Maps of Dust Infrared Emission for Use in Estimation of Reddening and Cosmic Microwave Background Radiation Foregrounds

We present a full-sky 100 μm map that is a reprocessed composite of the COBE/DIRBE and IRAS/ISSA maps, with the zodiacal foreground and confirmed point sources removed. Before using the ISSA maps, we remove the remaining artifacts from the IRAS scan pattern. Using the DIRBE 100 and 240 μm data, we have constructed a map of the dust temperature so that the 100 μm map may be converted to a map proportional to dust column density. The dust temperature varies from 17 to 21 K, which is modest but does modify the estimate of the dust column by a factor of 5. The result of these manipulations is a map with DIRBE quality calibration and IRAS resolution. A wealth of filamentary detail is apparent on many different scales at all Galactic latitudes. In high-latitude regions, the dust map correlates well with maps of H I emission, but deviations are coherent in the sky and are especially conspicuous in regions of saturation of H I emission toward denser clouds and of formation of H2 in molecular clouds. In contrast, high-velocity H I clouds are deficient in dust emission, as expected. To generate the full-sky dust maps, we must first remove zodiacal light contamination, as well as a possible cosmic infrared background (CIB). This is done via a regression analysis of the 100 μm DIRBE map against the Leiden-Dwingeloo map of H I emission, with corrections for the zodiacal light via a suitable expansion of the DIRBE 25 μm flux. This procedure removes virtually all traces of the zodiacal foreground. For the 100 μm map no significant CIB is detected. At longer wavelengths, where the zodiacal contamination is weaker, we detect the CIB at surprisingly high flux levels of 32 ± 13 nW m-2 sr-1 at 140 μm and of 17 ± 4 nW m-2 sr-1 at 240 μm (95% confidence). This integrated flux ~2 times that extrapolated from optical galaxies in the Hubble Deep Field. The primary use of these maps is likely to be as a new estimator of Galactic extinction. To calibrate our maps, we assume a standard reddening law and use the colors of elliptical galaxies to measure the reddening per unit flux density of 100 μm emission. We find consistent calibration using the B-R color distribution of a sample of the 106 brightest cluster ellipticals, as well as a sample of 384 ellipticals with B-V and Mg line strength measurements. For the latter sample, we use the correlation of intrinsic B-V versus Mg2 index to tighten the power of the test greatly. We demonstrate that the new maps are twice as accurate as the older Burstein-Heiles reddening estimates in regions of low and moderate reddening. The maps are expected to be significantly more accurate in regions of high reddening. These dust maps will also be useful for estimating millimeter emission that contaminates cosmic microwave background radiation experiments and for estimating soft X-ray absorption. We describe how to access our maps readily for general use.


INTRODUCTION
In the past 15 years two NASA missions have revolutionized our knowledge of the di †use interstellar medium. The path-breaking Infrared Astronomy Satellite (IRAS) mission of 1983 led to the Ðrst full-sky maps of the di †use background radiation in four broadband infrared channels, centered at 12, 25, 60, and 100 km, with a D5@ beam. The DIRBE experiment (Di †use Infrared Background Experiment) on board the COBE satellite imaged the full sky in 10 broad photometric bands from 1 to 240 km with a beam of This experiment, for all but the shortest wave-0¡ .7. length channels, was active for 42 weeks in 1989È1990 before its 4He cryogen was exhausted. Although IRAS was optimized to detect point sources and sources of small angular extent et al. it has been possible to (Beichman 1988), create large-area sky maps from the IRAS data stream (ISSA images ;et al. Striping artifacts from Wheelock 1994). time variation in the zodiacal foreground emission have largely been Ðltered out of the individual ISSA maps, but artifacts remain and the zero point has large-scale drifts across the images. The DIRBE experiment had a much better control of the absolute calibration and has two further channels to the submillimeter than does IRAS (at 140 and 240 km), but the maps have much lower angular resolution.
The di †use emission seen in these experiments is a superposition of zodiacal foreground emission, dust and molecular emission from the interstellar medium (ISM), point sources from within and beyond the Milky Way, and possibly a di †use extragalactic background. Separating these components has proven to be a most difficult task (Hauser The zodiacal emission dominates the 12 and 25 km 1996). channels and is a quite serious contaminant at 60 km, while 525 SCHLEGEL, FINKBEINER, & DAVIS Vol. 500 its contribution to the 100 km maps is less strong. Modeling of the zodiacal emission has proven quite complex et (Reach al. One must consider that the zodiacal dust in the 1996). ecliptic plane has a distribution of temperatures, that the tilt of EarthÏs orbit relative to the midplane of the dust is readily detectable, and that the zodiacal emission shows a strong dependence on solar elongationÈinformation that is lost if one uses only the annual-averaged map of the DIRBE data.
