Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_Constraints_on_coronal_turbulence_models_from_sour.txt
|
an artificially low value of 1 m, instead of being computed self-consistently from Eq 4.
|
Processed_Constraints_on_coronal_turbulence_models_from_sour.txt
|
scattering angle by around two orders of magnitude as compared to the plane wave case.
|
Processed_Constraints_on_coronal_turbulence_models_from_sour.txt
|
for scattered solar radio emission and plane wave results should be used with great caution.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Night sky brightness is a major source of noise both for Cherenkov telescopes as well as for wide-angle Cherenkov detectors. Therefore, it is important to know the level of night sky brightness at potential sites for future experiments. The measurements of night sky brightness presented here were carried out at Fowler’s Gap, a research station in New South Wales, Australia, which is a potential site for the proposed TenTen Cherenkov telescope system and the planned wide-angle Cherenkov detector system HiSCORE.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
August 2010. Brightness levels were measured for a range of different sky regions and in various spectral bands.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The term “night sky brightness” (NSB) refers to the residual light that is present in the night sky during dark, It poses a major source of noise for moonless nights. ground-based astronomical observations, and good astro- nomical sites are therefore characterised by their low level of NSB (“darkness”) (Garstang, 1989).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
In the visible light regime, the main contributions of NSB are from air-glow, direct and scattered starlight and zodiacal light. Cities, but also facilities like ports or min- ing sites, produce a high level of light pollution, and the NSB near such locations is often dominated by artificial light. The pollution can extend many kilometres beyond the boundaries of the city, as can be seen in the world- wide atlas of night sky brightness (Cinzano et al., 2001). At good observatory sites, however, the contribution of ar- tificial sources to the total NSB should be less than 1%, and the atlas provides a good first check for appropriate sites.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
A comprehensive overview about sources and measure- ments of NSB is given in Leinert et al. (1998). In Benn and Ellison (1998) a long term measurement of the NSB levels at the La Palma observatory is presented, including a de- tailed study of the spectrum of NSB using a spectrograph mounted at one the the observatory’s telescopes. In the preparation of the H.E.S.S. telescope system, two candi- date sites in La Palma and in Namibia have been examined in respect of their NSB levels, using a photon counting pho- tomultiplier and a portable telescope mount (Preuß et al., 2002). A similar measurement has been conducted by Mir- zoyan and Lorenz (1994) using the first HEGRA telescope at La Palma.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
In this paper, the night sky brightness is examined in the light of its effects on air Cherenkov light detectors for gamma-ray astronomy. Cherenkov telescopes like MAGIC (Albert et al., 2008), VERITAS (Acciari et al., 2008) and H.E.S.S. (Aharonian et al., 2006) observe cosmic gamma- rays and charged cosmic rays using the air Cherenkov tech- nique: The primary particles interact with atoms in the upper atmosphere and produce a cascade of secondary, still highly energetic particles (air shower). These parti- cles (mainly the electrons) emit Cherenkov light, and the air shower can be observed by Cherenkov telescopes with a very fast and sensitive camera system. The reconstruction of the primary particle properties (direction, energy, and others) uses the image properties of the air shower and, in some systems, its time evolution.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
NSB adds a constant level of background into each pixel of the camera, and its fluctuations limit the sensitivity to very weak signals. The spectral region from 280 nm to about 500 nm is the most relevant source of noise, as both the Cherenkov light spectrum and the spectral sensitivity of the photomultiplier tubes used in Cherenkov telescopes peak in the blue.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
NSB is an even greater concern for wide-angle Cheren- kov detectors like AIROBICC (Karle et al., 1995), TUNKA (Ptuskin, 2008) and the proposed HiSCORE detector (Tlu- czykont et al., 2009). These instruments do not use an imaging system but integrate light from a large solid angle (up to 1 sr), which leads to a much higher level of NSB compared to Cherenkov telescopes, where the solid angle (10−5 sr). of a single pixel is of In its simplest form, these detectors consist of an array of large photomultipli- ers aligned towards zenith, where the spacing between the detectors can be up to a few hundred meters. The recon- struction of the primary particle properties is done using the lateral photon distribution at detector level, the arrival times of photons and the duration of the Cherenkov light signals (Hampf et al., 2009).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
