HUBBLE IMAGING OF THE IONIZING RADIATION FROM A STAR-FORMING GALAXY AT Z = 3.2 WITH ${f}_{\mathrm{esc}}\gt 50 \%$ ^*

Cosmic reionization is a major episode in the history of the universe and the search for ionizing sources is one of the main goals of modern observational cosmology (Robertson et al. 2010). Star-forming galaxies and active galactic nuclei (AGNs) have been proposed to be the dominant sources of ionizing radiation, possibly active at different cosmic epochs (Haardt & Madau 2012). While the redshift evolution of the ultraviolet luminosity function of star-forming galaxies is relatively well measured up to $z\simeq 7\mbox{--}8$, showing that the bulk of the ultraviolet luminosity density is dominated by the faint galaxy population ($L\lt {L}^{\star }$, e.g., Bouwens et al. 2015), the number density of high-redshift and faint AGNs is still highly uncertain (e.g., Georgakakis et al. 2015; Giallongo et al. 2015). In addition, understanding the mechanisms of reionization (and post-reionization $z\lt 6$) hinges on assessing how the escape fraction of ionizing radiation, ${f}_{\mathrm{esc}}(\mathrm{LyC})$, changes as a function of luminosity and redshift. Because of the intergalactic medium (IGM) opacity, direct observation of ionizing radiation during reionization is not feasible (Prochaska et al. 2010). A strategy to make progress is to identify LyC sources at lower redshift, e.g., $z\sim 3\mbox{--}3.5$, and study which of their observed properties can be used as predictors of an LyC leakage. Recent advances have been made by looking at starburst galaxies in the local universe (Borthakur et al. 2014; Izotov et al. 2016) at about 10 Gyr after reionization ended ($z\sim 6$). However, it is more useful to identify LyC leakers at the highest redshifts possible ($z\sim 3$, one billion years after the end of reionization) because these galaxies are likely better analogs to those that reionized the universe. For instance, LyC emitters identified in the nearby universe exhibit redder UV slopes than $z\sim 6$ galaxies (Borthakur et al. 2014; Izotov et al. 2016).

Escaping LyC radiation from galaxies has been searched in recent years (Siana et al. 2010; Vanzella et al. 2010b; Mostardi et al. 2013, 2015; Nestor et al. 2013) and, to date, no spectroscopically confirmed detection of LyC has been reported at high redshift (Siana et al. 2015). Moreover, nothing is known about the spatial distribution of the emerging LyC radiation. The difficulty of directly and unambiguously identifying LyC radiation in distant galaxies is due to a combination of several effects: (1) superposition of foreground sources can produce false LyC detections (Vanzella et al. 2010b, 2012) and, in this respect, Hubble Space Telescope (HST) observations are crucial (Mostardi et al. 2015; Siana et al. 2015); (2) the intergalactic transmission in the ionizing domain is stochastic (Inoue et al. 2014); (3) the geometrical distribution of the neutral gas in galaxies and the relatively short duty cycle of ${f}_{\mathrm{esc}}(\mathrm{LyC})$ over cosmic time adds further stochasticity to the LyC visibility (Wise et al. 2014; Cen & Kimm 2015); (4) given the current sensitivity limits of large telescopes, it is only possible to probe an interesting dynamic range of LyC/non-ionizing UV ratios in galaxies with $L\gt 0.5{L}^{\star }$, a category for which the escape fraction of ionizing radiation has been shown to be intrinsically low ($\lt 5 \% \mbox{--}10 \%$, Vanzella et al. 2010a; Mostardi et al. 2015; Siana et al. 2015; Grazian et al. 2016). These aspects make the search for escaping ionizing radiation challenging and can explain the current low detection rate at high redshift. Nevertheless, the identification of examples with escaping LyC radiation lying in the tail of the large ${f}_{\mathrm{esc}}(\mathrm{LyC})$ values, though rare (Vanzella et al. 2010a), represents the only empirical method we have to increase our physical insight into the mechanisms that allow ionizing photons to escape, thus providing unique reference for studies at $z\gt 6$.

Here, we report on LyC emission arising from a distant galaxy (z = 3.212) unambiguously confirmed with HST observations. Throughout this paper, the AB magnitude system ($\mathrm{AB}=31.4-2.5\mathrm{log}({f}_{\nu }/{nJy}$) and a cosmology of ${{\rm{\Omega }}}_{\mathrm{tot}},{{\rm{\Omega }}}_{M},{{\rm{\Omega }}}_{{\rm{\Lambda }}}=1.0,0.3,0.7$ with ${H}_{0}=70$ km s⁻¹ Mpc⁻¹ are used.

