OH Megamasers in HI Surveys: Forecasts and a Machine Learning Approach to Separating Disks from Mergers
DDraft version February 26, 2021
Typeset using L A TEX twocolumn style in AASTeX63
OH Megamasers in H I Surveys: Forecasts and a Machine Learning Approach to Separating Disksfrom Mergers
Hayley Roberts, Jeremy Darling, and Andrew J. Baker Center for Astrophysics and Space Astronomy, Department of Astrophysical and Planetary Science, University of Colorado, 389 UCB,Boulder, CO 80309-0389 Department of Physics and Astronomy, Rutgers, The State University of New Jersey, 136 Frelinghuysen Road, Piscataway, NJ08854-8019
AbstractOH megamasers (OHMs) are rare, luminous masers found in gas-rich major galaxy mergers. Inuntargeted neutral hydrogen (H I ) emission-line surveys, spectroscopic redshifts are necessary todifferentiate the λ rest = 18 cm masing lines produced by OHMs from H I
21 cm lines. Next generationH I surveys will detect an unprecedented number of galaxies, most of which will not have spectroscopicredshifts. We present predictions for the numbers of OHMs that will be detected and the potential“contamination” they will impose on H I surveys. We examine Looking at the Distant Universe withthe MeerKAT Array (LADUMA), a single-pointing deep-field survey reaching redshift z HI = 1 . ,as well as potential future surveys with the Square Kilometre Array (SKA) that would observelarge portions of the sky out to redshift z HI = 1 . . We predict that LADUMA will potentiallydouble the number of known OHMs, creating an expected contamination of 1.0% of the survey’s H I sample. Future SKA H I surveys are expected to see up to 7.2% OH contamination. To mitigatethis contamination, we present methods to distinguish H I and OHM host populations withoutspectroscopic redshifts using near- to mid-IR photometry and a k-Nearest Neighbors algorithm.Using our methods, nearly 99% of OHMs out to redshift z OH ∼ . can be correctly identified. Atredshifts out to z OH ∼ . , 97% of OHMs can be identified. The discovery of these high-redshift OHMswill be valuable for understanding the connection between extreme star formation and galaxy evolution. INTRODUCTIONOH megamasers (OHMs) are luminous 18 cm masersfound in (ultra-)luminous infrared galaxies ([U]LIRGs)produced predominantly by major galaxy mergers. Thedominant masing lines occur at 1667 and 1665 MHzin the ground state of OH. These masing lines haveisotropic luminosities between − L (cid:12) and linewidths ranging from 10 to 1000 km s − due to Dopplerbroadening (Darling 2005). This rare phenomenon hasonly been discovered in roughly 110 galaxies out to red-shift z = 0 . (Darling & Giovanelli 2002a).OHMs are associated with high molecular gas density, n ( H ) ∼ cm − , and strong far-IR radiation, makingthem markers for some of the most extreme star forma-tion observed in our local universe (Darling 2007; Lock-ett & Elitzur 2008). These galaxies can provide signifi-cant information about how extreme star formation re- Corresponding author: Hayley [email protected] lates to galaxy evolution, particularly once discovered athigher redshifts. As products of gas-rich major mergers,OHMs can also provide an independent measure of thegalaxy merger rate at a specific evolutionary stage. Fur-ther, the masing lines present in OHMs can be utilizedas Zeeman magnetometers, providing in-situ measure-ments of magnetic fields in nearby galaxies (Robishawet al. 2008; McBride et al. 2014). OHMs are useful astro-nomical tools for understanding many aspects of galaxyevolution.However, in untargeted emission-line surveys for neu-tral hydrogen (H I ), an OH emission line at z OH can“spoof” a 21 cm H I line at z HI if ν HI / (1 + z HI ) = ν OH / (1 + z OH ) (Briggs 1998) where ν HI = 1420 . MHzand ν OH = 1667 . MHz. For example, restframe H I cor-responds to OH at a redshift of z OH = 0 . , while H I atredshift z HI = 0 . would correspond to OH at a redshiftof z OH = 0 . . These two lines have similar linewidthsin their respective environments: H I in spiral galaxiesand OH in major galaxy mergers. Distinguishing be-tween these lines often requires knowledge of a galaxy’s a r X i v : . [ a s t r o - ph . GA ] F e b H. Roberts, J. Darling, A. J. Baker spectroscopic redshift to determine the rest wavelengthfor an observed emission line (e.g., Hess et al. 2021). Forgalaxies that do not have spectroscopic redshifts, disen-tangling H I from OH is not straightforward. This am-biguity particularly becomes a problem for high-redshiftuntargeted line surveys, which will not have prior spec-troscopic redshifts for many of their detected galaxies.Despite the versatility and capacity of OHMs to serveas tools for studying galaxy evolution, line confusion inH I surveys can serve as a source of contamination forthe survey’s main goals. Nonetheless, next-generationH I surveys, such as those with the Square Kilometre Ar-ray (SKA) and its precursors, will be able to detect manynew OHMs at redshifts never before reached. Findingthese unique galaxies in the pool of H I disk detectionswill enable an exciting new era in OHM science.LADUMA (Looking at the Distant Universe withMeerKAT Array; Blyth et al. 2016) is a survey withthe MeerKAT radio interferometer, a precursor instru-ment for the SKA, that will be susceptible to OH/H I confusion. LADUMA will be the deepest neutral hy-drogen survey to date and is expected to detect H I outto redshifts z HI = 1 . . LADUMA’s main science goalsare related to studying the neutral atomic gas contentof galaxies, meaning that OHM detections will contami-nate its H I samples. At low redshift, spectroscopic red-shifts are generally known, so the contamination rateshould be small. At greater distances where fewer spec-troscopic redshifts are currently known and OHM preva-lence increases due to the elevated merger rate, the con-tamination rate threatens to be higher.This paper presents predictions for the numbers ofOH megamasers that will be detected by LADUMA inSection 2 and by other H I surveys in Section 3. Wethen present methods for distinguishing OH from H I inLADUMA and other untargeted line surveys using near-to mid-IR photometry in Section 4. We discuss theseresults and their limitations in Section 5 and summarizeour conclusions in Section 6.Throughout this work, we assume a flat Λ CDM cos-mology with H = 70 km s − Mpc − , Ω m = 0 . , and Ω Λ = 0 . . OH MEGAMASERS IN LADUMAIn this section, we present our predictions for the num-ber of OHMs that will be detected in LADUMA. A fewinputs are needed to enable these predictions: an OHluminosity function, the volume of the survey, and thegalaxy merger rate. The following subections will covereach of these. We then make use of this information topredict the number of OH megamasers and the resultantOHM contamination rate in other H I surveys. 2.1. The OH Luminosity Function