In these detectors the photon intensity due to NSB is high enough to not only disturb the air shower recon- struction but also to produce a non-negligible current in the photomultiplier anode, which requires the operation of these photomultipliers at low gain.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
This work presents NSB measurements conducted at the Australian research facility Fowler’s Gap1 (New South Wales, coordinates: 31◦ 05’ S, 141◦ 43’ E) in the context of a campaign to find a suitable site for the proposed Aus- tralian Cherenkov telescope system TenTen (Rowell et al., 2008) as well as the planned wide-angle Cherenkov detec- tor HiSCORE. The site is about hundred kilometres away from the nearest city (Broken Hill), and from maps pre- sented in Cinzano et al. (2001) the artificial light pollution is expected to be negligible.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
data (G. Thornton et al 2011, in preparation). Also, the property of the research facility provides sufficient space for a large detector field like HiSCORE.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Fowler’s Gap is at an elevation of about 180 m, which is far lower than the detector sites examined by the groups cited above. Simulations indicate that a low altitude site is very well suited for detectors that are designed for the ultra high energy regime (> 30 TeV) of gamma-rays, like It is however not a priori clear HiSCORE and TenTen. how NSB varies with altitude, and all studies so far have concentrated on high altitude sites. This study is therefore an important step towards the evaluation of a low altitude site for Cherenkov astronomy.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The measurements presented here were carried out for the summer sky (14th and 15th February 2010) and the winter sky, including the Galactic Centre region (12th and 13th August 2010).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The detector used follows the concept outlined in Preuß et al. (2002) with a few modifications to obtain a better calibration and a wider coverage of wavelengths. The de- tector and its calibration are described in detail in section 2. Section 3 outlines the mathematical procedures used to obtain the NSB levels from the measurements. Section 4 summarises the NSB levels measured in Fowler’s Gap for various regions of the sky. The measured spectrum of NSB is presented in section 5 and the use of band-pass fil- ters to suppress NSB is discussed in section 6. Finally, the implications of the measured light levels for the TenTen telescope system as well as for the HiSCORE detector are discussed in section 7.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
To measure the night sky brightness levels, a photo- multiplier module with photon counting capability (Hama- matsu HC124-3) is used. The module incorporates a R268P bi-alkali photomultiplier tube (PMT), a Cockroft-Walton high-voltage generator, an amplifier, a pulse shaper and a discriminator into a single, rugged metal casing. The threshold is internally set to discriminate noise from the single photon events, and signals are given according to the TTL standard. The PMT electronics can handle rates up to about 1 MHz, and the dark count rate is in the order of 200 Hz.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The PMT is fixed in a metal tube, which in turn is installed on a standard telescope mount (Meade LX D55). The telescope mount can be controlled by a hand-held con- troller or by a computer using a serial connection. A long plastic tube (inner diameter 17 mm, length 580 mm) is positioned in front of the PMT in order to reduce the rate of the NSB photons to rates between 50 kHz and several hundred kHz. The set-up is shown in figures 1 and 2.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
programme. Additionally, this programme controls the high voltage of the PMT and records the ambient tem- perature as well as the position of the telescope mount in equatorial coordinates. All data are logged into an ASCII file for later analysis.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The supply voltages for the PMT module and the tem- perature readout electronics are generated by four 12 V rechargeable lead batteries.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
where ǫ(θ) is the angle-dependent transmission through the plastic tube (Preuß et al., 2002). The transmission ǫ(θ) is calculated by a ray-tracing algorithm, which con- tains the (a priori unknown) reflectivity of the black plas- tic tube as a free parameter. An analytical calculation has been performed for the case of no reflection in order to cross-check the ray-tracing simulation. In both cases the reflectivity has been assumed to be independent of angle and wavelength.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