A full-sky map of the di †use radiation serves many purposes, and we here report our e †orts to combine the 100 km maps of IRAS and DIRBE in a manner designed to be accessible to the general astrophysical community. To construct this map, we have removed the zodiacal foreground emission from the DIRBE maps, removed striping artifacts from the IRAS/ISSA maps, subtracted conÐrmed point sources, and combined the maps in such a way as to preserve the DIRBE calibration and IRAS resolution. We have used the ratio of 100 km to 240 km emission to deduce a dust color temperature, allowing us to translate the 100 km emission to a column density of radiating dust. (A preliminary version of this work appeared in Such a Schlegel 1995.) procedure may be inadequate toward complex, dense clouds at low Galactic latitude, but in most directions the emission is dominated by dust within a single environment and radiation Ðeld.
Since the di †use emission in the infrared is a direct measure of the column density of the interstellar dust, such a map can be used as a measure of extinction for extragalactic objects. As we shall argue below, the new dust map has better angular resolution and better control of systematics than possible with the reddening maps of Burstein & Heiles hereafter BH). The essence of the (1978, 1982 was to assume that variations in the dustÈtoÈneutral gas ratio can be adequately modeled by the smoothed mean background of galaxies (at northern declinations) or that this ratio is constant (at southern declinations). We make no assumptions about variations in the dust-to-gas ratio ; neither do we need make any assumption concerning the correlation of di †use ionized gas with neutral hydrogen. We only make the weaker assumption that the distribution of dust grain sizes is everywhere the same, since the relationship between UV/optical extinction and far-IR emission depends on the grain size distribution & Lee & Anderson (Draine 1984 ;Draine 1985 ;& Draine & Cardelli Guhathakurta 1989 ;Mathis 1992). The di †use gas, where our results apply, is thought to have reasonably uniform dust properties, with values of R V 4 Since neither the Burstein-Heiles nor A(V )/E(B[V ) B 3.1. our assumptions are true for all environments, a comparison between the di †erent methods is of interest.
To calibrate the extinction curve, we extensively explored the anticorrelation of counts in cells of the Automatic Plate Measuring Facility (APM) galaxy survey (Maddox et al. with 100 km emission, Ðnding a consistent 1990a, 1990b) calibration in di †erent directions of the sky. Due to complications in translating this calibration into E(B[V ), we shall defer this discussion to a separate paper et (Finkbeiner al. 1997).
Since we Ðnd that the 100 km maps are well correlated with H I emission at high latitudes, we do not expect the new extinction maps to deviate grossly from the Burstein-Heiles reddening estimates when averaged to the same scales. But many dusty regions have Ðlamentary structure with large Ñuctuations in extinction estimates over angular scales much smaller than the resolution of the BH maps. Furthermore, at low Galactic latitudes, regions such as Orion and Ophiuchus might be saturated in H I emission or partially ionized, whereas the dust remains optically thin.
Another important application for maps of the di †use emission is comparison with cosmic microwave background radiation (CMBR) Ñuctuation maps. Recent analyses of the COBE/DMR Ñuctuations et al. (Kogut 1996) show a positive correlation with DIRBE maps, particularly in the North Polar Spur region. Similar correlations are seen between the Saskatoon CMBR maps and the IRAS maps et al. et al. et al. (Oliveira 1997 ;Leitch 1997 ; Ja †e This is unexpected, since these 40 GHz CMBR 1997). experiments are not sensitive to dust emission, but are sensitive to free-free emission from the warm ionized medium (WIM ;McCullough, & Van Buren Gaustad, 1996 ;Dennison, & Topsana This implies that Simonetti, 1996). the WIM is at least partly correlated with the cold dusty medium and that our dust maps could be used as a model for both components Alternatively, the (Reynolds 1995). observed CMBR correlations might be explained by small spinning dust grains producing continuum emission in the 10È100 GHz range & Lazarian in which case (Draine 1997), the dust maps should correlate extremely well with CMBR observations over the entire sky.
This paper is organized as follows : describes the Section 2 processing of the DIRBE data. How the weekly averaged DIRBE maps are trimmed and then reconstituted into annual-average maps that are more suitable for our simple treatment of the zodiacal light removal. We Ðt a zodiacal light model by maximizing the correlation between the DIRBE maps and H I in regions of high Galactic latitude and low Ñux.