To create a prediction for the number of OHMs thatwill be found in the LADUMA survey, we need to in-tegrate the OH luminosity function (OHLF) over thevolume and luminosity limits of LADUMA. The OHLFpresented in Darling & Giovanelli (2002b) was con-structed from OHMs detected by Arecibo and is validfor . < log( L OH /L (cid:12) ) < . and . < z OH < . .An obvious source of uncertainty is the extrapolation ofthe OHLF to cover redshifts out to z OH = 1 . andluminosities as low as L (cid:12) for the LADUMA survey.The OHLF is defined as the number of OHMs withluminosity ( L OH ) per unit comoving volume ( Mpc ) perlogarithmic interval in L OH , and can be parameterizedas Φ( L OH ) = b L a OH . (1)The values of a and b presented in Darling & Gio-vanelli (2002b) were determined using an error-weightedleast-squares fit. We use a Markov chain Monte Carlo(MCMC) method to refit the OHLF using the data pre-sented in Darling & Giovanelli (2002b) to account forthe correlations between fit parameters. We use emcee ,a Python package for implementing MCMC (Foreman-Mackey et al. 2013). The refit OHLF is Φ( L OH ) = (3 . ± . × − ( L OH /L (cid:12) ) − . ± . Mpc − dex − . (2)The samples from the MCMC allow us to create a repre-sentative sample of possible OHLF parameters for pre-dictions of the number of OH megamasers that will beobserved. This is the OHLF that will be used through-out this work. We later compare how the differencebetween the two OHLFs changes the number of OHMspredicted to be detected.2.2. Volume of LADUMA Field
LADUMA will be observing an area encompassing theExtended Chandra Deep Field South (E-CDFS), cover-ing 0.9 deg at z = 0 . The field of view of MeerKATwill increase with redshift due to the array’s larger pri-mary beam at lower frequencies. The total volume of thefield can be calculated starting with the equation for co-moving volume (equation 28 from Hogg 1999) shownin equation (3), where E ( z ) = (cid:112) Ω M (1 + z ) + Ω Λ , D H = c/H , and D A is the angular diameter distance: dV c = D H (1 + z ) D A E ( z ) d Ω dz. (3)To account for an increasing field of view at higherredshifts, we write d Ω in equation (3) in terms of the H Megamasers in H I Surveys Redshift ( z OH ) O H Lu m i no s i t y S en s i t i v i t y ( L ) L BandUHF and L Band OverlapUHF Band
Observed Frequency (MHz)
Figure 1.
Isotropic 5 σ luminosity sensitivity L OH as afunction of redshift and observed frequency. The verticalblack lines show the low- and high-redshift limits of OH de-tection for LADUMA. The three different curves come fromthe two bands LADUMA will use, UHF and L bands, andthe frequency range where they overlap. redshift-dependent field of view of a telescope with pri-mary beam diameter . c (1+ z ) / ( ν D ) , where D is thediameter of a single dish (for MeerKAT, D = 13 . )and ν is the rest frequency of the line being observed.This diameter yields a redshift-dependent solid angle: Ω = π (cid:34) . c (1 + z ) ν D (cid:35) . (4)Integrating equation (3) over d Ω and substituting inequation (4) yields the differential volume for the LAD-UMA field: dV = π (cid:18) . (cid:19) (cid:18) D A ν D (cid:19) D H (1 + z ) (cid:112) Ω m (1 + z ) + Ω Λ dz. (5)Integrating equation (5) over the redshift limits of theLADUMA survey for detecting OH ( z OH = 0 . − . ), we obtain a total volume of V = 0 . Gpc .2.3. Sensitivity of LADUMA
The luminosity sensitivity for integrating the OHLFchanges with redshift and with MeerKAT band (eitherthe UHF or L band). The sensitivity is calculated us-ing equation (6), where ∆ S ν is the interferometric sen-sitivity, SEFD is the system equivalent flux density asreported by MeerKAT , η corr is the correlator efficiency, MeerKAT SEFD values can be accessed via the SARAOMeerKAT specifications page. N ant is the number of antennas, N pol is the number ofpolarizations observed, ∆ ν is the bandwidth, and ∆ t isthe integration time: ∆ S ν = SEFD η corr (cid:112) N ant ( N ant − N pol ∆ ν ∆ t . (6)We assume that η corr = 1 , all N ant = 64 antennaswill be operating, and N pol = 2 . The assumed inte-gration times are 333 hours for the L band and 3091hours for the UHF band, with 3424 hours for the fre-quency range where the receivers overlap. We also as-sume a velocity width of ∆ V =
150 km s − , the widthof the average OH line at z = 0 , converted to hertzusing ∆ ν/ ( ν OH / (1 + z )) = ∆ V /c . As LADUMA willbe applying Briggs weighting to obtain a well-behavedsynthesized beam, an extra noise penalty of 1.45 is in-cluded in sensitivity calculations. We also include theeffects of primary beam attenuation away from the phasecenter, assuming two-dimensional Gaussian beams withhalf-power points defined as . × . λ/D .Using the sensitivity at each redshift, we calculate the σ luminosity limit as L min = 4 πD L S ν ∆ ν where D L is the luminosity distance. Figure 1 shows how the lumi-nosity sensitivity changes with redshift for LADUMA.2.4. Dependence on the Major Merger Rate
An important element in the calculation of the num-ber of OHMs is the galaxy merger rate, since OHMsarise in merging galaxies. This consideration leads usto introduce a factor of (1 + z ) γ in equation (7), where γ is determined by the galaxy merger rate (in the senseof merging events per comoving volume) and will be re-ferred to as the merger rate evolution coefficient . Previ-ous studies, based on both observations and simulations,have used different conventions for defining and param-eterizing a “merger rate,” and drawn different conclu-sions that depend on type of merger, redshift, mass, andmany other factors (e.g., Lotz et al. 2011; Rodriguez-Gomez et al. 2015; Mundy et al. 2017; Mantha et al.2018; Duncan et al. 2019; O’Leary et al. 2020, and ref-erences therein). The relevant merger rate for OHMsis one corresponding to gas-rich major mergers, whoseevolution remains poorly constrained. Therefore, we se-lect an intermediate estimate for this merger rate evo-lution coefficient of γ ∼ . (Rodriguez-Gomez et al.2015; Mundy et al. 2017), which will be assumed whennot stated otherwise. In Section 2.5, we present OHMcalculations for a conservatively large range of possible γ values, . ≤ γ ≤ . , which is slightly larger thanthe range suggested by recent studies ( . (cid:46) γ (cid:46) . )(Mundy et al. 2017; Ferreira et al. 2020). H. Roberts, J. Darling, A. J. Baker
50 100 150 200
Number of OH Megamasers F r a c t i on o f T r i a l s Median ±1 Errors
Figure 2.
