSIRO / Swinburne ; NASA

Wide-Field Radio Astronomy and the Dynamic Universe Bryan Gaensler / @SciBry! Dunlap Institute! !

with Cleo Loi, Kitty Lo,! Martin Bell, Keith Bannister,! Paul Hancock, Tara Murphy! and the MWA Collaboration! REPORTS

signal-to-noise ratios (SNRs) to yield astrophysi- probably requires the source to be in a high- close to that of FRB 010724, the FRBs presented cally interesting constraints for either parameter scattering region, for example, a galactic center. here have much higher and randomly distributed and show no evidence of scattering. The second possibility is scattering because DMs. Three of these FRBs are factors of >3 Our FRBs were detected with DMs in the of turbulence in the ionized IGM, unassociated narrower than any documented peryton. Last, the range from 553 to 1103 cm−3 pc. Their high Galactic with any . There is a weakly constrained characteristic scattering shape and strong disper- latitudes ( b > 41○,Table1)correspondtolines empirical relationship between DM and measured sion delay adherence of FRB 110220 make a of sight throughj j the low column density Galactic scattering for in the MW. If applicable to case for cold plasma propagation. ISM corresponding to just 3 to 6% of the DM the IGM, then the observed scattering implies The Sun is known to emit frequency-swept −3 measured (10). These small Galactic DM con- DMIGM > 100cm pc (2, 15). With use of the radio bursts at 1 to 3 GHz [typeIIIdm (17)]. These The Dynamictributions areRadio highly supportive of an extragalac- Skyaforementioned model of the ionized IGM, this bursts have typical widths of 0.2 to 10 s and tic origin and are substantially smaller fractions DM6 equates to z > 0:11 (2, 12, 13). The prob- positive frequency sweeps, entirely inconsistent Murphy et al. than those of previously reported bursts, which ability of an intervening galaxy located along the with measurements of W and a for the FRBs. et al, 1987) after et Fiedler al. 2013, (Murphy › Variability ↔ unique view of extreme −physics3 were 15% of DM = 375 cm pc for FRB 010724 line of sight within z ≈ 1is≤0.05 (16). Such a Whereas FRB 110220 was separated from the responsible for ESEs via scattering, absorption and refrac- −3 (4)and70%ofDM=746cm pc for FRB galaxy could be a sourceQ0954+658 of scattering andat dis-2.7 &Sun 8.1 by 5.6°,GHz FRB 110703 was detected at night tion (Draine 1998). Additionally, H i observations would › Variability at radio wavelengths010621 (5). persion, but the magnitude would be subject to and the others so far from the Sun that any event scattering Extreme reveal any associated neutral structures, Faraday rotation - no extinction, extremeThe sensitivity non-Galactic DM contribution,& resolution DME,is the same inclination dependence as described for solar radiation should have appeared in multi- measurements may constrain the magnetic field strength the sum of two components: the intergalactic asourcelocatedinthediskofaspiralgalaxy. ple beams. These FRBs were only detected in a within the structures, and high-resolution observations could 2 ) - small fields of view;medium e.g. (IGM; VLA DM IGM= )andapossiblehostgal-0.2 deg It is important to be sure that FRBs are not a single beam; it is therefore unlikely they are of

Jy confirm predicted changes in the apparent structure of the axy (DM ). The intervening medium could be terrestrial source of interference. Observations at solar origin. Host source. While there have been previous efforts in conduct- (relatively purelyrare; intergalactic Metzger and could et also al. include 2015) a con- Parkes have previously shown swept frequency Uncertainty in the true position of the FRBs › Explosions tribution from an intervening galaxy. Two op- pulses of terrestrialFlux( origin, dubbed “perytons.” within the frequency-dependent gain pattern of ing such observations (Clegg et al. 1996; Lazio et al. 2000, - supernovae, GRBs,tions orphan are considered afterglows according to the proximity These are symmetric W >20mspulses,which the telescope makes inferring a spectral index, and 2001a), only one source has been observed in such a manner (B1741 038), and, for some of the measurements, only a sin- - mass-loss history, beamingof the source to the center of a host galaxy. imperfectly mimic a dispersive sweep (2, 8). Al- hence flux densities outside the observing band, − If located at the center of a galaxy, this may be though perytons peak in apparent DM near difficult. A likely off-axis position changes the in- gle observation during the actual ESE could be obtained. An ahighlydispersiveregion;forexample,lines 375 cm−3 pc (range from ~200 to 420 cm–3 pc), trinsic spectral index substantially. The spectral all-sky survey could detect multiple ESEs during the course › Propagation (relativelyof sightcommon; passing through Lo the centralet al. regions 2013) of of a year, allowing ample opportunity to trigger follow-up –like could lead to DMs in Time (days) observations with other telescopes and at other wavelengths. - scintillation, extreme scattering−3 events on September 2, 2013 excess of 700 cm pc in the central ~100 pc (11), Figure 1. An extreme scattering event in Q0954 658 at 2.7 GHz (lower) + - probes of turbulence;independent baryons of the line-of-sight in the inclination.IGM In and 8.1 GHz (upper) adapted from Fiedler et al. (1987b). For clarity, an offset 2.2.3 Interstellar Scintillation of AGNs this case, DME is dominated by DMHost and re- of 1 Jy has been added to the top trace. The strong frequency dependence Observations of decimetre variability in extragalactic sources quires FRBs to be emitted by an unknown mecha- of ESEs and the necessity of regular sampling of the light curves over a (e.g. Hunstead 1972) led quickly to the recognition that › Accretion long period is evident. At ASKAP frequencies ( 1 GHz), the amplitude of nism in the central region, possibly associated ∼ the sources needed to have extremely compact components, the flux density decrease would likely have been even larger, and the flux et al. 2013) (Thornton