The di †use cirrus emission is expected to have a reasonably uniform temperature, since the ultraviolet radiation Ðeld that is heating the dust must be a smooth function of angular position, except within dense clouds. Fortunately, the emission at 100 and 240 km is expected to be dominated by grains sufficiently large to be in equilibrium with the radiation Ðeld (e.g., & Anderson Draine 1985 ; & Draine so that the 100 km/240 km Guhathakurta 1989), color ratio can provide a useful measure of grain temperature. In we describe how we measure the tem-°2.4 perature in the presence of substantial noise in the 240 km map. This information is used to translate the 100 km emission to dust column density.
The inclusion of the IRAS data is discussed in Before°3. we can combine the IRAS/ISSA map with the DIRBE map, we must Ðrst deal with the ISSA boundaries and discontinuities. The ISSA maps are destriped versions of the original IRAS scans, but further destriping is possible and essential for high angular resolution studies. We use a matched set of window functions that low-pass Ðlter the DIRBE map to 1¡ resolution and high-pass Ðlter IRAS/ISSA maps on the same scale. The combined map is then simply the sum of these two Ðltered maps. details the removal of Section 4 extragalactic objects and point sources from the maps, leaving behind only a map of the di †use dust emission.
In we present a simple procedure for normalizing the°5 reddening per unit Ñux density of dust emission. We correlate the residual of the B[V versus index for 389 Mg 2 elliptical galaxies against the estimated reddening and adjust the normalization until the slope of the relation is zero. The resulting calibration has an uncertainty of only 10%. We also show that the maps have an accuracy of 16% in predicting reddening, which is twice as good as the older BH procedure.
A summary of all these reprocessing procedures in presented in Readers uninterested in the details of the°6. analysis should skip directly to this point. General discussion and conclusions are presented in°°and 78 .
discusses the contamination level of stars that Appendix A remain in the maps.
describes how reddenings Appendix B and extinction in di †erent optical and infrared Ðlters can be related.
describes how the Ðnal maps can be Appendix C accessed.

T he DIRBE Data Set
The DIRBE data set is in the public domain and was obtained from the National Space Science Data Center (NSSDC) on For a complete description of the CD-ROM.1 DIRBE instrument, see et al. Both Annual Boggess (1992). Average Skymaps and maps averaged over each week of the mission are provided. The data are binned to 393,216 0¡ .32 pixels on the COBE Quadrilateralized Spherical ] 0¡ .32 Cube. (For a description of this "" skycube ÏÏ projection, see & OÏNeill As described below, we create new Chan 1975.) annual-average 25 and 100 km skymaps from the 41 weekly maps in skycube format in which the weighting function of each week is forced to be the same for both channels. We then reproject the data to standard equal-area polar projections oversampled to 512 pixels from b \ 0¡ to 90¡ (0.0250 deg2 pixel~1).
In order to preserve total Ñux to a good approximation, we initially projected to a Lambert (polar) mapping with 4 ] 4 smaller pixels that were then binned to the 0.0250 deg2 resolution. (Note that the surface brightness of the cirrus is preserved in this reprojection.) These maps for each passband b are denoted by and are in units of MJy sr~1. D b , We use the 25 km map as a model of the zodiacal light and the 100 and 240 km maps for measuring the temperature and column density of the dust.

Zodiacal L ight Removal
The DIRBE maps are severely contaminated by zodiacal emission from dust in the solar system, otherwise known as the IPD (interplanetary dust). The IPD primarily reradiates thermally at D280 K in the infrared (for l0 emissivity). Its emission per unit column density is substantially higher than that of 20 K Galactic dust : using an a \ 2 emissivity law, the 100 km emission of the IPD is larger by a factor D105 than that of Galactic dust for an equivalent column density. The brightest zodiacal emission is in the ecliptic plane, at the level of D10 MJy sr~1. Based upon our solutions, this corresponds to an extinction at the negligible level of A(B) B 10~6 mag. Thus, the removal of zodiacal light emission is critical, as it makes no measurable contribution to optical extinction.
The 25 km maps, where the IPD emission peaks, are used to model the IPD contamination at the longer wavelengths. Before removing this contamination, the DIRBE annualaverage data must Ðrst be regenerated so that the data at all wavelengths are sampled at the same time and through the same IPD. The zero points of the zodiacal contamination are constrained by H I maps at high Galactic latitude.

Dependence on T ime and Solar Elongation Angle
Removal of the zodiacal light is a difficult problem. The observing strategy of the COBE satellite resulted in a lower signal-to-noise ratio in the ecliptic plane where the zodiacal light is most prominent. The number of high-quality DIRBE observations per pixel ranged from D200 at the ecliptic equator to D800 near the ecliptic poles. The leastsampled regions were approximately cones of 90¡ opening angle centered at ecliptic (j, b) \ (100¡, 0¡) and (j, b) \ (280¡, 0¡). The DIRBE maps exhibit a striping pattern perpendicular to the ecliptic plane ; scans adjacent in the sky correspond to temporally separated instrument tracks, taken at varying solar elongation angles, through di †erent column densities of IPD.