Kernel density estimation of the predicted num-ber of OHMs to be found by LADUMA, assuming γ = 2 . ,using 10,000 samples from the MCMC fit to the OHLF pa-rameters. The expected number of OHM detections is +21 − . OHMs present a unique and independent way to mea-sure the galaxy merger rate. Once we have a secure sam-ple of OHMs, we will be able to provide an estimated γ for gas-rich major mergers. As more H I surveys takedata, this method will be a robust way for tracing thecosmic history of major mergers.2.5. MCMC Calculation of N(OH)
Using the MCMC fit samples for the OHLF discussedin Section 2.1, we can integrate over volume, luminosity,and merger rate to get a prediction of the number of OHmegamasers to be detected at the 5 σ level in LADUMAas shown in equation (7): N = (cid:90) L max L min (cid:90) V total Φ( L OH ) (1 + z ) γ dV d log L OH . (7)Figure 2 shows a distribution of possible values for thenumber of OHMs using 10,000 samples from the MCMCfit to the OHLF, assuming that γ = 2 . . The medianvalue is 82.99, and the associated 16 th and 84 th per-centiles are 66.49 and 103.89. The 16 th and 84 th per-centiles are commonly associated with − σ and +1 σ limits (for a Gaussian distribution) respectively, andthey will be referred to as such for the remainder ofthis paper. The mean of the distribution is . . Thenumber of OH megamasers we expect to detect withLADUMA is therefore +21 − . This total would nearlydouble the number of known OHMs.As discussed in Section 2.4, the exact value of themerger rate evolution parameter γ is poorly constrained.Figure 3 presents the above calculation for values of γ ranging from 0.0 to 3.0. If γ is assumed to be 0.0, and Merger Rate Evolution ( ) N u m be r o f O H M s Figure 3.
Number of OHM detections in LADUMA versusmerger rate evolution parameter γ . The fiducial γ = 2 . isdenoted by a red X marker. 0 H U J H U 5 D W H ( Y R O X W L R Q 1 X P E H U R I 2 + 0 V 7 K L V : R U N ' D U O L Q J * L R Y D Q H O O L Figure 4.
Comparison of OHM detection rate in LAD-UMA versus merger rate evolution coefficient for the OHLFspresented in this work and in Darling & Giovanelli (2002b).The dashed and dotted lines show the median values, andthe shaded areas represent ± σ uncertainties. the merger rate does not increase with redshift, then itis expected that LADUMA would detect +5 − OHMs.We can compare the impact of the OHLF presented inthis work to the OHLF in Darling & Giovanelli (2002b)by calculating the numbers of OHMs implied by thetwo. Figure 4 shows how the number of OHMs varies forthe two OHLFs. When γ = 2 . , the MCMC approachadopted here implies a factor of 1.4 fewer detections. Ingeneral, the larger uncertainties from calculations withthe OHLF in Darling & Giovanelli (2002b) are due to thelarger uncertainties in that paper’s OHLF parameters. OH CONTAMINATION IN OTHER H I SURVEYSIn this section, we generalize the above calculationsfor H I surveys underway or planned at other radio tele- H Megamasers in H I Surveys Table 1. H I Survey Parameters for N (OH) CalculationSurvey Redshift Range Sky Area σ Sensitivity( z HI ) (deg ) (µJy)LADUMA 0.0–1.45 0.90 a b CHILES 0.0–0.45 0.32 a b SKA1 Medium deep 0.0–0.50 20 52 b SKA1 Deep 0.35–1.37 1 a b SKA1 All-sky 0.0–0.50 20,000 525 b a Single-pointing survey comoving volume calculations include an ex-panding field of view at higher redshifts and lower observing frequen-cies. b LADUMA and SKA1 have published frequency-dependent sensitiv-ities, which we employ for calculations in this paper. The valuespresented here are the mean sensitivities across the entire range ofthe observed frequencies.