In order to further cross-check the simulation and to obtain the reflectivity, ǫ(θ) has also been determined ex- perimentally. For this, the set-up was pointed towards a point-like light source in about twenty meters distance in a dark indoor corridor. Rays reaching the front opening of the plastic tube from that distance are nearly parallel, the maximum difference of angles is less than 3 arc-minutes. Using the telescope mount controls the tube can be pointed to various angles away from the light source, and for each angle the intensity of the transmitted light is recorded by the PMT module. The results of the calculation, the ray- tracing simulation and the measurements, all normalised to the intensity at parallel incident light, are shown in fig- ure 3.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The analytical calculation and the simulation result for the case of no reflection agree very well. Simulations were carried out for different reflectivities in one percent steps, and the result of the simulation with 11% reflectivity de- scribes the experimental values best. Towards higher an- gles the measured transmissions are slightly above the ray- tracing values; this is mainly due to stray light being re- flected into the device from the corridor walls and residual light in the corridor. At small angles the measured values are higher than expected as well. The reason for this is unknown; however, the contribution of these values to the effective solid angle is small.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
In order to check the accuracy of the frequency mea- surement with LabView, several cross-checks with a stand- alone rate meter were performed, and no discrepancy was found up to 1 MHz.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
To check the linearity of the count rate to the incident light intensity, the device has been pointed at a bright re- gion of the sky and several combinations of neutral density filters have been used to reduce the rate by known ratios. The deviations from linearity were within 4%.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The stability has been tested in various long runs (12 to 24 hours) of the device in a dark room. A LED was switched automatically to several defined intensities in reg- ular intervals and the count rate of the PMT module was recorded. Deviations throughout the night were less than 1%, and deviations between different runs were smaller than 2%.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Count rate at PMT: This is the count rate that was recorded by the device during the measurements. It is given in order to allow direct comparison of results of future campaigns using the same device.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Photoelectron intensity at PMT: This value gives the above count rate in units of (s sr m2)−1, using the area and the solid angle given in Eq. 1 and 3. This can be used directly as input for detector sim- ulations, assuming that the PMT used here and the ones used in the experiments have sufficiently sim- ilar spectral response and peak quantum efficiency. The systematic uncertainty of this value is caused by uncertainties in the solid angle, the effective area of the tube and slight changes in the dark count rate, and is estimated to be about 5%.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Table 1: The effective wavelength window size wx for the four John- son filters that were used during the spectral measurements, calcu- lated by Eq. 7. The effective wavelength midpoints are taken from Binney and Merrifield (1998).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
is found. It is estimated that this value has an un- certainty of about 10%, resulting in an uncertainty of the integral photon intensity of 15%.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
While the integral photon intensity allows the easi- est comparison to other studies taken with different instruments (e.g., CCD cameras at telescopes), it is subject to larger systematic uncertainties introduced by the uncertainties in the quantum efficiency and the assumed night sky spectrum.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
with the PMT quantum efficiency ǫPMT(λ) and the trans- mission function of the filter Tx(λ). The transmission func- tions were taken from the manufacturer’s data sheets; a cross-check with a spectrograph gave consistent results. Table 1 shows the calculated values of wx for the four Johnson filters.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