Measurements for each week of the mission were used to construct weekly SKYMAPS. The DIRBE Annual Average Skymaps are a combination of the weekly maps, averaged by the number of times that each pixel was observed in a weekly map. Unfortunately, the whole sky is not scanned each week, and some parts of the scanned region are omitted in certain weeks as a result of the interference from the Moon and planets. Furthermore, not all the detectors were on at all times.
A cursory glance at the weekly skymaps reveals a serious problem of combining data in this way. Near the ecliptic plane the 25 km Ñux exhibits a strong dependence on solar elongation angle e. During the 41 weeks, data were typically taken between e \ 64¡ and 124¡, and the 25 km Ñux varies by a factor of 2 over this range. Comparison with the 100 km IPD Ñux also reveals a strong temperature gradient with respect to e. This is easily explained by assuming a radial gradient of IPD density from the Sun, because a line of sight passing close to the Sun must traverse more (and hotter) dust than a line of sight at larger solar elongation. As a further complication, because the 10 passbands were not always observed simultaneously, the weekly skymaps of the weighting function do not always match across all channels. Contributions to the Annual Average Skymaps were weighted according to the number of observations made in each channel, in spite of the strong dependence on elongation angle.
In cases where the measurement noise is overwhelmed by the time-dependent systematic variations, this method of averaging is inappropriate. The weights assigned to sky pixels are thus not always the same for the 25 and 100 km weekly maps, leading to weighted Annual Average Skymaps that cannot be directly compared to each other. To rectify this, we recombine the 41 weekly skymaps using the same weight for the same pixels in di †erent passbands. We use the error maps for the 100 km passband as weights in all other passbands. Because the IPD temperature gradient is strongest at low e, we further delete all data in each weekly map taken at e \ 80¡. The resulting maps have noise and artifacts that are somewhat worse than the Annual Average Skymaps from the NSSDC ; however, the artifacts are the same at both wavelengths. The resulting 25 km map is appropriate for modeling the IPD contamination at other wavelengths. The spatial-temporal variations of zodiacal light are complicated and difficult to model analytically. Our pro-Vol. 500 cedure for removing the zodiacal light assumes that the 100 km map correlates linearly with H I at high Galactic latitudes and low Ñux levels et al. Note that (Boulanger 1996). one should also expect some dust within the extensive ionized H`zones and that the column density ratio N(H`)/ N(H I) has considerable scatter in di †erent directions The relationship between the di †use (Reynolds 1990). ionized and neutral hydrogen in any direction is complex some of the H`is associated with neutral (Reynolds 1995) ; gas, while other ionized clouds adjoin neutral regions. We are, of course, interested in measuring the dust associated with both components, and to the degree that there is dust within H`regions not associated with neutral regions, the scatter of dust and H I will be increased. The degree of scatter between neutral gas and dust is therefore of considerable interest, but this scatter should not bias our procedure.
We consider the 25 km map as a template of the unwanted zodiacal foreground, as the 25 km channel is so thoroughly dominated by the IPD. In a few regions where the Galactic cirrus contributes more than a few percent to the 25 km emission, we interpolate the 25 km map in bins of constant ecliptic latitude. To Ðrst order, we subtract scaled versions of the 25 km map in order to minimize the scatter between the H I and zodiacal-corrected 100 and 240 km emission. We further reÐne this correction with the addition of a quadratic term to account for temperature variations in the IPD.
Fitting blackbody functions to the zodiacal light from 12 to 60 km, we Ðnd that the color temperature does not vary by more than 10% from ecliptic equator to the poles. This allows us to make a Ðrst-order correction under the assumption that the IPD is at a constant temperature. The DIRBE 25 km map, is directly scaled to model the D 25 , zodiacal contamination at longer wavelengths. The raw maps for these other DIRBE passbands, The coefficient determines the scaling of the zodiacal A b emission from the 25 km DIRBE passband to passband b. The zero-point o †set is used to account for a multitude B b of nonzodiacal contributions to the 25 and 100 km maps, either from the Galaxy or from extragalactic light. Although the 25 km map is relatively free of point sources, the 10 brightest sources have been masked to their local averages (see Note that these sources were not excluded  in each passband. The errors in this relation are systematic and difficult to express analytically. Therefore, we take the uncertainties to be proportional to the H I Ñux. The r i Ðtting procedure is performed on the maps binned to 2¡ .5 and limited to those areas of the sky with low emission, H I \ 2 0 0Kk ms1 or N(H) \ 3.7 ] 1018 cm~2 (19% of the sky). Two dust clouds with unusual temperature and the region immediately around the very bright source NGC 253 are omitted from the Ðts. The results of these Ðts for b \ 100, 140, and 240 km are presented in and a Table 2, scatter diagram is shown in both (a) before and (b) Figure 1 after the correction.