- X-ray binaries, AGN,with quasars the supermassive black hole located there. Fast radio bursts density would likely have increased to an even higher value during the start be affected by interstellar scintillation, or both (Armstrong,

If outside a central region, then elliptical host ) - disk/jet connection of the event. Spangler, & Hardee 1977; Spangler et al. 1989). The need galaxies (which are expected to have a low electron Jy for extremely compact components, apparently violating the density) will not contribute to DME substantially,

so-calledwww.sciencemag.org inverse Compton catastrophe, became even more › Magnetospheres and DMHost for a spiral galaxy will only contrib- Awide-areasurveywithdailycadencewouldallowusto( Flux acute with the discovery of ‘flickering’ (Heeschen 1984), and ute substantially to DM if viewed close to edge- - pulsars, , flare stars,E exoplanets address the physical composition and origin of the objects later intra-day variability. on [inclination, i > 87○ for DM > 700cm−3pc; - reconnection, particleprobability accelerationi > 87○ ≈ 0:05]. The chance of all responsible for ESEs in a comprehensive manner. The MASIV 5-GHz Very Large Array (VLA) survey ð Þ (Lovell et al. 2003, 2008) and other observations (Kedziora- four FRBs coming from edge-on spiral galaxies Population statistics: One of the more comprehensive ESE −6 Chudczer et al. 1997; Dennett-Thorpe & de Bruyn 2002; is therefore negligible (10 ). Consequently, if the surveys (Fiedler et al. 1994b) provided a total coverage of › The Unknown Bignall et al. 2003, 2006) have established that the fast— sources are not located in a galactic center, DMHost approximately 600 source-years; VAST has the potential to - fast radio bursts (1would every likely be 10 small, seconds!) and DMIGM dominates. intra-day and inter-day—flux density variations at centime- achieve a coverage that is larger by a factor of roughly 300, Downloaded from Assuming an IGM free-electron distribution, which tre wavelengths exhibited by a subset of compact quasars are searching for ESEs towardsTime sources (ms with) much lower flux takes into account cosmological redshift and as- predominantly caused by ISS rather than intrinsic variability. densities. In turn, an increased coverage will constrain the sumes a universal ionization fraction of 1 (12, 13), Intra-day variability refers to sources whose characteristic spatial density and distribution of refracting lenses in the the sources are inferred to be at redshifts z =0.45 timescale is less than 24 h, while inter-day variability refers Galaxy, including whether they are associated with particular to 0.96, corresponding to comoving distances of to sources whose timescale exceeds this. Galactic structures (Fiedler et al. 1994a). A larger number of 1.7 to 3.2 Gpc (Table 1). Scintillating AGNs are of astrophysical interest because ESEs will also allow investigation into any potential diversity In principle, pulse scatter-broadening mea- the small angular sizes that they must possess in order to in light-curve shapes, and whether any such diversity can be surements can constrain the location and strength exhibit rapid ISS require brightness temperatures near, or understood in terms of simple models (Clegg et al. 1998) of an intervening scattering screen (14). FRBs possibly in excess of, the 1012 K inverse Compton limit for in- or the extent to which more complex models of the density 110627, 110703, and 120127 are too weak to coherent synchrotron radiation. In several cases over the past enable the determination of any scattering; how- structures, background sources, or both, are required. decade, it appeared that the Doppler boosting factors required ever, FRB 110220 exhibits an exponential scat- Real-time characterisation of ESEs: Astraightforwardap- in order to reconcile the apparent brightness temperature with tering tail (Fig. 1). There are at least two possible sources and locations for the responsible scatter- proach to elucidating the nature of the structures responsible the inverse Compton limit were well in excess of the typical ing screens: a host galaxy or the IGM. It is pos- for ESEs would be to conduct additional observations of a values determined from Very Long Baseline interferometric sible that both contribute to varying degrees. source while it is undergoing an ESE. Events like those seen observations of quasars. Estimates of the Doppler boosting towards quasars B0954 658 and B1749 096 (Fiedler et al. factors are subject to considerable uncertainty related to the For screen-source, Dsrc,andscreen-observer, + + D ,distances,themagnitudeofthepulsebroad- 1994b) or B1741 038 (Fiedler et al. 1992; Clegg, Fey, & distance of the scattering material from the Earth, but many obs − ening resulting from scattering is multiplied by Fiedler, 1996) should be seen at the rate of roughly 60 per of the most rapid IDVs appear to require Doppler factors of a 2 the factor DsrcDobs= Dsrc + Dobs .Forascreen Fig.year 1. The in frequency-integrated an all-sky monitoring flux densities forsurvey the four with FRBs. The1mJyRMSsen- time resolutions match the few tens (Macquart et al. 2000; Rickett, Kedziora-Chudczer, and source located inð a distant galaxy,Þ this effect levelsitivity, of dispersive with smearing in10 the in central progress frequencychannel(0.8,0.6,0.9,and0.5ms,respectively). at any given∼ moment. Thus, in &Jauncey2002;Bignalletal.2006;Macquart&deBruyn contrast to the∼ GBI monitoring program, for which only one 2007). ESE was identified while it was ongoing, VAST will enable More recently, it appears that the fastest manifestations of 54 5JULY2013real-time VOL341 follow-up.SCIENCE www.sciencemag.org IDV are associated with sources whose lines of sight intersect Some key follow-up measurements could be made at more local, albeit much more inhomogeneous, patches of higher frequencies, particularly in the optical to X-ray bands, turbulence than previously supposed. Thus, there has been a which should directly reveal the dense neutral gas clouds transfer of our ignorance away from the of the AGN

PASA, 30, e006 (2013) doi:10.1017/pasa.2012.006 Transient Event Rate at 1.4 GHz The Astrophysical Journal,768:165(20pp),2013May10 Mooley et al.