References —CHILES (Fernández et al. 2013), WALLABY &DINGO-Deep (Duffy et al. 2012), APERTIF (Apertif Survey Plan),SKA1 (Braun et al. 2019; Staveley-Smith & Oosterloo 2014) scope arrays, particularly the Australian Square Kilo-metre Array Pathfinder (ASKAP), the Very Large Ar-ray (VLA), the APERture Tile In Focus (APERTIF)facility, and Phase I of the SKA. The VLA and ASKAPhave H I surveys underway or planned. At the VLA, theCOSMOS H I Large Extragalactic Survey (CHILES), asingle-pointing survey, is currently analyzing data (Fer-nández et al. 2016). WALLABY (Widefield ASKAP L-band Legacy All-sky blind surveY) is in the process ofobserving pilot fields with ASKAP alongside DINGO(Deep Investigation of Neutral Gas Origins; Duffy et al.2012). DINGO will have two tiers, Deep and UltraDeep. For our analysis, we consider only DINGO-Deep.APERTIF is also currently collecting data on the West-erbork Synthesis Radio Telescope (WSRT) and will ex-ecute multiple surveys at different depths. We considerthe Medium Deep Survey (MDS) as described in theAPERTIF Survey Plan. The SKA is a future telescopethat will come in two phases, with the first (SKA1) cov-ering ∼
10% of the total collecting area of the second(SKA2; Abdalla et al. 2015). Due to the uncertaintiesin the schedule for SKA2, we consider only possible mid-frequency SKA1 surveys for this analysis. The calculation of numbers of OHMs for these H I sur-veys uses equation (7) above. For each survey, we cal-culate L min as a function of redshift using the reportedsensitivity and we calculate the volume using redshiftranges and sky coverage. All assumptions made aboutthe survey or telescope for these calculations are pre-sented in Table 1. The sensitivity column assumes avelocity width of 150 km s − as done for the LADUMAcalculations. Concepts for the SKA1 surveys come fromStaveley-Smith & Oosterloo (2014). Each of the threefiducial surveys (medium wide, medium deep, and deep)assumes a total observing time of 1,000 hours, while theall-sky commensal survey assumes an observing time of10,000 hours. Staveley-Smith & Oosterloo (2014) notethat angular resolutions finer than 10" are only accessi-ble to the SKA for high column densities — a limitationthat especially applies to the SKA1 Deep survey, whosenominal angular resolution is 2". We have not modeledthe effects of resolution on the detectability of OHMs,but we note that at high resolutions, there could be abias in favor of detecting (more compact) OHMs relativeto (more extended) H I emitters. For surveys featuring single pointings, we have cal-culated comoving volumes assuming that sky area in-creases ∝ (1 + z H I ) due to the increasing size of theprimary beam at lower frequencies. This calculation ap-plies to CHILES, LADUMA, and a hypothetical SKA1Deep survey and is noted in Table 1. For surveys cover-ing larger sky areas through the use of multiple point-ings across contiguous patches, sky area will be higherat z H I > than at z H I = 0 , but the change will beless dramatic because only the pointings that lie at theedges of the contiguous patches will contribute. Becausethis effect will be small in a fractional sense (smaller forlarger sky areas), and will depend on the detailed distri-bution of patch sizes, we do not correct for it. We alsochoose to omit primary beam attenuation when predict-ing OHM contamination of other H I surveys, since wecite values for the numbers of H I detections that do notinclude this consideration (Staveley-Smith & Oosterloo2014).Table 2 presents the number of OHMs predicted tobe detected in each survey for merger rate evolution co-efficient γ = 2 . . Table 2 also presents the number ofH I sources each survey expects to detect. The contam-ination column is the ratio of OHM detections to H I This paper’s predictions ignore the effects of radio frequency in-terference, which can vary for different sites and different ar-ray configurations. Here too, it may in practice be systemati-cally easier to recover more compact OHMs than more extendedH I emitters in frequency ranges where RFI precludes the use ofshort-baseline data. H. Roberts, J. Darling, A. J. Baker
Table 2.
Predicted 5 σ OHM and H I Detections for Untargeted H I Surveys Survey N (OH) N (H I ) N (OH)/ N (H I )LADUMA . +2 . − . × × . +0 . − . % CHILES . +1 . − . × − × . +0 . − . % WALLABY . +2 . − . × × . +0 . − . % DINGO-Deep . +2 . − . × × . +0 . − . % APERTIF MDS . +0 . − . × × . +0 . − . % SKA1 Medium wide . +0 . − . × . × . +0 . − . % SKA1 Medium deep . +3 . − . × . × . +0 . − . % SKA1 Deep . +0 . − . × . × . +2 . − . % SKA1 All-sky . +1 . − . × . × . +0 . − . % Note — N (OH) values assume merger rate evolution coefficient γ = 2 . . detections, which can be related to the fraction of an“H I sample” that will actually be OHMs mistaken forH I sources if spectroscopic redshifts are unavailable.One noteworthy aspect from Table 2 is the muchhigher rate of contamination for LADUMA and theSKA1 Deep survey compared to the other surveys.These are distinctly different from the other H I surveysdue the fact that they extend to significantly higher red-shift. We therefore infer that H I detections dominateat low redshifts (i.e., z HI (cid:46) ) for all surveys. However,the OH detection density surpasses the H I detectiondensity at redshifts above z HI ∼ — an effect that ispronounced for single-pointing surveys that have muchlarger relative fields of view at high versus low redshifts.An earlier conclusion in the same vein was reached byBriggs (1998), who predicts that OHM contaminationin H I surveys will increase with redshift. We explorehow that contamination depends on redshift using bothLADUMA and an expanded version of the SKA1 Deepsurvey for comparison. The SKA1 Deep survey pre-sented in Staveley-Smith & Oosterloo (2014) only coversa frequency range of 600–1050 MHz that corresponds toonly a portion of the full SKA1 mid-band (Braun et al.2019). For the purpose of exploring OH contaminationversus redshift, we assume a deep survey that exploitsthe full range of the mid-band and therefore covers a fre-quency range of 600–1420 MHz. Equation (7) is used toestimate how the number of OHMs varies with redshift.To estimate the number of H I detections per redshift in-terval, we use the following equation from Obreschkowet al. (2009): dN/dz = 10 c z c e − c z , (8) where the c i are parameters specific to each H I survey.LADUMA’s values are interpolated from Obreschkowet al. (2009) for each redshift using sensitivities calcu-lated from equation (6) as the limiting integrated flux,and assuming linewidths of 100 km s − , allowing us todetermine H I detection rate versus redshift. We follow asimilar method for the SKA1 Deep survey using sensitiv-ity values presented in Braun et al. (2019). We calculatehow number density for OH and H I varies with redshift,assuming redshift bins of dz = 0 . . Obreschkow et al.(2009) note that dN/dz will be ≥ underestimatedfor z ≤ z c where z c depends on the limiting flux. Forboth LADUMA and SKA1 Deep, on average, z c ∼ . .Therefore, the numbers of H I detections are slightly un-derestimated for z HI ≤ . .Figure 5 shows how the detection rate of OH and H I varies with redshift. H I detections dominate at low red-shift for both surveys. At higher redshifts, the OH de-tection rate and OH fraction grow significantly. Thiscomparison also demonstrates how the OHM contami-nation rate depends on survey parameters, as discussedin Section 3. For LADUMA, OHMs will not outnumberH I source at any redshift probed by the survey (i.e.,for any z HI ≤ . ); in comparison, SKA1 Deep’s sen-sitivity as a function of frequency will yield a numberof OH detections surpassing that of H I detections for z HI ≥ . .Briggs (1998) first presented this issue of OH contam-ination in untargeted H I surveys. Results from thatpaper demonstrated that by redshift z HI ∼ I detectionsfor a much shallower survey ( σ = 5 mJy) than LAD-UMA or SKA1 Deep. The predicted transition from anH I -dominated to an OH-dominated sample is qualita-tively consistent with our findings, although the preciseredshift at which this happens depends on survey pa-rameters. In general, the deeper a survey observes overa given frequency range, the more H I emitters it willdetect relative to OHMs. LADUMA and SKA1 Deepare both deeper than the hypothetical Briggs (1998)survey, allowing them to detect more H I sources athigher redshifts and thus to push out the projected red-shift at which OH detections outnumber H I detections.The fact that this transition occurs at a lower redshift( z H I = 1 . ) for SKA1 than for LADUMA owes to thefact that LADUMA’s sensitivity improves at higher red-shifts (due to its distribution of observing time), in con-trast to the SKA1 Deep sensitivity. H Megamasers in H I Surveys N O H / N H I dN OH / dz dN HI / dz N OH / N HI Redshift ( z HI ) N u m be r LADUMA N O H / N H I Redshift ( z HI ) N u m be r SKA1 Deep
Figure 5.