Temperature variations in the IPD make the linear correction of inadequate for our purposes. Based equation (1) on 12 km/25 km and 25 km/60 km ratios, the color tem-  perature of the IPD varies by 5% for an a \ 1 emissivity law or by 10% for an a \ 2 emissivity law. Because the 100 km band is on the Rayleigh-Jeans side of the spectrum, this results in a 5% or 10% error in zodiacal light. As the zodiacal light contributes up to 10 MJy sr~1 at 100 kmi nt h e ecliptic plane, this results in absolute errors of 1 MJy sr~1, with a strong ecliptic latitude dependence. A constanttemperature linear model of zodiacal dust removal is a good Ðrst approximation, but it is clearly too simplistic. Because the neighboring passbands at 12 and 60 km are a complex mix of zodiacal and nonzodiacal contributions, it is desir- able to model the variation in IPD temperature using only the 25 km map, rather than some linear combination of passbands. Thus, we are driven to higher order Ðts using only the 25 km map.
The temperature variations in the IPD are most strongly a function of ecliptic latitude b. An adequate second-order function for modeling the variations in IPD emission at longer wavelengths includes a term that is linear in the product of and the mean 25 km emission at each The above is referred to as the quadratic correction. Solutions are again found via using the same region equation (2) of the sky as before. Results are presented in Since Table 2. the linear and quadratic contributions are not orthogonal, there is no particular physical meaning to the coefficients derived in the quadratic Ðts ; they are merely the coefficients that yield the minimum s2.
The result of the quadratic correction is shown in Figure  At 100 km the quadratic correction is clearly superior to 1c. the linear correction, reducing the rms scatter from 19% to 16% (for H I \ 2 0 0Kk ms1). At 240 km the di †erence between the two Ðts is negligible. This is both because the 240 km map has a higher inherent noise and because the zodiacal light correction at such long wavelengths is smaller. Even though there is no obvious di †erence in Ðtting methods at 240 km, we apply the quadratic correction to both passbands in order to avoid introducing systematic errors into the determination of the dust temperature. In principle, the Ðt should be made between the H I gas and the total column density of dust. The total column is a complicated function of the 100 and 240 km maps, and the zodiacal light coefficients for both those maps need be Ðtted simultaneously. This nonlinear Ðt with six parameters is not stable, primarily because the DIRBE 240 km signal-tonoise ratio is too low at high latitudes. For this reason, we have performed the Ðt independently for the 100 and 240 km maps. As long as the temperature of the dust does not vary signiÐcantly and systematically with ecliptic latitude, then the Ðt coefficients for our zodiacal model should be valid. As we discuss in variations in the Galactic dust°2.3, temperature at the high latitudes used in these Ðts are hardly measurable above the noise. This least-squares minimization procedure is used only to remove the zodiacal foreground. Our measure of the best removal is minimal scatter between the far-IR dust emission and 21 cm emission. There could well be an additional component of dust that is not correlated with the H I emission, such as dust associated with di †use H`regions. In such a situation, our procedure for zodiacal foreground removal should remain unbiased, but the interpretation of the Ðt coefficients would be di †erent. The slope would signify A b the amount of dust associated with neutral gas mixed with ionized gas, while the o †set would be the dust associated B b with instrumental o †set, the cosmic infrared background (CIB), and dust emission associated with a uniform sphere of di †use H`.
This method of zodiacal light correction is by no means deÐnitive. A more thorough examination of the problem is currently under way by the DIRBE team et al. (Kelsall They are performing a detailed modeling of the IPD 1998). in three-dimensional space that includes the dependence of dust temperature and density on the distance from the Sun, as well as speciÐc dust bands, and the inclination of the dust cloud with respect to the EarthÏs orbit. However, our model is adequate for the present analysis. In the worst case, assuming that all of the residual scatter in the corrected 100 km plots at H I \ 1 0 0Kk ms1 is caused by imperfectly subtracted zodiacal light, the rms of this error is 0.05 MJy sr~1. This corresponds to an error of 0.004 mag in A(B).

2.3.
Zodiacal Removal to Isolate the CIB As a digression at this point, the question of the CIB was a major motivating factor for the DIRBE experiment. The extraction of the CIB has proven to be a considerable challenge and to date, CIB detections from (Hauser 1996), DIRBE data have been reported as upper limits. The regression analysis described above is one method to approach such a measurement, provided that the zodiacal model is adequate. The coefficient B of the regression Ðts is a constant term that remains after the zodiacal foreground dust and all Galactic dust correlated with H I emission is removed. This term is not likely to be an instrumental o †set, since DIRBE had a cold-load internal chopper, and the entire telescope was cold. A signiÐcant detection of B is either a measure of the CIB or a result of some unknown instrumental artifact.