During the measurements in Fowler’s Gap, the device was pointed at several selected dark regions in the sky. Each region was measured several times each night, and the results were averaged. The differences between the measurements were less than 5%, and probably mostly due to the different zenith angles of the individual measure- ments (unfortunately, not enough data could be collected during this campaign to evaluate the zenith angle depen- dence systematically).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The upper part of table 2 gives the coordinates and the measured brightness levels at several dark regions that were observed during the two campaigns in February and August 2010. Most values found here are slightly lower than the one given in Preuß et al. (2002) and higher than in Mirzoyan and Lorenz (1994), but consistent with both within the systematic uncertainties. Further comparisons with other measurements, which were taken in certain pho- tometric bands, are given in section 5.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The dark regions 7 and 8 (Virgo and Libra constella- tions) are significantly brighter than the rest of the dark regions, even though they are well away from the galactic plane and other bright regions. However, since they were close to the horizon during the time of observation and their position on the sky is coincident with the ecliptic, it is possible that the higher photon intensity is at least partly due to zodiacal light. Comparing with measurements of the brightness of zodiacal light, e.g. by Bernstein et al. (2002), one finds that the observed increase of brightness is in the expected order of magnitude.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
As many known and potential high energy gamma-ray sources are located in brighter regions of the sky like the galactic plane, the night sky brightness levels in these re- gions are of great interest for this study. The lower part of table 2 shows some measurements from brighter regions of the sky.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Figure 4 shows two scans across the plane in the Ca- nis Major region, one of the darker regions of the galactic plane. The two scans, which were taken along a decli- nation of -22◦ in two successive nights, are in reasonable agreement. The maximum photon intensity in this scan is 1012 photons/(s sr m2), almost twice as bright about 3.5 as in the darkest regions of the sky.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Figure 4: Scan along a declination of -22◦ through the Canis Ma- jor region. The different markers present measurements from two different nights during the February campaign.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Major area and almost four times as high as in the darkest regions of the sky.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The situation is similar in the Galactic Centre / Sagit- tarius region, one of the brightest regions in the sky, a scan of which is shown in figure 6. The very bright outlier at 17.8 hours is probably due to a single bright star, while the actual peak of the scan is found at 17.3 hours with 1012 photons/(s sr m2). This confirms a brightness of 7 the measurement of Preuß et al. (2002) who found that the Galactic Centre region is up to four times brighter than the darkest regions of the sky.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The statistical errors in the data points of figures 4 to 6 are negligible. The systematic error is assumed to be 15%, mainly due to uncertainties in the instrument calibration (see section 3).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The spectral composition of the NSB has been exam- ined by using standard astronomical Johnson filters in front of the entry window of the set-up. The measured rates in U, B, V, and R band are converted to a differ- ential photon intensity at the effective mid-wavelength of the filters using Eq. 6.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Table 2: Night sky brightness levels observed in February (F) and August (A) 2010 from Fowler’s Gap, and references from Preuß et al. (2002) ([1], Namibia and La Palma) and Mirzoyan and Lorenz (1994) ([2], La Palma). The upper part of the table contains selected dark regions, the lower part brighter regions of the sky. The photoelectron and photon intensities are given per sr and m2 (see section 3 for definition of these numbers). LMC denotes the Large Magellanic Cloud and GC the Galactic Centre.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Figure 5: Scan along the right ascension of 11:30 hours through the Southern Cross region, taken in February 2010.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Figure 6: Scan along the declination of -30◦ through the Galactic Centre region, taken in August 2010.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
sea level in Namibia and 2200 m a.s.l. at La Palma, while the measurements from Fowler’s Gap were taken at 180 m a.s.l., and during a medium solar activity.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
photon intensities observed in V and R band. On the other hand, the lower altitude of Fowler’s Gap as compared to all other shown measurements should lead to stronger absorption of night sky brightness, especially at shorter wavelengths, which can explain the lower intensities seen in U and B band.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