An LyC candidate at z = 3.212 in the GOODS-Southern field, named Ion2 (GDS-ID 033203.24-274518.8), was first identified by Vanzella et al. (2015). This candidate is a $\sim {10}^{9}{M}_{\odot }$ galaxy with SFR $=\;15.6\pm 1.5\;{M}_{\odot }\;{{\rm{yr}}}^{-1}$ (sSFR$\;\sim \;10\;{\mathrm{Gyr}}^{-1}$), EW(Lyα)$\;=\;94\pm 20$ Å, and EW($[{\rm{C}}\;{\rm{III}}]\lambda 1909$)$\;=\;{18}_{-5}^{+9}$ Å (de Barros et al. 2016). The compact star-forming region, showing strong [O iii]λλ4959,5007 emission lines (with a rest-frame equivalent width of 1500 Å) and a large Oxygen ratio $[{\rm{O}}\;{\rm{III}}]\lambda 5007$/$[{\rm{O}}\;{\rm{II}}]\lambda 3727$ ($\geqslant 10$ at 1σ, de Barros et al. 2016) measured with Keck/MOSFIRE, makes Ion2 the highest redshift "Green Pea" galaxy the currently known and, accordingly to the photoionization models (Jaskot & Oey 2013; Nakajima & Ouchi 2014), an ideal candidate LyC emitter. A plausible spectroscopic LyC detection was subsequently discussed in de Barros et al. (2016). However, the presence of a close companion not resolved with ground-based spectroscopy and imaging (02, see Figure 1) cast some doubts on the association of the observed flux with Ion2,and thus on the reliability of the LyC leakage. Thanks to the Lyβ and Lyγ detection in the Very Large Telescope (VLT)/VIMOS Ion2 spectrum (see Figure 6 panel (A), or de Barros et al. 2016, Figure 2), the spectroscopic redshift of component A is unambiguously z = 3.2. The 4 hr exposure is not enough to detect such absorption features with a signal-to-noise ratio (S/N) of ${\rm{S}}/{\rm{N}}\gt 5$ (as observed here) for component B, which at these wavelengths is mag = 27.2 in the continuum.¹¹ Therefore these features arise from component A at z = 3.2. The [O iii] emission is also mainly associated with component A because the K-band magnitude is dominated by the strong [O iii] lines and the K-band spatial emission is more prominent from A. This is shown in Figure 2 where the VLT/HAWKI K-band image has been extracted from the HUGS survey (Fontana et al. 2014).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

To confirm the LyC emission, we obtained HST WFC3/UVIS images in the F336W filter, which corresponds to rest-frame $730\mbox{--}890\;\mathring{\rm A}$. Seventeen 2800 s (1-orbit) dithered exposures were taken for a total integration time of 47.6 ks. The charge transfer efficiency (CTE) of the WFC3/UVIS CCDs has degraded significantly due to radiation damage and this was particularly problematic in images with low background. To mitigate the effects of poor CTE, we increased the background to $\sim 12{e}^{-}$ pix⁻¹ with a post-flash LED (Biretta & Baggett 2013). Furthermore, we placed the target near the read-out edge of the CCD so that the electron transfer occurred across only ∼300 of the 2048 pixels. Finally, we applied a pixel-based CTE correction.¹² As the dark current was nearly half of the total background, proper dark subtraction was critical. Unfortunately, the STScI darks were only a single value and did not allow us to capture the gradient and blotchy pattern in the dark current (Teplitz et al. 2013). Also, because the darks were not CTE-corrected, more than half of the hot and warm pixels were not properly masked (Rafelski et al. 2015). To mitigate these issues, we adopted the dark processing method explained in detail in Rafelski et al. (2015). First, we CTE-corrected all of the raw dark images in the anneal cycle of our visits and removed cosmic rays. We then found and masked the hot pixels that appeared throughout the anneal cycle and made a mean super dark from all of the darks in the anneal cycle. We then dark subtracted the science images while masking the hot pixels that were found to exist at the time of the observation.