Since the zodiacal dust is conÐned to a thin zone in the ecliptic plane of the solar system, along directions away from the ecliptic plane, all the zodiacal foreground is approximately 1 AU from the Sun. It is therefore reasonable to assume the zodiacal foreground emission of our reconstructed yearly maps to be the same temperature in all directions for sufficiently high o b o. Thus, a linear model should be adequate for removal of the zodiacal emission. However, one cannot look strictly at the ecliptic poles, since the 25 km maps vary only by a factor of 3.5 from the ecliptic plane to the pole ; a suitable lever arm in b is therefore essential for decoupling the isotropic CIB from the smoothly varying zodiacal foreground.
We have performed linear regression analyses for the 100, 140, and 240 km maps as a function of the lower limit on ecliptic latitude o b o using the procedures described in°T he results are listed in The quoted errors 2.2.2. Table 3. represent the formal 95% conÐdence intervals and do not  1995). of this temperature range, the emission at 100 km will di †er by a factor of 5 for the same column density of dust. Therefore, it is crucial to use temperature information to recover the dust column density. There are three useful DIRBE bands for computing the Galactic dust temperature, centered at 100, 140, and 240 km. The median signal-to-noise ratio per pixel for these maps at o b o [ 10¡ is 164, 4.2, and 4.9, respectively. Because of the high S/N at 100 km, we will use that map as a spatial template and multiply it by a temperature correction factor derived from the 100 km/240 km ratio. The 140 km map is not used because of its large noise level compared to the nearby 100 km map. We attempt no temperature correction in regions of very low emission, since our procedure will introduce noise, and the exact temperature of the dust in these dilute regions is less important.

FIR Spectrum of the Dust
Estimation of the dust color temperature and column density is complicated by our lack of knowledge about dust opacities and temperature distributions. Our procedure works only if the large grains responsible for the emission at 100È240 km are in thermal equilibrium with the interstellar radiation Ðeld, and the smaller, transiently heated dust grains make a negligible contribution to emission at these wavelengths. Fortunately, these assumptions appear to be in agreement with our current understanding of interstellar dust emission & Draine et (Guhathakurta 1989 ;Sodroski al. 1997).
Because the ISM is optically thin in the FIR, the intensity is given by where T is the grain temperature, o is the mass density, is B l the Planck function, is the opacity, and the integral is i l over the physical path length. Since the dust particles are small (radii a \ 0.25 km) compared to FIR wavelengths, the opacity does not depend upon the details of the particle i l size distribution. However, the opacity does depend upon the nature of the material. In general, the opacity follows a power law in the FIR, The & Lee model predicts that the FIR emis-Draine (1984) sion is predominately from graphite with aB2.0. Layerlattice materials, such as amorphous carbon, have opacities aB1.0, and silicates have aB1.5. We have chosen a \ 2.0 in deriving our dust temperatures. Though this choice directly a †ects the temperature estimation, we show that our Ðnal dust column densities are completely insensitive to it. The reason for this insensitivity is that the DIRBE passbands are sufficiently near the Wien tail of the dust spectrum.
"" Classical ÏÏ grains with radii a [ 0.05 km exposed to ambient starlight retain an equilibrium temperature. In we assume such an equilibrium and Ðt each line of°2.4.2 sight with a single color temperature. However, each line of sight through the Galaxy may pass through several regions at di †erent equilibrium temperatures. By ignoring multiple temperature components, we systematically underestimate the true column density. We have modeled this e †ect by Ðtting a single color temperature to emission from 18 K dust added to emission from another region at temperature For di †erent choices of and its mass fraction we T B .
, compare the recovered dust column to the true column density For the range K, the true (Fig. 2).
15 \ T B \ 21.5 column is at most underestimated by 10%. A factor of 2 error would result only in contrived circumstances with K or K. As the majority of the sky spans [ 33 only a 5¡ range of temperature, our single-temperature model is deemed satisfactory.