Figure 13. Normalized 1.4 GHz differential radio-source counts for persistent sources from de Zotti et al. (2010) and the normalized areal density of transients (or limits) as a function of the flux density for various surveys at this frequency. The Bannister et al. (2011)surveyat0.84GHziscoloreddifferentlythantheotherMooley et al. (2013) surveys. Most of the surveys are upper limits (wedge symbols) and the sampled phase space is shown by the gray shaded area. Upper limits from Frail et al. (1994) and Bower & Saul (2011) do not explore any new part of the phase space (non-gray area), and hence have been left out of this diagram. Our upper-limit is labeled as “E-CDFS.” Three surveys have transient detections so far, the 2σ error bars for which are shown according to Gehrels (1986). Note that Thyagarajan et al. (2011) and Bannister et al. (2011) may have identified a few strong variables as transients (see Section 5.2), which would make their detections move downward on this plot. The black solid line is the model for AGNs and star-forming galaxies from Condon (1984). Lines of constant areal density are shown as blue dotted lines. The horizontal dashed lines are estimates for the areal density for known and expected classes of long-duration radio transients taken directly from Frail et al. (2012). The areal density for Swift J1644+57-like tidal disruption events has been modified according to Berger et al. (2012)toreflecttheirtruerateat1.4GHz.Upperlimitsfrom the ASKAP-VAST surveys are estimated to be an order of magnitude or more below the rate of orphan gamma-ray burst afterglows, and to have an rms sensitivity ranging between 10 µJy and 0.5 mJy. (A color version of this figure is available in the online journal.)

5.2. Limits on Transient Areal Density and Rate adurationtdur less than the shortest time between epochs of 1 2 1 (>0.21 mJy) < 268(tdur/0.5day)− deg− yr− . We searched our multi-epoch data for transients but found ℜ Our upper limit on the areal density of transient sources at none. The search was conducted on each image out to a radius sub-mJy levels can be compared with the predictions based of 21.′5fromthepointingcenter.Thesingle-epochareaoutto on previous surveys. The Bower et al. (2007)surveyisa 2 2 that radius is 0.40 deg ,oratotalareaof20deg for all 49 useful benchmark since their areal density dominates all known epochs. However, the sensitivity of the VLA antennas is not classes of transients. Adopting their measured two epoch rate uniform across this area. The primary beam response is well 2 of κ(>0.37 mJy) 1.5deg− and assuming a Euclidean described by a Gaussian with a half-width to half-maximum of source distribution= (i.e., γ 3/2), we predict κ(>0.21 mJy) 15 ,fallingto20%responseatoursearchradiusof21.5. At 2 = = ′ ′ 3.5deg− at the flux density limit of our current survey. the pointing center, the 7σ flux density limit was approximately An alternative way to look at our results is to compare our 210 µJy for each epoch. null detection to the expected number of Bower et al. transients In order to calculate a limit on the areal density of any expected in our data set. We use the parameterization of Fender putative transient population, we follow Ofek et al. (2011)and &Bell(2011)forthepredictedBoweretal.transientrateasa parameterize the source number-count function as a power law function of flux density, of the form γ κ(>S) κ0(S/S0)− , (4) κ Sν = log 2 1.5log 5.13, (5) deg− = − Jy − where S is the peak flux density, κ(>S)istheskysurface ! " ! " density of sources brighter than S, κ0 is the sky surface density where κ is the snapshot rate, and Sν denotes the detection of sources brighter than S0,andγ is the power-law index of threshold of the observations at the pointing center (i.e., 7σ the source number-count function. We assume for simplicity a 210 µJy). Integrating both sides of Equation (5)overthe= homogeneous source distribution in a Euclidean universe so that azimuthal angle and in θ out to 21.′5, we get about 0.42 transients γ 3/2. The one-sided 2σ upper limit on the areal density is per epoch if the Bower et al. (2007)transientsarereal.Since three= events (Gehrels 1986). Therefore, using Equation (C5) we have 49 epochs, we expect to have about 21 Bower et al. in Ofek et al. (2011), we find that the 2σ upper limit on transients in our E-CDFS data set. 2 areal density to a flux limit of 210 µJy is 18.0 deg− per Our search on the E-CDFS field suggests that the areal density epoch. Given that we have 49 epochs, the 2σ upper limit of radio transients is an order of magnitude or more below 2 2 on the areal density is κ(>0.21 mJy) < 0.37 deg− .Wecan the rate measured by Bower et al. (2007; i.e., <0.37 deg− 2 further estimate an upper limit on the transient rate assuming versus 3.5 deg− ). Alternatively, we find a 2σ upper limit of <3

18 Ant Schinckel (CSIRO) SKA ASKAP LSST DECam SKA ASKAP Space Hubble Telescope , ASKAP, SKA ASKAP, , AperTIF MWA

MWA, LOFAR, LOFAR, MWA, widefields, rapid surveys, enormousdatarates

MWA

- -

Manysmall antennas orfocal plane arrays new → detection algorithms real-time → processing dedicated → supercomputers MWA

Wide-Field Radio Astronomy Radio Wide-Field › Wide-Field Radio Astronomy

› ASKAP Tucana survey (6/36 dishes) › MWA Transients Survey (Heywood et al. 2015) (MWATS; Bell et al. 2014) - 150 deg2 in 12 hrs at 1.4 GHz - 16000 deg2 in 10 hrs at 150 MHz - 3 epochs to 1 mJy; 2000 sources - 24 epochs to 10 mJy; - all-sky image database: ~50 PB 20000 sources

Ian Heywood / ACES / CSIRO David Kaplan / MWATS Transient Event Rate at 1.4 GHz The Astrophysical Journal,768:165(20pp),2013May10 Mooley et al.

The Astrophysical Journal,768:165(20pp),2013May10 Mooley et al.