Projected numbers of OHMs and H I sources andthe OHM fraction (N OH / N HI ) vs redshift z HI for LADUMA(above) and SKA1 Deep (below). Left axis shows numbersof objects (plotted in black); the dashed curve indicates howthe number of OH detections changes with redshift, and thesolid curve indicates the same for H I detections. Right axisshows the ratio of OHM detections to H I detections (plot-ted in red). The vertical dotted lines indicate the redshiftswhere the numbers of OH and H I detections are equal. Thediscontinuities in the LADUMA curves originate from theoverlap between the L and UHF bands. All calculations as-sume galaxy merger rate evolution parameter γ = 2 . .4. IDENTIFYING OH MEGAMASERS INUNTARGETED H I SURVEYSDistinguishing an OH from an H I line is currentlyonly done using the optical spectroscopic redshift of anobject to determine an observed line’s rest wavelength.Next-generation H I surveys will observe orders of mag-nitude more objects than previous surveys, as shown inTable 3, most of which will not have spectroscopic red-shifts available. For that reason, we explore machinelearning as a way to distinguish OH from H I emissionlines using ancillary data. 4.1. Creating OH and H I Models for DistinguishingPopulations
Fitting OHM Host Galaxy SEDs
The limited number of OHMs creates serious limi-tations in understanding the OH population and howit differs from H I hosts. We therefore fit the spec-tral energy distributions (SEDs) of 111 OHMs usingMulti-wavelength Analysis of Galaxy Physical Proper-ties (MAGPHYS; Da Cunha et al. 2008), a softwarepackage that fits galaxy SEDs using physical parame-ters of galaxies at the same redshifts and in the samephotometric bands.MAGPHYS fits SEDs from far-UV to far-IR, so weuse photometry from that range for fitting OHM hostSEDs. In total, we use photometry from eight sources:the Galaxy Evolution Explorer ( GALEX ; Martin et al.2005), the Sloan Digital Sky Survey (SDSS; Stoughtonet al. 2002), the Two Micron All-Sky Survey (2MASS;Skrutskie et al. 2006), the
Wide-field Infrared Survey Ex-plorer ( WISE ; Wright et al. 2010), the Infrared ArrayCamera (IRAC) and Multiband Infrared Photometer for
Spitzer (MIPS) (both on
Spitzer ; Werner 2005), the
In-frared Space Observatory ( ISO ; Kessler & Habing 1996),and the
Infrared Astronomical Satellite ( IRAS ; Beich-man et al. 1988). In total, from these sources, we useup to 33 bands to fit OHM SEDs. We omit
WISE band1 (3.4 µm) if
Spitzer
IRAC band 1 (3.6 µm) exists fora given galaxy because the introduction of both causespoor fits, and IRAC tends to have smaller uncertaintiesthan
WISE . The same is done for
WISE and IRAC band2 (4.6 and 4.5 µm, respectively).In Figure 6, we present examples of these SED fits. Forcomparison, we use the Atlas of Galaxy SEDs (Brownet al. 2014), which includes a total of seven OHM hostgalaxy SEDs. The fits we present are imperfect matchesto the complete Brown et al. (2014) SEDs, and we em-phasize that our SED fits were done with the limitedscope of reproducing observations of OHMs in our par-ticular wavelength regime of interest, UV to mid-IR. Ourfits are limited in wavelength outside this range, partic-ularly in the far-IR. Although far-IR photometry wouldprovide very useful information, data are sparse and un-available for our objects of interest, so we have chosento omit far-IR photometry.
H. Roberts, J. Darling, A. J. Baker Wavelength ( m) F l u x ( e r g / s / c m / m ) NGC 3690
Published (Brown et al. 2014)SED Fit (This Work)Photometry ( N = 0.97) 10 Wavelength ( m) F l u x ( e r g / s / c m / m ) Arp 220
Published (Brown et al. 2014)SED Fit (This Work)Photometry ( N = 0.63) Figure 6.
Examples of SED fits with MAGPHYS (black line). The published SED comes from Brown et al. (2014) and isshown as a thick grey line. Red dots denote the photometry used in a given fit.