Another constituent of the dust that we have ignored is very small grains (VSGs), with radii km. These a [ 0.005 FIG. 2.ÈRatio of recovered vs. true column density of dust using a single-temperature Ðt to two components. A fraction of dust at temf B perature is added to 18 K dust. The recovered column density is always T B lower than the true column density, with contours spaced in units of 0.1. VSGs are stochastically heated by the interstellar radiation Ðeld (ISRF) and cannot be characterized by a single temperature. VSGs undergo a sudden rise in temperature following absorption of a visible or UV photon, followed by a nearly continuous decline in temperature as the heat is radiated away in the form of many FIR photons. Such grains are assumed to be responsible for the stronger than expected Galactic cirrus emission observed by IRAS at 12È60 km. The contribution from VSGs at the wavelengths that we use to Ðt the temperature (j[100 km) is small & Anderson & Draine (Draine 1985 ;Guhathakurta 1989). Thus, the VSGs do not spoil our derivation of the column density due to classical grains. VSGs would be signiÐcant to our dust maps only if their density were much higher than expected and their distribution di †erent from that of classical grains in the di †use ISM.

Recovering a Temperature Map
The 100 and 240 km maps are too noisy to recover a reliable and independent dust color temperature at each DIRBE pixel. In the regions of low S/N a Ðltering scheme is required. One option is to smooth the maps with a large beam, using the resulting maps to construct a heavily smoothed temperature map. This simple technique has the disadvantage of throwing away small-scale information in the compact dusty regions where the S/N ratio is high. The spirit of most Ðltering methods is to produce a map with a fairly uniform S/N ratio. An optimal scheme would smooth the data where there is no signal (and very little dust) and leave the data unsmoothed in the regions with large signal (and more dust). However, such schemes (e.g., the minimum-variance Wiener Ðlter) have the disadvantage of overweighting low-S/N pixels near high-S/N pixels in such a way that bright sources have noisy halos.
One way to avoid artifacts such as halos around bright sources is to Ðlter the maps on a local basis at each pixel. We assume that regions of very little dust have a uniform temperature. This allows us to combine the DIRBE maps with background averages taken at high Galactic latitudes, weighting in such a way that we preserve S/N.
To generate the temperature correction map from the zodiacal lightÈcorrected DIRBE maps, we Ðrst remove 13 bright point sources from the DIRBE images (including M31, the LMC, and the SMC ; see We smooth Table 1). both the 100 and 240 km maps with a Gaussian Ðlter of which determines the resolution of the FWHM \ 1¡ .1, resulting temperature map.
The dust e †ective color temperature is recovered from the ratio of intensities at 100 and 240 km. Using the Ðltered DIRBE maps from we deÐne the ratio map equation (7), where Each Ðltered map is a weighted average of the local D b S (Gaussian-smoothed) DIRBE Ñux and the background average Ñux determined at high Galactic latitude. The background average Ñux is determined separately in the D1 b Q north and south Galactic hemispheres. The weight function W determines how to Ðlter the resulting R between the T ), Ðtted using K b (a, eq. (11). Fits assume an a \ 2 emissivity model. local quantity and the background average. Where W is small, the resulting R will have the high-latitude background value, as appropriate in regions of low S/N. In regions of high S/N the weight function approaches unity, and R reduces to the local value of the Ñux density ratio. The advantages of this method are that it has no leakage from neighboring pixels and that the scale of the smoothing is constant. Therefore, noise generated by the ratio of two small uncertain numbers is suppressed. The background averages are assessed in each hemisphere by simply averaging the 100 and 240 km Ñux The ratios are 0.661 in the north and 0.666 in the south.
We determine the weight function W by a minimumvariance analysis of the resulting ratio map R. For this purpose we use a map of the standard deviation of measured 240 km Ñux at each point reduced by the e †ecr 240 , tive binning of the Gaussian smoothing of the maps. For the 240 km map the DIRBE errors are in the range 0.5 \ MJy sr~1, which is much larger than for the 100 r 240 \ 2.0 km map. Because the detector noise is negligible for the 100 km data, we construct a weight function that minimizes the variance in the reciprocal of the color ratio d2(1/R), where is an assumed mean color ratio. We use R T R T \ 0.55, but the results are insensitive to this choice. Minimization of this function with respect to W leads to a quadratic equation whose solution is a map of the weight function The resulting W-function is close to the simple expression for typical noise levels in the DIRBE To convert from the ratio map R to e †ective color temperature, we need to know the emissivity model for the dust and the frequency response of the DIRBE instrument v l passbands W (l). This information is combined to form a color correction factor for a passband b, using an Appendix B of the Explanatory Supplement DIRBE (1995) has tabulated for all their bands in the domain K b (a, T ) 0 \a¹2 and 10 ¹ T ¹ 2 ] 104 K. These color correction factors are well Ðtted by a functional form Coefficients for these Ðts at a \ 2.0 are presented in Table 4. The Ðts are accurate to 1 % at all temperatures. The measured Ñux in passband b, of dust emitting thermally with D b , a la emissivity is where is the actual intensity at frequency b. I b Calculating the ratio R(a, T ) that DIRBE would measure for a given temperature and emissivity model, we have We interpolate the function for 3000 temperatures K b (a, T ) and invert yielding T (R). The emissivity equation (13), model a \ 1.5 yields a color temperature D2 K hotter than a \ 2.0, but the e †ect of this on the recovered dust column density is not signiÐcant, aside from an overall multiplicative normalization. Aside from this normalization, the Ðnal column density maps vary only at the 1% level for an a \ 1.5 versus an a \ 2.0 emissivity model. For all calculations that follow, we have chosen a \ 2.0, in agreement with the & Lee dust model. The resulting Draine (1984) temperature map is shown as The light regions on Figure 3. the temperature map denote warmer dust, while the darker regions denote cooler dust where the 100 km Ñux underestimates the column density of dust. Note that most of the sky is a neutral grey color, while several prominent highlatitude molecular cloud regions show up as dark Ðlaments. A few isolated white (hot) spots correspond to the LMC, SMC, and O star regions near Ophiuchus and Orion.