We processed all of our CTE-corrected raw data using STSDAS task CALWF3, including subtraction of our new super darks. These calibrated images were then combined using Astrodrizzle (Gonzaga et al. 2012), which performs background subtraction, cosmic-ray rejection, and geometric distortion correction. The final combined F336W image had a pixel scale of 003 and was astrometrically aligned with the three-dimensional-HST (3D-HST) F606W image (Skelton et al. 2014) with a precision of 48 mas. As a by-product, Astrodrizzle produced an inverse variance image that we used to derive the uncertainties in the photometry. We performed aperture photometry using a growth-curve methodology. The estimated uncertainty was corrected for the correlated noise introduced by the drizzling process (Casertano et al. 2000). Finally, we corrected our photometry for Galactic extinction (0.041 mag; Schlafly & Finkbeiner 2011).

The spectroscopic detection of the LyC emission reported by de Barros et al. (2016) is uncertain because of the presence of a secondary component (component B). If component B were at a lower redshift, then it could have contaminated the spectrum of the primary component, mimicking LyC emission. The HST imaging clearly solves the problem of the close neighbor by showing that the LyC emission arises only from component A (see Figure 1). The LyC emission is unambiguously confirmed at ${\rm{S}}/{\rm{N}}=10$ with magnitude $m({\rm{F}}336{\rm{W}})=27.57\pm 0.11$. Also, for the first time, it provides spatial mapping of the ionizing radiation from a galaxy. From deep multi-wavelength observations, including Chandra 6 Ms image, HST optical (GOODS, Giavalisco et al. 2004) and near-infrared (CANDELS, Grogin et al. 2011), Spitzer (3.6, 4.5, 5.8, 8.0, 24 μm), and wide spectroscopic coverage from the U to K bands (VLT and Keck), we found Ion2 to be a low-mass ($\leqslant {10}^{9}{M}_{\odot }$), low-metallicity ($\sim 1/6{Z}_{\odot }$), star-forming galaxy (de Barros et al. 2016).

We derived the absolute escape fraction quantity, ${f}_{\mathrm{esc}}(\mathrm{LyC})$, with the usual formulation (e.g., Vanzella et al. 2012):

$$\begin{eqnarray}{f}_{\mathrm{esc}}(\mathrm{LyC}) & = & \displaystyle \frac{{(L1500/L800)}_{\mathrm{INT}}}{{(f1500/f800)}_{\mathrm{OBS}}}\\ & & \times \displaystyle \frac{1}{T{(\mathrm{IGM})}_{{\rm{F}}336{\rm{W}}}}\times {10}^{-0.4\times {A}_{1500}},\end{eqnarray} \tag{ 1 }$$

which is related to the relative escape fraction, ${f}_{\mathrm{esc},\mathrm{rel}}(\mathrm{LyC})$, as ${f}_{\mathrm{esc},\mathrm{rel}}(\mathrm{LyC})={10}^{0.4{A}_{1500}}{f}_{\mathrm{esc}}(\mathrm{LyC})$. Given the low dust attenuation derived in de Barros et al. (2016), $E(B-V)$ $\;\lt \;0.04$ (${A}_{1500}\lt \;0.4$), the relative and absolute fractions are very similar in this case (they coincide if $E(B-V)$ = 0 or ${A}_{1500}=0$). For simplicity, we assume no dust attenuation. The f1500/f800 is the observed flux density ratio ($=14.60$) calculated on component A only (where the LyC arises) using the same circular apertures of 02 radius and subtracting component B from the HST/ACS bands before running SExtractor on A, as described in Vanzella et al. (2015). T(IGM)_F336W is the IGM transmission for the F336W filter, which is obtained by convolving the 10,000 IGM transmissions at z = 3.2 (Inoue et al. 2014) with the F336W transmission curve (as described in Vanzella et al. 2010a, 2015). The intrinsic luminosity density ratio L1500/L800_INT is observationally uncertain and must be estimated from stellar population synthesis models.

Although we have a clear detection at ($\lambda \lt 890$ Å), it is not possible to compute a precise value of ${f}_{\mathrm{esc}}(\mathrm{LyC})$, which depends on the IGM transmission and the intrinsic luminosity density ratio, L1500/L800_INT. Therefore we derive a range of plausible values taking advantage of the detected LyC signal ($m({\rm{F}}336{\rm{W}})=27.57\pm 0.11$). A range of possible IGM transmissions is obtained by assuming that the minimum theoretical intrinsic flux density ratio is 1 (e.g., Siana et al. 2007) and that ${f}_{\mathrm{esc}}(\mathrm{LyC})\leqslant 1$ (Vanzella et al. 2012). Under these assumptions, the IGM transmission T(IGM)_F336W cannot be lower than 0.068 (otherwise $m({\rm{F}}336{\rm{W}})$ would be brighter than the magnitude at 1500 Å). In Figure 3, we show the resulting ${f}_{\mathrm{esc}}(\mathrm{LyC})$ distribution adopting T(IGM)${}_{{\rm{F}}336{\rm{W}}}\gt 0.068$ and the intrinsic ratio (L1500/L800)${}_{\mathrm{INT}}=5$. The latter is an intermediate value between the predictions for an instantaneous burst and the predictions for a constant star formation history (Siana et al. 2007): given the ionizing photon production rate (see next section) and the age derived from SED fitting (de Barros et al. 2016), we expect the intrinsic ratio to be between ∼3 and ∼7. In particular, the observed LyC emission suggest the presence of O-type stars, and consequently an interval of time since the onset of the last burst of a few million years.