Emulating OHM Host Photometry
We use our SED fits to model OHM host galaxy ob-servations for various missions and surveys as a functionof redshift. We use PYPHOT, a package for calculat-ing an object’s photometry from its SED. It calculatesthe photometry using a given filter’s transmission curve, T ( λ ) , by calculating the photon number flux: N tot = 1 hc (cid:90) λ f λ λ T ( λ ) dλ, (9)where f λ is the flux density as shown above in Figure 6.The greatest benefit of having SED fits is the abil-ity to redshift them and mimic observations at higherredshifts. This scaling is done by adjusting the restwavelength and the flux density by the inverse squareof the luminosity distance and a redshift factor ( ∝ D − L (1 + z ) − ) and then re-“observing” the OHM host.This method is used to create synthetic observations outto the desired redshifts.4.1.3. Emulating H I Host Galaxy Photometry
For consistency, H I host galaxy photometry is createdsimilarly. However, instead of fitting SEDs, we use 57SEDs published in the Atlas of Galaxy SEDs (Brownet al. 2014) that have previous H I detections, a pop-ulation of mainly spiral galaxies. Since these sourcesare not drawn from a strictly H I -selected sample, they PYPHOT’s documentation can be accessed at https://mfouesneau.github.io/docs/pyphot/. This number was originally 58 SEDs; however, after some exam-ination, NGC 7674 seems to behave much more like an OHM inthe near- to mid-IR like just a spiral galaxy, despite looking likea classic spiral morphologically. We attempted to do follow-upobservations to determine if it potentially possessed both emis-sion lines; however the OH line cannot be observed due to RFI.We therefore removed this galaxy from our H I SED sample. may not behave identically to samples from untargetedH I surveys, although we expect differences to be mod-est. These SEDs are also redshifted, and photometry is“measured” using PYPHOT.4.2. Machine Learning to Distinguish OH from H I To aid in determining if an emission line is an H I orOH detection, we use machine learning algorithms to de-termine the likelihood of the line’s classification. We em-ploy a k -Nearest Neighbors ( k -NN) algorithm that clas-sifies objects based on a plurality vote of their neighbors’classes, where neighbors are determined within some pa-rameter space (Goldberger et al. 2005). k -NN classifi-cation is a non-parametric method and a lazy learning algorithm. Lazy learning means that the algorithm it-self does not make assumptions or generalizations basedon the training data, but instead uses those data tomake direct decisions about the testing data. Algorithm“optimization” is purely done by our choices about thenearest-neighbor algorithm parameters used. The costof using a lazy learning algorithm is the computationtime in the testing phase. However, we are not testingon data sets large enough for slow speed to be prob-lematic. This context makes the k -NN classification arobust and transparent method for our purposes.The final classification parameters are WISE magni-tudes and colors as well as the observed line frequency.Suess et al. (2016) demonstrate that
WISE photometrycan separate OHM and H I populations at low redshift.We also choose WISE because of its all-sky coverage,allowing it to be applicable to many different H I sur-veys. Section 4.2.1 discusses the use of WISE magni-tudes and colors to distinguish H I and OH populations,as well as the limitations of using WISE and its simi-larities to IRAC. Section 4.2.2 presents similar exercisesusing IRAC data, which have significant coverage over
H Megamasers in H I Surveys W1-W2 W OHHI 0.00.51.01.5 R ed s h i ft ( z H I ) Figure 7.
Predicted
WISE magnitude versus color for anOHM host galaxy (stars) and an H I galaxy (circles). H I redshift is denoted by color with redshift increasing withlightness. The corresponding OH points lie at the same ob-served frequency but are at a higher actual redshift. Thegrey dashed line represents the detection limit of WISE (theregion above the line is undetectable by
WISE band W1). the LADUMA field but are otherwise less broadly ap-plicable for other H I surveys.4.2.1. OH and H I Classification Using WISE
The analysis in Suess et al. (2016) is done with low-redshift ( z < . ) objects and uses WISE bands W1,W2, W3, and W4 (3.4, 4.6, 12, 22 µm). W3 and W4are very insensitive compared to W1 and W2; thus, thismethod is limited by both object brightness and red-shift. We focus on using machine learning to sort usingonly W1 and W2 magnitudes, W1 − W2 color, and theobserved line frequency.One of the cuts from Suess et al. (2016) is done in thecolor-magnitude space of W1 versus W1 − W2 (or [3.4]versus [3.4] − [4.6]). We use this same parameter spacefor the k -NN algorithm. Examples of redshift evolutionin this space are shown in Figure 7 for an OHM hostand an H I source. These data are “measured” from theirSEDs and show how OH-H I separability varies with red-shift.We use the 57 H I and 111 OHM host SEDs to testand train the k -NN algorithm. Each SED is redshifted tothe maximum redshift detectable by LADUMA ( z HI =1 . and z OH = 1 . ), with WISE photometry being“measured” roughly every dz = 0 . . We then removeany data too faint to have a σ WISE detection. k -NN algorithms require feature scaling or parameternormalization, since the algorithm is inherently built onthe distances between a data point and its neighbors.Therefore, we normalize each parameter from 0 to 1. k -NN classification algorithms are dependent on a fewparameters that can be optimized for a given case. Pa-rameters that were varied and tested for our purposesinclude the number of neighbors that is included in theplurality vote on an object’s classification ( k ), whetherneighbors are weighted by their distances, and how dis-tance between objects is calculated ( p ). The distancebetween points is defined by the Minkowski distance oforder p : D ( X, Y ) = (cid:16) n (cid:88) i =1 | x i − y i | p (cid:17) p , (10)where X and Y are two points in an n -dimensional pa-rameter space. Euclidean distance is recovered for p = 2 .All algorithm optimization for this work is done bymaximizing OH recall . In machine learning classifica-tion, two metrics that are often considered when opti-mizing are precision and recall , both of which scale from0 to 1 (1 being the best score). Precision is the fractionof positive identifications that are correct. By optimiz-ing precision, the number of false positives (or Type Ierrors) is minimized.
Recall , conversely, is the fractionof positives that were correctly identified. When recallis optimized, the number of false negatives (or Type IIerrors) is minimized (Sammut & Webb 2011). Theseterms correspond to the familiar astronomical conceptsof sample purity and completeness. In our case, a posi-tive identification is the classification of a galaxy as anOHM host. We choose to optimize OH recall due to therarity of OHMs and the desire to not miss any poten-tial candidates. Although this approach increases thenumber of false positives, this algorithm does add in-formation, and any positive identification it makes canmotivate follow-up observations for confirmation.The algorithm parameter exploration is shown in Fig-ure 8. The x -axis shows a wide range of choices forthe number of neighbors used, the lines plotted show afew choices for Minkowski distance, and the two panelsshow the difference between weighting and not weight-ing neighbors by distance. Each unique combinationof parameters is tested using a five-fold cross-validationtest. This process involves randomly sorting our datainto two sets: training and testing data. The trainingdata build the algorithm and the testing data determinehow successful the algorithm is. This split was donefive times, randomly splitting data each time, for eachcombination of parameters, and the final OH recall wasdetermined by averaging the five individual OH recallvalues. In total, 2,000 k -NN algorithms were tested.The results of these parameter tests give the highestOH recall for large numbers of neighbors ( k > ).Although an OH recall of 1 would be ideal, this result0 H. Roberts, J. Darling, A. J. Baker k neighbors O H r e c a ll Uniform Weighting k neighbors Weighting by Distance p = 1.0 p = 2.0 p = 3.0 Figure 8.