ASKAP

10-4

10-5 Figure 13. Normalized 1.4 GHz differential radio-source counts for persistent sources from de Zotti et al. (2010) and the normalized areal density of transients (or limits) as a function of the flux density for various surveys at this frequency. The Bannister et al. (2011)surveyat0.84GHziscoloreddifferentlythantheotherMooley et al. (2013) Figuresurveys. 13. MostNormalized of the surveys 1.4 GHz are upper differential limits radio-source (wedge symbols) counts and for the persistent sampled sources phase space from deis shown Zotti et by al. the (2010 gray) shadedand the area. normalized Upper limitsareal density from Frail of transients et al. (1994 (or) limits)and Bower as a & function Saul (2011 of the) do flux not density explore for any various new part surveys of the at phase this frequency. space (non-gray The Bannister area), and et hence al. (2011 have)surveyat0.84GHziscoloreddifferentlythantheother been left out of this diagram. Our upper-limit is labeled surveys.as “E-CDFS.” Most Threeof the surveys haveare upper transient limits detections (wedge symbols) so far, the and 2σ theerror sampled bars for phase which space are shown is shown according by the gray to Gehrels shaded (1986 area.). Upper Note thatlimits Thyagarajan from Frail et al. ( 19942011) and BowerBannister & Saul et al. (2011 (2011)) do may not have explore identified any new a few part strong of the variables phase space as transients (non-gray (see area), Section and hence5.2), have which been would left makeout of their this diagram. detections Our move upper-limit downward is labeled on this asplot. “E-CDFS.” The black Three solid surveys line is the have model transient for AGNs detections and star-forming so far, the 2σ galaxieserror bars from for Condon which are (1984 shown). Lines according of constant to Gehrels areal density (1986). are Note shown that Thyagarajan as blue dotted et lines.al. (2011 The) andhorizontal Bannister dashed et al. lines (2011 are) estimates may have for identified the areal a density few strong for known variables and as expected transients classes (see of Section long-duration5.2), which radio would transients make taken their directly detections from move Frail downward et al. (2012 on). The this plot.areal The density black for solid Swift line J1644+57-like is the model tidal for AGNs disruption and events star-forming has been galaxies modified from according Condon to (1984 Berger). Lines et al. of (2012 constant)toreflecttheirtruerateat1.4GHz.Upperlimitsfrom areal density are shown as blue dotted lines. The horizontalthe ASKAP-VAST dashed lines surveys are estimates are estimated for the to areal be an density order of for magnitude known and or expected more below classes the of rate long-duration of orphan gamma-ray radio transients burst taken afterglows, directly and from to have Frail an et al. rms (2012 sensitiv). Theity arealranging density between for Swift 10 µJy J1644+57-like and 0.5 mJy. tidal disruption events has been modified according to Berger et al. (2012)toreflecttheirtruerateat1.4GHz.Upperlimitsfrom the(A color ASKAP-VAST version of thissurveys figure are is estimated available toin bethe an online order journal.) of magnitude or more below the rate of orphan gamma-ray burst afterglows, and to have an rms sensitivity ranging between 10 µJy and 0.5 mJy. (A color version of this figure is available in the online journal.) 5.2. Limits on Transient Areal Density and Rate adurationtdur less than the shortest time between epochs of 1 2 1 (>0.21 mJy) < 268(tdur/0.5day)− deg− yr− . We searched5.2. Limits our on multi-epoch Transient Areal data Density for transients and Rate but found adurationℜ Our uppertdur limitless on than the the areal shortest density time of transient betweensources epochs of at 1 2 1 none. The search was conducted on each image out to a radius sub-mJy(>0.21 mJy)levels< can268( betdur compared/0.5day)− withdeg the− yr predictions− . based We searched our multi-epoch data for transients but found ℜ of 21.′5fromthepointingcenter.Thesingle-epochareaoutto onOur previous upper surveys.limit on the The areal Bower density et al. of ( transient2007)surveyisa sources at none. The search was conducted2 on each image out2 to a radius that radius is 0.40 deg ,oratotalareaof20deg for all 49 sub-mJyuseful benchmark levels can since be their compared areal density with the dominates predictions all known based of 21.5fromthepointingcenter.Thesingle-epochareaoutto epochs.′ However, the sensitivity of the VLA antennas is not onclasses previous of transients. surveys. Adopting The Bower their et measured al. (2007 two)surveyisa epoch rate that radius is 0.40 deg2,oratotalareaof20deg2 for all 49 uniform across this area. The primary beam response is well usefulof κ(> benchmark0.37 mJy) since1 their.5deg areal2 densityand assuming dominates a Euclidean all known epochs. However, the sensitivity of the VLA antennas is not − described by a Gaussian with a half-width to half-maximum of classessource distribution of transients.= (i.e., Adoptingγ 3/2), their we measured predict κ( two>0. epoch21 mJy) rate uniform across this area. The primary beam response is well 2 15′,fallingto20%responseatoursearchradiusof21.′5. At of3.5degκ(>02.37at mJy) the flux density1.5deg= limit− ofand our assuming current survey. a Euclidean= described by a Gaussian with a half-width to half-maximum of − = the pointing center, the 7σ flux density limit was approximately sourceAn alternative distribution way (i.e., toγ look3/ at2), our we results predict isκ to(> compare0.21 mJy) our 15 ,fallingto20%responseatoursearchradiusof21.5. At 2 = = 210′ µJy for each epoch. ′ 3null.5deg detection− at the to flux the expected density limit number of our of Bowercurrent et survey. al. transients the pointing center, the 7σ flux density limit was approximately In order to calculate a limit on the areal density of any expectedAn alternative in our data way set. to We look use at the our parameterization results is to compare of Fender our 210 µJy for each epoch. putative transient population, we follow Ofek et al. (2011)and null&Bell( detection2011)forthepredictedBoweretal.transientrateasa to the expected number of Bower et al. transients In order to calculate a limit on the areal density of any parameterize the source number-count function as a power law expectedfunction of in flux our data density, set. We use the parameterization of Fender putativeof the form transient population, we follow Ofek et al. (2011)and &Bell(2011)forthepredictedBoweretal.transientrateasa parameterize the source number-count functionγ as a power law function of flux density,κ S of the form κ(>S) κ0(S/S0)− , (4) log 1.5log ν 5.13, (5) = deg 2 = − Jy − γ κ− S where S is the peakκ(>S flux) density,κ0(S/Sκ0()>S− ,)istheskysurface(4) ! " ! ν " = log 1.5log 5.13, (5) density of sources brighter than S, κ is the sky surface density where κ is thedeg snapshot2 = rate,− and SJydenotes− the detection 0 ! − " !ν " whereof sourcesS is brighter the peak than fluxS0 density,,andγ isκ(>S the)istheskysurface power-law index of threshold of the observations at the pointing center (i.e., 7σ densitythe source of sources number-count brighter function. than S, κ We0 is assume the sky for surface simplicity density a where210 µJy).κ is Integrating the snapshot both rate, sides and ofSν Equationdenotes the (5)overthe detection= ofhomogeneous sources brighter source than distributionS0,and inγ ais Euclidean the power-law universe index so that of thresholdazimuthal of angle the and observations in θ out to at 21 the.5, we pointing get about center 0.42 (i.e., transients 7σ ′ = theγ source3/2. The number-count one-sided 2 function.σ upper limitWe assume on the for areal simplicity density is a 210per epochµJy). if Integrating the Bower both et al. sides (2007 of)transientsarereal.Since Equation (5)overthe homogeneousthree= events (Gehrels source distribution1986). Therefore, in a Euclidean using universe Equation so (C5) that azimuthalwe have 49 angle epochs, and in weθ out expect to 21 to.′5, have we get about about 21 0.42 Bower transients et al. γin Ofek3/2. et The al. one-sided (2011), we 2σ findupper that limit the on 2 theσ upper areal density limit on is pertransients epoch in if our the E-CDFS Bower et data al. ( set.2007)transientsarereal.Since = 2 threeareal eventsdensity (Gehrels to a flux1986 limit). Therefore, of 210 µJy using is 18.0 Equation deg− (C5)per weOur have search 49 epochs, on the E-CDFS we expect field to suggests have about that 21the Bowerareal density et al. inepoch. Ofek Given et al. that (2011 we), have we find 49 epochs,that the the 2σ 2upperσ upper limit limit on transientsof radio transients in our E-CDFS is an data order set. of magnitude or more below 2 2 2 arealon the density areal density to a flux is κ limit(>0.21 of mJy) 210 µ0.21 mJy) < 0.37 deg− .Wecan the rate measured by Bower et al. (2007; i.e., <0.37 deg− 2 further estimate an upper limit on the transient rate assuming 18 versus 3.5 deg− ). Alternatively, we find a 2σ upper limit of <3 18 Challenge I : Ionosphere