PMTs are much less sensitive at longer wavelengths than in the blue, so that the light in V and especially in R band does not contribute as much to the overall brightness seen by the PMT as the U and B bands. Figure 8 shows the spectra before the unfolding of the PMT response, i.e.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Figure 7: Averaged spectra of dark regions 1, 2, 3, 6, and 7 as mea- sured from Fowler’s Gap with astronomical Johnson filters (U, B, V, and R bands) and measurements from Preuß et al. (2002) (”PMT”) and Benn and Ellison (1998) (”La Palma CCD“). Statistical un- certainties are indicated by errorbars, systematic uncertainty is 15% (see section 3).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
the light levels actually seen by the PMT, for an average of dark regions (the same as in the plot of figure 7) and bright regions (the Crux region, the LMC and the Monoceros region as defined in table 2). The highest contributions come from the B and V band, which are of roughly the same strength even though the photon intensity in V band is almost three times higher than in B band (see figure 7). There is less contribution from the U band and even less from the R band, even though the NSB is dominated by long wavelengths. There is no significant difference in the shape of the spectrum between dark and bright regions of the sky.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Table 3 shows the measured brightness in B and V band in different units. The magnitude in V band of a dark re- gion of the sky is often used as a measure for the darkness of an astronomical site. A value of 21.6 mag/arcsec2 can be considered to indicate a good dark site (Garstang, 1989) and compares to the most renowned observatory sites of the world: Measured values for the European Southern Observatory in Chile are between 21.69 and 21.91, and for the McDonald Observatory in the United States between 21.54 and 21.92 (Leinert et al., 1998). In those measure- ments all stars brighter than magnitude 13 were excluded, which was not possible in the measurements at Fowler’s Gap, so that a comparable measurement would give an even (slightly) better value at Fowler’s Gap.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Figure 8: Spectra of dark and bright regions as measured in Fowler’s Gap, without unfolding the PMT response (i.e., the light levels ac- tually seen by the PMT). Statistical uncertainties are indicated by errorbars, systematic uncertainty is 15% (see section 3).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
noise ratio. Figure 9 shows the photoelectron intensity at the PMT caused by NSB and by Cherenkov light in the four used photometric bands. The Cherenkov light has been simulated taking into account atmospheric absorp- tion between the height of a typical air shower maximum and the observation altitude, and the PMT quantum effi- ciency. Both spectra are normalised for better comparison. It can be seen that the Cherenkov light is strongest in U band and is constantly decreasing with wavelength, while the NSB peaks in the B band and does not fall off as quickly towards higher wavelengths as the Cherenkov light.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
It is therefore reasonable to assume that the signal to noise ratio can be improved if long wavelengths are blocked, e.g. by a low-pass filter in front of the PMT. Table 4 shows the fraction of Cherenkov light and NSB present in the various bands. If a filter blocks the V and R band but transmits the wavelengths in U and B band, about 92% of the Cherenkov but only 57% of the NSB will survive, resulting in an improvement of the signal to noise ratio of about 1.2. This requires however an ideal low-pass filter that reaches 100% transmission at short wavelengths.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Table 3: Measured NSB levels at Fowler’s Gap in B and V band in different units.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Table 4: Fraction of Cherenkov light and NSB present in the used photometric bands (using the same numbers as the plot in figure 9). The last column shows the gain in the signal to noise ratio which is achieved if only the respective wavelength bands are transmitted to the PMT. Shown are the numbers for dark regions as measured at Fowler’s Gap; however, the numbers for bright regions are almost identical.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Figure 9: Normalised spectrum of the dark regions (the same as in figure 7), and simulated Cherenkov light intensities in the four used bands (U, B, V, R). Both spectra contain atmospheric absorption and the photomultiplier response function (see section 6 for details).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
study (see Eq. 5), the expected photoelectron rates range from 53 MHz up to 190 MHz. The numbers can increase additionally if a bright star is in the field of view of the particular pixel. These photoelectron rates are in the same regime as those for other Cherenkov telescope systems, e.g. H.E.S.S. typically sees a NSB photoelectron rate of 100 MHz (Aharonian et al., 2004).