Parameter exploration for our k -Nearest Neighbors ( k -NN) classification algorithm. We vary the number ofneighbors considered in voting ( x -axis), whether distance is weighted (left and right plots), and how distance is determinedusing the Minkowski metric (the lines plotted). Each test is done using five-fold cross-validation. comes at the cost of very low precision and defeats thepurpose of the k -NN method by classifying based onthe value of an algorithmic parameter instead of po-sition relative to neighbors. Large numbers of neigh-bors also make for very computationally expensive al-gorithms. One common approach is to select k = √ N ,where N is the number of data points, but for our data( k ≈ ), this choice of k is near the lowest value ofOH recall. Another common approach is to select k = 1 or another low number. For small data sets, always as-suming the nearest object has the same classificationcan introduce noise. However, with a sufficiently largedata set, this trend is less problematic. Figure 8 indi-cates that a small k achieves a recall of over 0.98 fordistance-weighted learning. We stress that there existsno optimal k for all purposes, since each k -NN optimiza-tion varies based on the properties of the data (Altman1992).We choose our number of neighbors to be k = 3 , basedon the above considerations. Increasing a small amountabove k = 1 also reduces noise while maximizing OHrecall. For weighting and Minkowski metric, we chooseto weigh votes by distance and use standard Euclideandistance ( p = 2 ).The final trained and tested k -NN algorithm resultsare shown in the top panel of Figure 9. The blue andred points show correctly identified OH and H I hostgalaxies respectively, and the black stars are misiden-tified objects. The incorrect identifications concentratewhere OH and H I sources overlap the most at the same observed frequency, and indicate where there will be themost confusion. The final OH recall is and theOH precision is . In other words, for redshifts lessthan z ∼ . , we expect to identify 98.5% of OH lines inH I surveys, thereby mitigating the impact of contami-nation.We repeat this process for another Suess et al. (2016)parameter space, W1 − W2 versus W2 − W3, as well asthe observed line frequency. This alternative approachsignificantly limits the number of available detectionsbecause of the inclusion of the comparatively less sensi-tive W3 band. This approach however leads to a higherOH precision, since mid-IR data are relevant for distin-guishing between these populations. The results of thistest are shown in the bottom panel of Figure 9. Preci-sion and recall from this test are compared to those forother tests in Table 3.4.2.2.
OH and H I Classification Using IRAC
As discussed previously,
WISE is insensitive to galax-ies at redshifts above z ∼ . . Although having sig-nificantly less sky coverage than WISE , Spitzer ’s IRACbands 1 and 2 are very similar to
WISE bands 1 and 2but are much more sensitive and can detect OHM andH I host galaxies over the full redshift ranges probed byboth LADUMA and SKA1. We therefore perform an ex-ercise similar to that in Section 4.2.1 using IRAC data.Throughout this paper, IRAC bands are referred to bytheir wavelengths in microns (e.g., IRAC [3.6] denotesthe 3.6 µ m band magnitude). H Megamasers in H I Surveys W1-W2 W OHHIIncorrectClassification
W2-W3 W - W Figure 9.
Final results from training and testing our k -NN algorithm using WISE
W1 versus W1 − W2 (above)and W1 − W2 versus W2 − W3 (below). Blue points indicateOH host testing points that were correctly identified, andred indicate the same for H I sources. Black stars show themisidentified objects, either OH misidentified as H I or viceversa (3.6% of objects in top panel, 1.9% of objects in bottompanel). Note that the OH and H I markers are partiallytransparent to show overlapping. We use a parameter space analogous to the first testfor the IRAC k -NN algorithm ([3.6] versus [3.6] − [4.5]).(As a reminder, WISE uses Vega-based magnitudes,whereas IRAC uses AB magnitudes.) Since IRAC is sen-sitive to the entire redshift range of our H I and OHM Table 3.
Machine Learning Results for Distinguishing H I Emis-sion Lines from OH MegamasersMission Features OH Recall OH Precision
WISE
W1, W1 − W2, ν WISE W1 − W2, W2 − W3, ν − [4.5], ν − [4.5], [5.8] − [8.0], ν Note —The features column indicates what data were used to dis-tinguish H I from OHM host galaxies. Each row includes the ob-served frequency ( ν ) of the line in question (OH or H I ) to assistin separation of sources. hosts, we do not perform any detection cuts. Resultsfrom this exercise are presented in Table 3 and visual-ized in Figure 10. Comparing to the analogous WISE space, this test has the same OH precision, but OH re-call suffers slightly. However, achieving an OH recall of0.979 is still a powerful tool when it comes to sortingOHM hosts from H I hosts, and the ability to probe tohigher redshifts has strong appeal.We also consider IRAC [3.6] − [4.5] versus [5.8] − [8.0].Stern et al. (2005) suggest that cuts in this space canseparate active galaxies from normal galaxies. We per-form algorithm optimizations similar to those mentionedpreviously before training and testing. We present thefinal results in Table 3 and Figure 10. Despite hav-ing information from the [8.0] band, this test performsslightly worse than the previous tests in both OH recalland precision, indicating that the overlap between OHMand H I hosts is greater in this parameter space than inthe previous alternatives.Being able to probe the full redshift range of LAD-UMA is beneficial, but inevitably introduces more con-tamination, as indicated by the slightly reduced OHrecall. However, it is worth noting that despite beinglower, these recall values still exceed 95%. These IRACand WISE tests create a new framework for the processof separating OH and H I host populations. DISCUSSIONThe methods presented in this paper will be crucialto mitigating OHM contamination of H I emission-linesurveys. WISE provides the all-sky coverage needed forupcoming surveys that will be covering large portions ofthe sky, while IRAC has the deep-field coverage neededfor surveys such as LADUMA, which will be the deepestH I emission-line survey to date.One of the biggest shortcomings of these methods andcalculations is that they are based on known OHMs,2 H. Roberts, J. Darling, A. J. Baker [3.6]-[4.5] [ . ] OHHIIncorrectClassification [5.8]-[8.0] [ . ] - [ . ] Figure 10.