MWA time series (Natasha Hurley-Walker) › Spatial and temporal fluctuations in plasma density produce position shifts, defocusing, scintillation

› Huge instantaneous MWA field of view (~1000 deg2) - track 1000 source positions at 2-minute cadence - map vector offsets as function of time (Loi et al. 2015a, 2015b) - robust correction for ionospheric refraction

› Organized strips of alternating position shifts - bands of underdensity and overdensity - aligned with projection of Earth’s magnetic field - stereoscopic imaging: h = 570 ± 40 km → cylindrical density ducts, coupling ionosphere and plasmasphere via whistler waves → direct 4D visualization of bulk plasma drifts

Loi et al. (2015a) Challenge II : Source Finding

› Identification of interesting events needs to be catalogue-based, not image-based - missed/blended sources will trigger huge numbers of false1820 alarmsP. J. H a n c o ck e t a l . - 99% accuracy is notthe good small differences enough! between each of these source finders, and

in how to optimize the approach to avoid even the residual small et al. (2012) Hales level of incompleteness and FDR. In surveys such as EMU (Norris › BLOBCAT (Hales et al. et2012; al. 2011), github with an expected) 70 million sources, an FDR of even 1 per cent translates into 700 000 false sources. This clearly has an - flood-fill: superior toimpact gaussian on the study offitting rare or unusual behaviour. In particular we are interested in how the source-finding algorithm affects the final output catalogue at a level that is far more detailed than previously 1820 P.› J. AEGEAN, H a n c o ck e t a l . BANE & MIMASexplored. With this in mind we now delve into specific cases in which existing source-finding packages fail. the small differences (Hancock between each et of theseal. 2012, source finders, 2015; and github) in how to optimize the approach to avoid even the residual small6MISSEDSOURCES level of incompleteness- andAEGEAN: FDR. In surveys suchLaplacian as EMU (Norris for robust component separation et al. 2011), with an expected 70 million sources, an FDR of evenWe are now at the stage where we can consider the real sources that 1 per cent translates into- 700BANE: 000 false sources.fast This& accurate clearly haswere an missedbackground by the source-finding estimation packages, as well as the false impact on the study of rare or unusual behaviour. In particulardetections we that these programs generate. There are two populations are interested in how- the source-findingMIMAS: algorithm describe/combine/mask affects the finalof sources that are missed regions by one or more of the source-finding output catalogue at a level that is far more detailed than previouslypackages as will be discussed in Sections 6.1 and 6.2. 18201820 P.P. J. J. H H a a n n c c o o ck ck e e t t a a l l . . explored. With this in mind we now delve into specific cases in which existing source-finding packages fail. Image SFIND6.1 Isolated faint SEXtractor sources Selavy IMSAD AEGEAN thethe small small differences differences between between each each of of these these source source finders, finders, and and In the simulated image, for which no clustering was taken into inin how how to to optimize optimize the the approach approach to to avoid avoid even even the the residual residual small small 6MISSEDSOURCES account, 99.5 per cent of the islands contained a single source. levellevel of of incompleteness incompleteness and and FDR. FDR. In In surveys surveys such such as as EMU EMU (Norris (Norris The first population of sources that was not well detected by the etet al. al. 2011), 2011), with with an an expected expected 70 70 million million sources, sources, an an FDR FDR of of even evenWe are now at the stage where we can consider the real sources that source-finding algorithms are isolated faint sources. These sources 11 per per cent cent translates translates into into 700 700 000 000 false false sources. sources. This This clearly clearly has haswere an an missed by the source-finding packages, as well as the false have a true flux greater than the threshold, but have few or no pixels impactimpact on on the the study study of of rare rare or or unusual unusual behaviour. behaviour. In In particular particulardetections we we that these programs generate. There are two populations above the threshold due to the addition of noise. SFIND and SEXTRAC- areare interested interested in in how how the the source-finding source-finding algorithm algorithm affects affects the the final finalof sources that are missed by one or more of the source-finding TOR require an island to have more than some minimum number of outputoutput catalogue catalogue at at a a level level that that is is far far more more detailed detailed than than previously previouslypackages as will be discussed in Sections 6.1 and 6.2. pixels for it to be considered a candidate source. IMSAD, SELAVY and explored.explored. With With this this in in mind mind we we now now delve delve into into specific specific cases cases in in AEGEAN have no such requirement. The number of sources that are whichwhich existing existing source-finding source-finding packages packages fail. fail. 6.1 Isolated faint sources not seen in a catalogue due to the effects of noise can be calcu- In the simulatedHancock image, for et whichal. (2012) no clustering was takenlated into directly and is essentially the inverse problem to that of false Figure 8. Top left: a section of the simulated image. Remainder: the fitting 6MISSEDSOURCES6MISSEDSOURCES account, 99.5 per cent of the islands contained a single source. detections. A correction can be applied to any statistical measure residual for each of the source-finding algorithms. AEGEAN was the only The first population of sources that was not well detected byextracted the from the catalogue in order to account for these missed algorithm to fit all three sources, over both islands. WeWe are are now now at at the the stage stage where where we we can can consider consider the the real real sources sources that that sources. The only way to recover all sources with a true flux above werewere missed missed by by the the source-finding source-finding packages, packages, as as well well as as the the false falsesource-finding algorithms are isolated faint sources. These sources have a true flux greater than the threshold, but have few or no pixelsa given limit is to have a threshold that is well below this limit, detectionsdetections that that these these programs programs generate. generate. There There are are two two populations populations either by producing a more sensitive image or by accepting a larger the island and the fitting residual is inspected. If the fitting residual above the threshold due to the addition of noise. SFIND and SEXTRAC- ofof sources sources that that are are missed missed by by one one or or more more of of the the source-finding source-finding number of false detections. Since this noise affected population of meets some criterion then the fit is considered to be ‘good’ and TOR require an island to have more than some minimum number of packagespackages as as will will be be discussed discussed in in Sections Sections 6.1 6.1 and and 6.2. 