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The HiSCORE detector will consist of an array of non- imaging, angle integrating Cherenkov detectors. Every station will contain four detector modules, each consist- ing of a 20 cm diameter photomultiplier and a light col- lector (Winston cone) to increase the light sensitive area of the module. To derive the expected noise generated by NSB in one HiSCORE detector channel, the angular acceptance of the Winston cone is calculated using the simulation described in section 2.2. The upper opening is set at a diameter of 40 cm, the half opening angle at 30◦ and the reflectivity of the cone material at 90%, accord- ing to current design plans for the detector (Hampf et al., 2011). The effective solid angle is calculated with Eq. 2 to be Ω = 0.68 sr.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
To calculate the highest expected NSB rates in one HiSCORE module, a light band with a photon intensity profile as found in the scan of the Galactic Centre re- gion (figure 6) is used as light source for the ray-tracing simulation of the Winston Cone. This simulation gives a 1011 Hz, and a photoelectron rate of photon rate of 3.5 1010 Hz or 33 photoelectrons per nanosecond. For 3.3 comparison, a single bright star with magnitude zero adds less than 0.1 photoelectrons per nanosecond, so that even several bright stars within the field of view will not increase this level significantly.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The level of NSB determines the energy threshold of the detector system: First, the trigger threshold must be set high enough to keep the rate of false triggers from NSB fluctuations at a manageable level, second the accuracy of the reconstruction of events with low energy (and hence weak Cherenkov light emission) is limited by the signal to noise ratio. An upcoming paper by M. Tluczykont, D. Hampf, D. Horns et al. (2011) will present detailed sim- ulations of the HiSCORE detector, including a discussion of the effects of NSB on the trigger threshold and the re- construction.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
It can be concluded from the data that Fowler’s Gap is an excellent astronomical location in respect to the night sky brightness. However, long term measurements of the NSB in combination with extinction measurements are needed to rate the quality of the site more reliably. Cam- paigns for more measurements at this and other sites in Australia are planned.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The night sky brightness varies greatly with the region of the sky looked at. Regions on the galactic plane are more than three times brighter than the darkest regions of the sky. However, plane scans show that these high 10◦ from the light levels are restricted to within b = galactic plane, which means that the highest level of night sky brightness seen by the wide-angle Cherenkov detector HiSCORE will be at most twice that of dark regions.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
The night sky brightness is much stronger at longer wavelengths. However, standard photomultiplier sensitiv- ities peak at short wavelengths, so that the most domi- nant contribution for the background is from the B and V bands. The peak in Cherenkov light is at shorter wave- lengths, so that filters can be employed to improve the signal to noise ratio up to a factor of about 1.2.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Daniel Hampf likes to thank the German Federal Min- istry of Education and Research (BMBF) for its financial support (contract number 05A08GU1), as well as the Ger- man Academic Exchange Service (DAAD) for its support of the field studies in Australia under the ”Kurzstipendium f¨ur Doktoranden“ scheme.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Acciari, V. A., Beilicke, M., Blaylock, G., et al., Jun. 2008. VERITAS Observations of the γ-Ray Binary LS I +61 303. Astrophys. J.679, 1427–1432.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Aharonian, F., Akhperjanian, A. G., Aye, K., el al., Nov. 2004. Cali- bration of cameras of the H.E.S.S. detector. Astroparticle Physics 22, 109–125.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Aharonian, F., Akhperjanian, A. G., Bazer-Bachi, A. R., et al., Oct. 2006. Observations of the Crab nebula with HESS. Astr. Astro- phys.457, 899–915.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Albert, J., Aliu, E., Anderhub, H., et al., Feb. 2008. VHE γ-Ray Observation of the Crab Nebula and its Pulsar with the MAGIC Telescope. Astrophys. J.674, 1037–1055.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Bernstein, R. A., Freedman, W. L., Madore, B. F., May 2002. The First Detections of the Extragalactic Background Light at 3000, 5500, and 8000 ˚A. II. Measurement of Foreground Zodiacal Light. Astrophys. J.571, 85–106.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Binney, J., Merrifield, M., 1998. Galactic astronomy. Cinzano, P., Falchi, F., Elvidge, C. D., Dec. 2001. The first World Atlas of the artificial night sky brightness. Month. Not. Roy. Astr. Soc.328, 689–707.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Hampf, D., Tluczykont, M., Horns, D., Sep. 2009. Event reconstruc- tion with the proposed large area Cherenkov air shower detector SCORE. ArXiv 0909.0663.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
Hampf, D., Tluczykont, M., Horns, D., Apr. 2011. Simulation of the expected performance for the proposed gamma-ray detector HiSCORE. Proceedings of Science.