Final results from training and testing our k -NN algorithm using Spitzer
IRAC [3.6] versus [3.6] − [4.5](above) and [3.6] − [4.5] versus [5.8] − [8.0] (below). Bluepoints indicate OH host testing points that were correctlyidentified, and red indicate the same for H I sources. Blackstars show misidentified objects (3.4% of objects in top panel,2.9% of objects in bottom panel). Note that the OH and H I markers are partially transparent to show overlapping. which currently extend to a highest redshift of z OH =0 . (Darling & Giovanelli 2002a). This limitation hasforced us to make some extrapolations to obtain predic-tions for higher-redshift surveys. This approach is un-avoidable until we have higher-redshift data on both H I and OH populations. As more surveys are conducted, we will be able to update the OHLF and OH SED evolutionas well as provide tighter constraints on these calcula-tions and predictions.As discussed in Section 1, H I and OH sources canbe separated by spectroscopic redshift. It is thereforeworth recognizing that some objects will be readily iden-tifiable and that these objects will be crucial for help-ing classify those without redshifts. For LADUMA,the current largest source of spectroscopic redshifts isthe PRIsm MUlti-Object Survey (PRIMUS), with over32,000 redshifts in the field and in the relevant red-shift range (Coil et al. 2011). PRIMUS only detectsgalaxies out to z ∼ . , meaning that some of the mostpotentially contaminated (i.e., highest) redshift rangeswill have few spectroscopic redshifts available. Anothersource of spectroscopic redshifts soon to come online isthe Wide-Area VISTA Extragalactic Survey (WAVES),which will have two campaigns, WAVES-Wide (large-sky, low-redshift) and WAVES-Deep (small-area, high-redshift) (Driver et al. 2019). WAVES-Deep will haveseveral small patches, including one on the LADUMAfield. Slated to target 45,000 objects, WAVES-Deepwill also be crucial in identifying objects; however thecurrent estimates only show detections out to redshift z ∼ . . For future all-sky untargeted H I surveys suchas those on ASKAP or the SKA, WAVES-Wide aims toprovide 880,000 spectroscopic redshifts out to redshift z ∼ . . Although we may have many redshifts for iden-tifying objects as OH or H I sources, these redshifts areextremely limited where potential OH contamination isthe greatest threat. WISE and IRAC photometry were not the only datatested in Section 4 for the ability to separate OHM andH I hosts. We also tested other photometry for sepa-rability, focusing on data that have significant coveragein the LADUMA field. These include SDSS ugr , John-son UBV , and
HST
ACS, WFC3, and NICMOS bands.Figure 11 shows how each of these bands correlates withOH/H I classification. For each band, a Pearson corre-lation test was done for three redshift ranges. Bandsare grouped on the x -axis and then sorted by increasingwavelength. Figure 11 demonstrates the sorting value ofbands in the near- to mid-IR. This distinction is due tothe extreme star formation in OHM host galaxies, whichis detected in the IR. Optical bands are poorer candi-dates for separation, since they are less sensitive to highstar-formation rates in dusty, gas-rich systems. CONCLUSIONSWe present predictions for the numbers of OH mega-masers that will be detected in future untargeted H I surveys and explore how those numbers impact H I H Megamasers in H I Surveys J ohn s on U J ohn s on B J ohn s on V S D SS u S D SS g S D SS r A C S F W A C S F W A C S F W A C S F W A C S F P W F C F M W F C F WW F C F WW F C F W N I C F W N I C F W I R A C [ . ]I R A C [ . ]I R A C [ . ]I R A C [ . ] W I SE W W I SE W W I SE W W I SE W C o rr e l a t i on w i t h C l a ss i f i c a t i on z HI < 0.50.5 < z HI < 1.01.0 < z HI <1.5 Figure 11.
Correlation of band photometry with OH/H I classification using a Pearson correlation test. Bands on the x -axisare grouped by mission or type and then ordered by increasing wavelength. Each correlation is tested in three redshift ranges. source confusion over a range of redshifts up to z HI =1 . . To assist in untangling these populations, we alsopresent methods for estimating the likelihood that a linehas been identified as H I or OH. Below, we summarizeour predictions and discuss the implications of this workfor future H I surveys:1. LADUMA will likely triple the number of knownOHMs: we predict +21 − new detections. Largersurveys with telescopes such as the SKA1 will de-tect thousands more OHMs.2. The contamination these OHM detections willimpose on H I line surveys is highly dependenton redshift (and, secondarily, depth). In a lineflux-limited survey, OHMs are more abundant athigher redshift, while H I sources become sparser.For these high-redshift surveys, OH detections willoutnumber H I detections near redshift z HI ∼ . .3. Near- and mid-IR observations can assist in sep-arating H I from OHM emission lines, which wedemonstrate using a k -Nearest Neighbors machinelearning algorithm. We will be able to identifynearly 99% of OH lines for redshifts less than z ∼ . and 96% of lines at higher redshifts.Although OHM host galaxies represent a potentialcontamination for untargeted H I line surveys, these rareand interesting objects can be important scientific tools.As discussed in Section 2.4, OHM density can also pro-vide an independent measurement on the major mergerrate evolution paramater, γ , since OHMs serve as tracers of major galaxy mergers. These galaxies are signpostsof the most extreme star formation in our universe, sig-naling where the most massive starbursts are happening(Briggs 1998), and can even offer a way to measure in-situ magnetic fields using Zeeman splitting (Robishawet al. 2008; McBride et al. 2014). The methods presentedin this paper and follow-up observations will begin un-covering these galaxies and allowing us to characterizethem at higher redshifts and potentially create bettermethods for mitigating contamination in H I surveys.ACKNOWLEDGMENTSThis work has been supported by the National ScienceFoundation through grants AST-1814648 (to HR andJD) and AST-1814421 (to AB). This research has madeuse of the NASA/IPAC Extragalactic Database (NED),which is funded by the National Aeronautics and SpaceAdministration and operated by the California Insti-tute of Technology. We thank Sarah Blyth, NatashaMaddox, and Aaron Stemo for helpful insights and con-versations. We also thank the anonymous referee fortheir thorough and helpful comments which improvedthe presentation of this paper. Software:
Astropy (Robitaille et al. 2013), Mat-plotlib (Hunter 2007), Numpy (Van Der Walt et al.2011), Scipy (Virtanen et al. 2020), Scikit-Learn (Pe-dregosa et al. 2011), MAGPHYS (Da Cunha et al.2008), emcee (Foreman-Mackey et al. 2013), PYPHOT(https://github.com/mfouesneau/pyphot)4
H. Roberts, J. Darling, A. J. Baker