Zoom In Zoom Out Reset image size

In the most extreme case, adopting (L1500/L800)${}_{\mathrm{INT}}\gt 1$ and the maximum IGM transmission, the minimum ${f}_{\mathrm{esc}}(\mathrm{LyC})$ is 10%. However, as shown in Figure 3, the degeneracy introduced by the IGM stochasticity prevents us from placing strong constraints on ${f}_{\mathrm{esc}}(\mathrm{LyC})$ or (L1500/L800)_INT, as various combinations of ${f}_{\mathrm{esc}}(\mathrm{LyC})$, (L1500/L800)_INT, and T(IGM) can produce the observed F336W flux. Adopting (L1500/L800)${}_{\mathrm{INT}}=5$, an ${f}_{\mathrm{esc}}(\mathrm{LyC})$ of $50(100) \%$ is derived if T(IGM)_F336W is 34(73)%. It is worth noting that this source has been selected as a candidate LyC emitter with a method that eventually favors IGM and/or ISM transparencies (Vanzella et al. 2015). Therefore, an IGM transmission higher than the average along this line of sight would be not surprising.

4.1. The Ionizing Photon Production Rate

4.2. LyC Morphology

The LyC leakage is co-spatial with component A and nothing is detected from component B, where A and B refer to the brighter and the fainter blobs, respectively (see Figure 1, following the Vanzella et al. 2015 nomenclature). Ion2 is a compact but well-resolved galaxy in the F435W, F606W, F775W, and F850LP ACS images, corresponding to the rest-frame far-UV from 1030 to 2020 Å. The galaxy, however, does not appear to be resolved in the WFC3 F336W image. We took a well-exposed, nearby bright star and inserted it into a blank area of the image at random positions to simulate a number (N = 50) of realizations of the point-spread function (PSF), after rescaling it to the same flux as Ion2. We then aligned the star at the same position of the light centroid of Ion2 and subtracted it, in each case obtaining residuals consistent with the sky background (an example is shown in Figure 4). We have also compared the morphology of Ion2 in the F336W and F435W bands to test the possibility that the lower S/N in the former band is the reason for the apparent unresolved morphology, adopting the following procedure. Basic PSF-corrected morphological information from both components was extracted by Vanzella et al. (2015) using Galfit (Peng et al. 2010). Given their regular (symmetric) morphology, the Galfit modeling with a Gaussian light profile well reproduces both regions in all of the ACS images and shows that they are spatially resolved with effective radii of a few hundred parsecs. In particular, in Figure 5, we show the result obtained at 1000 Å rest-frame (F435W band) adopting a Gaussian shape. The best solution produces an effective radius of ${R}_{e}=340\pm 25$ pc, suggesting that the stellar radiation emerges from a compact and resolved region; the error has been calculated by running galfit on simulated images obtained by inserting the B-band model into random and free regions of the F435W-band image. Galfit fitting on the LyC image does not produce resolved solutions. In order to characterize the LyC morphology, we perform a simple test by subtracting the best-fit B435-band model from the F336W image. After normalizing the two images at the same flux peak and checking the residuals over a grid of dimming factors, it turns out that the LyC emission arises from a region smaller than what has been inferred from the F435W-band observations (see Figure 4); in particular, the spatially unresolved image in F336W has an effective radius $\lt 200$ pc. The negative residual flux seen in Figure 4 extends at least to 5 pixels radially from the center, and cannot be explained by the small variations in the PSFs at the two wavelengths. The observed compactness of the LyC emission is probably not surprising if we compare our HST resolution to the size of the super-star clusters observed in local starbursts, in which the O-type stars are spatially segregated toward the center within tens of parsecs (e.g., Annibali et al. 2015; James et al. 2016).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Another possible interpretation is that the LyC emission of Ion2 could originate from an AGN. This is unlikely and is discussed in Section 5.1.