|
Processed_Measurement_of_night_sky_brightness_in_southern_Au.txt
|
M., Samorski, M., Aug. 1995. Design and performance of the an- gle integrating cerenkov array airobicc. Astroparticle Physics 3, 321–347.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
We provide a novel notion of what it means to be interpretable, looking past the usual association with human understanding. Our key insight is that interpretability is not an absolute concept and so we define it relative to a target model, which may or may not be a human. We define a framework that al- lows for comparing interpretable procedures by linking it to important practical aspects such as accuracy and robustness. We char- acterize many of the current state-of-the-art interpretable methods in our framework por- traying its general applicability.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
shows how the interpretation of a seemingly complex model can aid in creating a simple rule.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
The above example defines interpretability as humans typically do: we require the model to be understand- able. This thinking would lead us to believe that, in general, complex models such as random forests or even deep neural networks are not interpretable. How- ever, just because we cannot always understand what the complex model is doing does not necessarily mean that the model is not interpretable in some other use- ful sense. It is in this spirit that we define the novel notion of δ-interpretability that is more general than being just relative to a human.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
What does it mean for a model to be interpretable? From our human perspective, interpretability typi- cally means that the model can be explained, a qual- ity which is imperative in almost all real applica- tions where a human is responsible for consequences of the model. However good a model might have performed on historical data, in critical applications, interpretability is necessary to justify, improve, and sometimes simplify decision making.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
We offer an example from the healthcare domain (Chang & Weiner, 2010), where interpretability is a critical modeling aspect, as a running example in our paper. The task is predicting future costs based on de- mographics and past insurance claims (including doc- tor visit costs, justifications, and diagnoses) for mem- bers of the population. The data used in (Chang & Weiner, 2010) represents diagnoses using ICD-9-CM (International Classification of Diseases) coding which had on the order of 15,000 distinct codes at the time of the study. The high dimensional nature of diagnoses led to the development of various abstractions such as the ACG (Adjusted Clinical Groups) case-mix sys- tem (Starfield et al., 1991), which output various map- pings of the ICD codes to lower dimensional categorical spaces, some even independent of disease. A particu- lar mapping of IDC codes to 264 Expanded Diagnosis Clusters (EDCs) was used in (Chang & Weiner, 2010) to create a complex model that performed quite well in the prediction task.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Figure 1. Above we depict what it means to be δ-interpretable. Essentially, our procedure/model is δ-interpretable if it improves the performance of TM by ≥ δ fraction w.r.t. a target data distribution.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
on the strength of the machine (i.e. program) required to recognize it, we look to define interpretability along analogous lines.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Based on this discussion we would like to define in- terpretability relative to a target model (TM), i.e. δ- interpretability. The target model in the most obvious setting would be a human, but it doesn’t have to be. It could be a linear model, a decision tree or even an entity with superhuman capabilities. The TM in our running healthcare example (Chang & Weiner, 2010) is a linear model where the features come from an ACG system mapping of IDC codes to only 32 Aggregated Diagnosis Groups (ADGs).
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Our model/procedure would qualify as being δ- interpretable if we can somehow convey information to the TM that will lead to improving its performance (e.g., accuracy or AUC or reward) for the task at hand. Hence, the δ-interpretable model has to trans- mit information in way that is consumable by the TM. For example, if the TM is a linear model our δ- interpretable model can only tell it how to modify its feature weights or which features to consider. In our healthcare example, the authors in (Chang & Weiner, 2010) need a procedure to convey information from the complex 264-dimensional model to the simple linear 32-dimensional model. Any pairwise or higher order interactions would not be of use to this model. Thus, if our ”interpretable” model came up with some mod- ifications to pairwise interaction terms, it would not be considered as an δ-interpretable procedure for the linear TM.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Definition 2.1. δ-interpretability: Given a target model MT belonging to a hypothesis class H and a target distribution DT , a procedure PI is δ-interpretable if the information I it communicates to MT resulting in model MT (I) ∈ H satisfies the following inequality: eMT (I) ≤ δ · eMT , where eM is the expected error of M relative to some loss function on DT .
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
The above definition is a general notion of inter- pretability that does not require the interpretable pro- cedure to have access to a complex model. It may use the complex model (CM) but it may very well act as an oracle conjuring up useful information that will im- prove the performance of the TM. The more intuitive but special case of Definition 2.1 is given below which defines δ-interpretability for a CM relative to a TM as being able to transfer information from the CM to the TM using a procedure PI so as to improve the TMs performance. These concepts are depicted in figure 1.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Definition 2.2. CM-based δ-interpretability: Given a target model MT belonging to a hypothesis class H, a complex model MC, and a target distribution DT , the model MC is δ-interpretable relative to MT , if there exists a procedure PI that derives information I from MC and communicates it to MT resulting in model MT (I) ∈ H satisfying the following inequality: eMT (I) ≤ δ · eMT , where eM is the expected error of M relative to some loss function on DT .