6.2. sources cannot be reduced by an improved source-finding algo- a single source is reported. If the residual is ‘poor’ then the fit is pixels for it to be considered a candidate source. IMSAD, SELAVY and rithm, and can be accounted for in a statistically robust way, we will redone with an extra component. Once either the fitting residual is AEGEAN have no such requirement. The number of sources that are consider this population to be non-problematic. found to be ‘good’ or some maximum number of components has 6.16.1 Isolated Isolated faint faint sources sources not seen in a catalogue due to the effects of noise can be calcu- been fit, the iteration stops and the extracted sources are reported. A lated directly and is essentially the inverse problem to that of false InIn the the simulated simulated image, image, for for which which no no clustering clustering was was taken taken into into Figure 8. Top left: a section of the simulated image. Remainder: the fittingdisadvantage of this method is that if the number of allowed Gaus- detections. A correction can be applied to any statistical measure account,account, 99.5 99.5 per per cent cent of of the the islands islands contained contained a a single single source. source. 6.2 Islandsresidual with multiple for each of sources the source-finding algorithms. AEGEAN was the onlysians (n) is poorly chosen, islands containing single faint sources extracted from the catalogue in order to account for these missed algorithm to fit all three sources, over both islands. can have a ‘better’ fitting residual when fit by multiple components, TheThe first first population population of of sources sources that that was was not not well well detected detected by by the the The second population of sources that is not well detected by the sources. The only way to recover all sources with a true flux above and source fragmentation occurs. When a source is fragmented it is source-findingsource-finding algorithms algorithms are are isolated isolated faint faint sources. sources. These These sources sources source-finding packages are the sources that are within an island of a given limit is to have a threshold that is well below this limit, difficult to extract the overall source parameters from the multiple havehave a a true true flux flux greater greater than than the the threshold, threshold, but but have have few few or or no no pixels pixels pixels thatthe contains island multiple and the components. fitting residual Examples is inspected. of such If the islands fitting residual either by producing a more sensitive image or by accepting a larger Gaussians that were used in the fitting of the source. In particular aboveabove the the threshold threshold due due to to the the addition addition of of noise. noise.SFINDSFINDandandSEXTRACSEXTRAC-- are shownmeets in Figs some 8–10. criterion If a source-finding then the fit algorithm is considered is unable to be to ‘good’ and number of false detections. Since this noise affected population of the source flux is not simply the sum of the flux of the fragments. If TORTORrequirerequire an an island island to to have have more more than than some some minimum minimum number number of of correctly characterizea single source multiple is reported. sources withinIf the residual an island, is some ‘poor’ or then all the fit is sources cannot be reduced by an improved source-finding algo- the chosen value of n is too small then not all of the sources within pixelspixels for for it it to to be be considered considered a a candidate candidate source. source.IMSADIMSAD,,SELAVYSELAVYandand of these sourcesredone will with be an missed. extra component. There are two Once approaches either the used fitting by residual is rithm, and can be accounted for in a statistically robust way, we will an island will be characterized. These uncharacterized sources will AEGEANAEGEANhavehave no no such such requirement. requirement. The The number number of of sources sources that that are are the tested algorithmsfound to be to ‘good’ extract or multiple some maximum sources from number an island of components of has consider this population to be non-problematic. contaminate the fitting of the previously identified sources resulting notnot seen seen in in a a catalogue catalogue due due to to the the effects effects of of noise noise can can be be calcu- calcu- pixels – iterativebeen fit, fitting the iteration and de-blending. stops and theEach extracted of these sources approaches are reported. A in a poor characterization of the island. latedlated directly directly and and is is essentially essentially the the inverse inverse problem problem to to that that of of false false FigureFigure 8. 8. TopTop left: left: a a section section of of the the simulated simulated image. image. Remainder: Remainder:can fail the the to fitting fittingdisadvantage characterize anof this island method of sources is that for if the different number reasons, of allowed Gaus- When the flux ratio of components within an island of pixels detections.detections. A A correction correction can can be be applied applied to to any any statistical statistical measure measure6.2 Islandsresidualresidual withmultiple for for each each of of sources the the source-finding source-finding algorithms. algorithms. AEGEANAEGEANandwaswas will the the now onlysians only be (discussedn) is poorly in detail. chosen, islands containing single faint sources extractedextracted from from the the catalogue catalogue in in order order to to account account for for these these missed missed algorithmalgorithm to to fit fit all all three three sources, sources, over over both both islands. islands. can have a ‘better’ fitting residual when fit by multiple components,becomes very large, an iterative fitting approach can fail. The cause The second population of sources that is not well detected by the sources.sources. The The only only way way to to recover recover all all sources sources with with a a true true flux flux above above and source fragmentation occurs. When a source is fragmentedof it is this failure is related to the performance of an ideal Gaussian source-finding packages are the sources that are within an island of fitting routine. Fig. 7 shows the fractional error in measuring the aa given given limit limit is is to to have have a a threshold threshold that that is is well well below below this this limit, limit, 6.2.1 Iterativedifficult fitting to extract the overall source parameters from the multiple pixels thatthe containsthe island island multiple and and the the components.fitting fitting residual residual Examples is is inspected. inspected. of such If If the the islands fitting fitting residual residual eithereither by by producing producing a a more more sensitive sensitive image image or or by by accepting accepting a a larger larger Gaussians that were used in the fitting of the source. In particularamplitude of a Gaussian. For high SNR sources, the absolute flux are shownmeetsmeets in Figs some some 8–10. criterion criterion If a source-finding then then the the fit fit algorithm is is considered considered is unable to to be be to ‘good’ ‘good’ and and numbernumber of of false false detections. detections. Since Since this this noise noise affected affected population population of of The first approachthe source to flux characterizing is not simply an the island sum of of the multiple flux of the com- fragments.error If can be orders of magnitude below the rms image noise, so correctly characterizeaa single single source source multiple is is reported. reported. sources