Independent of the nature of ionizing photons, the LyC detection implies a low column density of neutral gas along the line of sight, lower than ${10}^{17.2}$ cm⁻² (corresponding to $\tau (\mathrm{LyC})\lt 1$). de Barros et al. (2016) have shown for the first time empirical evidence linking high ${f}_{\mathrm{esc}}(\mathrm{LyC})$ with the compactness of the star-forming region, the large $[{\rm{O}}\;{\rm{III}}]$/[O ii] line ratio, the weak ultraviolet interstellar absorption lines, and the emerging Lyα emission close to the systemic redshift, as predicted by photoionization and radiative transfer models when the medium is considered to be optically thin (Jaskot & Oey 2013; Borthakur et al. 2014; Nakajima & Ouchi 2014; Verhamme et al. 2015; Izotov et al. 2016). The HST observations presented here unambiguously confirm those predictions.

5.1. Nature of the Source of the LyC Photons

In the following, we summarize a few key elements (de Barros et al. 2016) and add new empirical and theoretical evidence that favors the stellar origin of the ultraviolet light rather than an AGN origin.

• 1.  
  No high-ionization emission lines like ${\rm{N}}\;{\rm{V}}\lambda 1240$, ${\rm{C}}\;{\rm{IV}}\lambda 1550$, or $\mathrm{He}\;{\rm{II}}\lambda 1640$ have been detected at more than the 3σ level from the VLT/VIMOS MR spectrum (Figure 6), while the [Ciii]λ1909 line is clearly detected. From the ${\rm{C}}\;{\rm{IV}}\lambda 1550/[{\rm{C}}\;{\rm{III}}]\lambda 1909\lt 0.15$ and $[{\rm{C}}\;{\rm{III}}]\lambda 1909/\mathrm{He}\;{\rm{II}}\lambda 1640\gt 4.0$ line ratios, and following Feltre et al. (2016), this source is classified as a star-forming galaxy lying in the same region occupied by low-metallicity galaxies of Stark et al. (2014).
• 2.  
  Ion2 is spatially resolved in all of the ACS bands (340 ± 25 pc) implying that the stellar emission is dominating the observed range 1000–2000 Å rest-frame.
• 3.  
  The narrow width of the [O iii]λ5007 emission line ($\sigma =65$ km s⁻¹) is compatible with lower-redshift star-forming galaxies (Maseda et al. 2014), while the AGN population typically shows higher velocity dispersions ($\gt 200$ km ${{\rm{s}}}^{-1}$, e.g., Osterbrock & Mathews 1986).
• 4.  
  Ion2 is not detected in the X-ray in the 4 Ms CDFS (Cappelluti et al. 2016) and no evidence of emission is inferred from the recent Chandra pointings publicly available on the Chandra website¹³ , which increased the exposure time to ∼6 Ms (see Figure 6). This places a limit on the X-ray luminosity of ${L}_{X}\lesssim 3\times {10}^{42}\;\mathrm{erg}\;{{\rm{s}}}^{-1}$ at the $1\sigma$ limit. If Ion2 were a type I AGN, and assuming that the ${L}_{X}\mbox{--}{L}_{[{\rm{O}}{\rm{III}}]}$ relation observed locally is valid at z = 3, then the measured [O iii]λ5007 luminosity (de Barros et al. 2016) would imply a detection with ${\rm{S}}/{\rm{N}}\gt 100$ at the Ion2 position in the 6 Ms X-ray image, corresponding to ${L}_{X}\simeq {10}^{45}\;\mathrm{erg}\;{{\rm{s}}}^{-1}$ (Panessa et al. 2006; Ueda et al. 2015). The measured [O iii]λ5007 line luminosity ($2\times {10}^{43}\;\mathrm{erg}\;{{\rm{s}}}^{-1}$) and the non-detection in the 6 Ms X-ray would imply a high obscuration (if it were an AGN), formally with an equivalent column density ${N}_{H}\gt {10}^{25}{\mathrm{cm}}^{-2}$ (assuming solar metallicity). The estimated stellar mass from the LyC source is $\sim {10}^{8}\;{M}_{\odot }$ (see Section 5.2) and, following Maseda et al. (2014), the dynamical mass from the [O iii]λ5007 line width ($\sigma =65\;\mathrm{km}\;{{\rm{s}}}^{-1}$) is $\sim {10}^{9}{M}_{\odot }$. A lower limit on the BH mass of ${10}^{8}\;{M}_{\odot }$ can be estimated by converting the [O iii]λ5007 line luminosity into an AGN bolometric luminosity (Panessa et al. 2006; Lusso et al. 2012) and then assuming that the AGN is radiating at the Eddington limit. This mass would imply a very unusual and probably unrealistic ratio of M_BH/Total mass $\geqslant 0.1$ for this object.
• 5.  
  Adopting the expected L_X (from [O iii]λ5007 reported above), we expect a clear detection at 6 μm rest-frame (i.e., at 24 μm observed with Spitzer/MIPS) by assuming that the relation between L_X and $L(6\;\mu {\rm{m}})$ (Stern 2015) is valid. In particular, even when adopting an ${L}_{X}\gt {10}^{44}$ erg s⁻¹, we expect L(6 μm)$\;\gt \;{10}^{44}$ erg s⁻¹ that corresponds to mag $(24$ μm)$\;\lt \;20$(AB) and therefore a detection in the MIPS 24 μm image at ${\rm{S}}/{\rm{N}}\gt 10$. At the position of Ion2, no signal is detected in the MIPS $24\;\mu {\rm{m}}$ band down to AB ≃ 22.3 (de Barros et al. 2016).