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
inputs and outputs or could be through delving into the inner workings of the complex process. In addition, it is imperative that MT (I) ∈ H i.e. the information conveyed should be within the representational power of the TM.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
The advantage of this definition is that the TM isn’t tied to any specific entity such as a human and thus neither is our definition. We can thus test the utility linear, of our definition w.r.t. decision lists, etc.) given that a humans complexity maybe hard to characterize. We see examples of this in the coming sections.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
One may consider the more intuitive definition of δ- interpretability when there is a CM.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
We now clarify the use of the term Information I in the definition. In a normal binary classification task, training label y ∈ {+1, −1} can be considered to be a one bit information about the sample x, i.e. ”Which label is more likely given x?”, whereas the confidence score p(y|x) holds richer information, i.e. ”How likely is the label y for the sample x?”. From an information theoretic point of view, given x and only its training label y, there is still uncertainty about p(y|x) in the interval [1/2, 1] prior to training. According to our definition, an interpretable method can provide useful information I in the form of a sequence of bits or pa- rameters about the training data that can potentially reduce this uncertainty of the confidence score of the TM prior to training. Moreover, the new MT (I) is better performing if it can effectively use this informa- tion.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Moreover, a direct consequence of our definition is that it naturally creates a partial ordering of inter- pretable procedures relative to a TM and target distri- bution, which is in spirit similar to complexity classes for time or space of algorithms. For instance, if R+ denotes the non-negative real line δ1-interpretability ⇒ δ2-interpretability, where δ1 ≤ δ2 ∀δ1, δ2 ∈ R+, but not the other way around.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
We first extend our δ-interpretability definition to the practical setting where we don’t have the target dis- tribution, but rather just samples. We then show how this new definition reduces to our original definition in the ideal setting where we have access to the target distribution.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Our definitions above are motivated by the fact that when people ask for an interpretation there is an im- plicit quality requirement in that the interpretation should be related to the task at hand. We capture this relatedness of the interpretation to the task by requir- ing that the interpretable procedure improve the per- formance of the TM. Note the TM does not have to be interpretable, rather it just is a benchmark used to measure the relevance of the provided interpretation. Without such a requirement anything that one can elucidate is then an explanation for everything else, making the concept of interpretation pointless. Con- sequently, the crux for any application in our setting is to come up with an interpretable procedure that can ideally improve the performance of the given TM.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Our definition of δ-interpretability just focuses on the performance of the TM. However, in most practical applications robustness is a key requirement. Imagine a doctor advising a treatment to a patient. He better have high confidence in the treatments effect before prescribing.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
is that the newly created test sets are at least moder- ately probable w.r.t. the target distribution. Of course, in case of non-uniform loss functions the test sets on whom the expected loss is low are uninteresting. This brings us to the question of when is it truly interesting to study robustness.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
It seems that robustness is really only an issue when your test data on which you evaluate is incomplete i.e. it doesn’t include all examples in the domain. If you can test on all points in your domain, which could be finite, and are accurate on it then there is no need for robustness. That is why in a certain sense, low gen- eralization error already captures robustness since the error is over the entire domain and it is impossible for your classifier to not be robust and have low GE if you could actually test on the entire domain. The prob- lem is really only because of estimation on incomplete test sets (Varshney, 2016). Given this we extend our definition of δ-interpretability for practical scenarios.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
adversarial examples. But in practice this is not what people usually mean, when one talks about adversarial examples.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Given this, ideally, we should choose DR = DT so that we test the model mainly on important examples. If we could do this and test on the entire domain our Definition 3.1 would reduce to Definition 2.1 as seen in the following proposition.
|
Processed_A_Formal_Framework_to_Characterize_Interpretabilit.txt
|
Proposition 1. In the ideal setting, where we know DT , we could set DR = DT and compute the true errors, (δ, γ)-interpretability would reduce to δ- interpretability.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.