Ifwithin If the the residual residualan island, is is some ‘poor’ ‘poor’ or then then all the the fit fit is is sourcessources cannot cannot be be reduced reduced by by an an improved improved source-finding source-finding algo- algo- ponents isthe an choseniterative value one ofwhichn is toorelies small on thenthe notion not all of of a the fitting sources withinit may be expected that the maximum flux in the fitting residual of these sourcesredoneredone will with with be an an missed. extra extra component. component. There are two Once Once approaches either either the the used fitting fitting by residual residual is is rithm,rithm, and and can can be be accounted accounted for for in in a a statistically statistically robust robust way, way, we we will will residual. Thean islandfittingwill residual be characterized. is the difference These between uncharacterized the data and sourcesshould will also be at or below the rms image noise. However, the the testedfound algorithmsfound to to be be to ‘good’ ‘good’ extract or or multiple some some maximum maximum sources from number number an islandof of components components of has has considerconsider this this population population to to be be non-problematic. non-problematic. the modelcontaminate fit. In the iterative the fitting approach of the previously a single Gaussian identified is sources fit to resultingmain contribution to the flux seen in the fitting residual is not from pixels – iterativebeenbeen fit, fit, fitting the the iteration iteration and de-blending. stops stops and and the Eachthe extracted extracted of these sources sources approaches are are reported. reported. A A in a poor characterization of the island. C can fail todisadvantagedisadvantage characterizeof an of this thisisland method method of sources is is that that for if if the the different number number reasons, of of allowed allowed Gaus- Gaus- ⃝ 2012 The Authors, MNRAS 422, 1812–1824 When the flux ratio of components within an island of pixels C 6.26.2 Islands Islands with with multiple multiple sources sources and will nowsianssians be ( discussed(nn)) is is poorly poorly in detail. chosen, chosen, islands islands containing containing single single faint faint sources sources Monthly Notices of the Royal Astronomical Society ⃝ 2012 RAS becomes very large, an iterative fitting approach can fail. The cause cancan have have a a ‘better’ ‘better’ fitting fitting residual residual when when fit fit by by multiple multiple components, components, TheThe second second population population of of sources sources that that is is not not well well detected detected by by the the of this failure is related to the performance of an ideal Gaussian andand source source fragmentation fragmentation occurs. occurs. When When a a source source is is fragmented fragmented it it is is source-findingsource-finding packages packages are are the the sources sources that that are are within within an an island island of of fitting routine. Fig. 7 shows the fractional error in measuring the 6.2.1 Iterativedifficultdifficult fitting to to extract extract the the overall overall source source parameters parameters from from the the multiple multiple pixelspixels that that contains contains multiple multiple components. components. Examples Examples of of such such islands islands amplitude of a Gaussian. For high SNR sources, the absolute flux GaussiansGaussians that that were were used used in in the the fitting fitting of of the the source. source. In In particular particular areare shown shown in in Figs Figs 8–10. 8–10. If If a a source-finding source-finding algorithm algorithm is is unable unableThe to to first approach to characterizing an island of multiple com- error can be orders of magnitude below the rms image noise, so thethe source source flux flux is is not not simply simply the the sum sum of of the the flux flux of of the the fragments. fragments. If If correctlycorrectly characterize characterize multiple multiple sources sources within within an an island, island, some some or orponents all all is an iterative one which relies on the notion of a fitting it may be expected that the maximum flux in the fitting residual thethe chosen chosen value value of ofnnisis too too small small then then not not all all of of the the sources sources within within ofof these these sources sources will will be be missed. missed. There There are are two two approaches approaches used usedresidual. by by The fitting residual is the difference between the data and should also be at or below the rms image noise. However, the anan island island will will be be characterized. characterized. These These uncharacterized uncharacterized sources sources will will thethe tested tested algorithms algorithms to to extract extract multiple multiple sources sources from from an an island islandthe of of model fit. In the iterative approach a single Gaussian is fit to main contribution to the flux seen in the fitting residual is not from contaminatecontaminate the the fitting fitting of of the the previously previously identified identified sources sources resulting resulting pixelspixels – – iterative iterative fitting fitting and and de-blending. de-blending. Each Each of of these these approaches approaches inin a a poor poor characterization characterization of of the the island. island. C cancan fail fail to to characterize characterize an an island island of of sources sources for for different different reasons, reasons, ⃝ 2012 The Authors, MNRAS 422, 1812–1824 WhenWhen the the flux flux ratio ratio of of components components within within an an island island of of pixels pixels Monthly Notices of the Royal Astronomical Society C 2012 RAS andand will will now now be be discussed discussed in in detail. detail. ⃝ becomesbecomes very very large, large, an an iterative iterative fitting fitting approach approach can can fail. fail. The The cause cause ofof this this failure failure is is related related to to the the performance performance of of an an ideal ideal Gaussian Gaussian fittingfitting routine. routine. Fig. Fig. 7 7 shows shows the the fractional fractional error error in in measuring measuring the the 6.2.16.2.1 Iterative Iterative fitting fitting amplitudeamplitude of of a a Gaussian. Gaussian. For For high high SNR SNR sources, sources, the the absolute absolute flux flux TheThe first first approach approach to to characterizing characterizing an an island island of of multiple multiple com- com- errorerror can can be be orders orders of of magnitude magnitude below below the the rms rms image image noise, noise, so so ponentsponents is is an an iterative iterative one one which which relies relies on on the the notion notion of of a a fitting fitting itit may may be be expected expected that that the the maximum maximum flux flux in in the the fitting fitting residual residual residual.residual. The The fitting fitting residual residual is is the the difference difference between between the the data data and and shouldshould also also be be at at or or below below the the rms rms image image noise. noise. However, However, the the thethe model model fit. fit. In In the the iterative iterative approach approach a a single single Gaussian Gaussian is is fit fit to to mainmain contribution contribution to to the the flux flux seen seen in in the the fitting fitting residual residual is is not not from from