Zoom In Zoom Out Reset image size

It is worth noting a photometric excess (${\rm{S}}/{\rm{N}}\gt 5$) in the second IRAC channel at 4.5 μm (Figure 6) that we ascribed to possible He i $\lambda 10830$+Pγ lines with rest-frame equivalent width of ∼1100 Å. While these lines can only be observed with the James Webb Space Telescope (JWST), such strong He i $\lambda 10830$ emission up to a 1000 Å equivalent width has been observed in local compact H ii regions by Izotov et al. (2014) and in a local LyC emitter candidate (Verhamme et al. 2015).

Therefore, we conclude that the signal observed in the HST/F336W band is very likely due to stellar emission; the only (unlikely) alternative would be the presence of a heavily obscured AGN, located at the center of the star-forming region and hidden by gas and dust at all wavelengths but visible in the ionizing continuum, a possibility we deem very contrived. An intrinsically faint AGN could co-exist with dominant star formation activity that controls the transparency of the medium. In this regard, we already proposed in Vanzella et al. (2015) the possible hybrid configuration in which stellar and non-stellar (AGN) ionizing emission could co-exist in some systems and might explain the tension found between the UV excess and the stellar population synthesis models reported in the literature. In the present case, a partial AGN contribution could also explain the LyC leakage, although a more dedicated discussion of the physical mechanisms that lower the column density of neutral gas would be needed and this is beyond the scope of this work. Moreover, despite the excellent deep multi-wavelength coverage, dedicated observations will be necessary to detect a possible faint AGN. For example, monitoring for variability could be a powerful diagnostic. Currently, there is no observational evidence of nuclear activity in this object, while stellar emission is certainly detected in all of the HST images.

5.2. Constraining the Burst of Star Formation

While the LyC emitter reported here is rare among sources with similar luminosity (Vanzella et al. 2010a; Grazian et al. 2016), the non-ionizing multi-frequency properties observed in our galaxy and the confirmed LyC emission provide valuable prospects for the characterization of similar or fainter sources in the higher-redshift domain. In particular, the confirmed $z\gt 7.5$ galaxies (Finkelstein et al. 2013; Oesch et al. 2015; Zitrin et al. 2015) show particularly strong Oxygen and Balmer structure ([O iii]λλ4959,5007 + Hβ), at the same level as reported here (equivalent widths larger than 800–1000 Å rest-frame). It is premature to conclude that those sources effectively have an ${f}_{\mathrm{esc}}(\mathrm{LyC})\gt 0$, as optical line ratios are needed to perform a direct comparison and this must be postponed until JWST launch. Our result is currently a unique high-redshift reference, both in terms of large Oxygen line ratio and large line equivalent width, and needs to be extended to statistically significant samples, especially investigating the faint luminosity domain. However, it demonstrates the feasibility of the identification of ionizing sources during the reionization epoch.

We thanks the anonymous referee for a helpful report which greatly improved this manuscript. We thank F. Annibali and D. Schaerer for useful discussions. Part of this work has been funded through the INAF grants (PRIN INAF 2012). M.B. acknowledges support from the FP7 grant "eEASy" (CIG 321913). Support for program 14088 was provided by NASA through a grant from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555.