CC ⃝⃝20122012 The The Authors, Authors, MNRAS MNRAS422,422,1812–18241812–1824 CC MonthlyMonthly Notices Notices of of the the Royal Royal Astronomical Astronomical Society Society⃝⃝20122012 RAS RAS Challenge III : Source Classification

5.3 Classification method

› ASKAP event rate will be large, ~50 per night CV - need prompt classification & follow-up Lo et Lo al. (2013)

› Random forest machine-learning algorithm

(Lo et al. 2014; Farrell et al. 2015) Flux XRB - variable light curves taken from 3XMM - 869 identified sources used as training set 5.6 Catalogue of probabilistically classified XMM variable sources (AGN, CV, GRB, SSS, star, ULX, XRB) Time - input: time-series plus contextual features → 92% – 96% classification accuracy Lo et Lo al. (2013) › Apply algorithm to 2876 other 3XMM sources - compare to 101 sources with known IDs → agree in 93 cases; most others ambiguous - identification of ~20 “outlier” sources

Figure 5.2: Example light curves for the seven types of X-ray sources in our training set.

93 Figure 5.10: Confusion matrix from performing 10-fold cross-validationonthetrainingset using the RF classifier with time-series and contextual features. The color bar represents the true positive rate. The overall accuracy is 97%.

to be highly informative, with all four hardness ratios placed in the top 10 of most informative features. On the other hand, time-series features do not rankhighlyonthelist.Themost informative of the time-series features are features related to the Lomb-Scargle periodogram.

5.6 Catalogue of probabilistically classified XMM variable sources

5.6.1 Results

Using the entire training set, we constructed a RF classification model using the method de- scribed in Section 5.5.Thenweappliedthisclassificationmodeltothesetofunknown 2XMMi variable sources. For sources where there are more than one detection, we classified each de- tection separately and combined the results by averaging theoutputclassmembershipprobabil- ities. Table 5.4 shows the number of unknown sources classified as one of seven classes. The majority of the unknown sources are classified as stars. We also compiled a downloadable table of the class membershipprobabilities.Table5.5 shows a portion of that table.

105 Summary NASA › Time-domain astronomy is rapidly evolving

› Wide-field surveys: unique strengths & challenges - atmospheric distortions can now be precisely characterized … and contain new science

- robust source fitting and cataloging CSIRO / Swinburne - automatic classification of both

expected categories and outlier sources

› Goal: address major topics in fundamental physics and

- unbiased census of cosmic explosions Digital Berry/Skyworks NASA/Dana - propagation as unique probe of turbulence and baryonic matter

- high-time-resolution Universe: a new frontier

› 2020- : Exploration of full Dynamic Universe with the