We compute the column density of dust as the amount of emission we would expect in the DIRBE 100 km band if the dust were all at a reference temperature of K. The T 0 \ 18.2 column density DT is expressed as the 100 km Ñux multiplied by a temperature correction factor X : where A temperature map T is recovered with the function, The maximum deviation of these Ðts is 0.8% in the domain 10 \ T \ 32 K. The column density map DT, recovered from is our model for the Galactic dust at equation (14), DIRBE resolution.

FOLDING IN IRAS RESOLUTION
The infrared maps from the COBE satellite have the advantages of being well calibrated and well corrected for zero-point drift. However, with a resolution of FWHM, 0¡ .7 the COBE/DIRBE maps utterly fail to resolve important Ðlamentary details in the cirrus and detect only the very brightest point sources. Therefore, we resort to the use of the older IRAS satellite data to improve upon this limited angular resolution.
The IRAS satellite observed 97% of the sky in four passbands, detecting some 250,000 point sources, including D20,000 galaxies et al.
In addition to (Neugebauer 1984). these point sources, the satellite discovered the infrared cirrus, especially visible in the 100 km band et al. (Low As the instrument was designed for point source 1984). detection (di †erential photometry), it is far from optimal for the absolute photometry required to make sense of the cirrus. In particular, the zero point from one scan to the next drifts considerably, leaving stripes with power on a range of scales, from a few arcminutes to several degrees.
The IRAS Sky Survey Atlas (ISSA ; et al. Wheelock 1994) released by IPAC in 1991È1994 is the most useful presentation of the IRAS data for our purposes. The ISSA sky is divided into 430 plates and spaced roughly 12¡ .5]12¡ .5 every 10¡ in right ascension and declination. The ISSA maps su †er from the zero-point drifts as well as residual zodiacal light. Despite the best e †orts at IPAC to remove striping artifacts and zodiacal light, these features are still quite obviously present. This limits the usefulness of these maps for many applications.
We have reprocessed the ISSA data to create a 100 km map with the following properties : 1. ISSA plates are destriped, reducing the amplitude of the striping artifacts by a factor of D10.
2. A deglitching algorithm was employed to remove small-scale artifacts that are not conÐrmed on other IRAS scans.
3. IRAS missing data areas were Ðlled in with DIRBE data.
5. Angular resolution is close to the IRAS limit with FWHM \ 6@ .1.
6. Extragalactic objects were removed to a Ñux limit of The combination of our destriping algorithm and the DIRBE zero point in e †ect removes zodiacal contamination from the ISSA maps. discusses our Fourier des-Section 3.1 triping algorithm. Removal of deviant pixels, including asteroids, is described in describes the°3.2. Section 3.3 method used to combine DIRBE and IRAS data to form our Ðnal maps. The removal of 20,000 point sources is the topic of°4.
3.1. Destriping the ISSA Maps IPAC took steps to destripe the ISSA plates, including a zodiacal light model and local destriping algorithm. However, serious striping remains. Fortunately, the scanning strategy of the IRAS satellite gives sufficient information to remove most of the striping artifacts that remain. et al. have made important progress toward Cao (1996) destriping the maps and deconvolving the IRAS beam reponse in small regions of the sky. However, due to the complexity of their algorithm and large amount of computing time required, it is impractical at this time to apply such a method to the full sky. We have developed an algorithm that is straightforward to implement and can easily be applied to the full sky.
We consider each ISSA plate separately and manipulate it in Fourier space. One ISSA plate and its Fourier transform are seen in Figures and where zero wavenumber 4a 4b, is at the center. The rays emanating from the center are a clear signature of the striping in real space. The problem then becomes one of removing this excess power in Fourier space that is associated with the real-space stripes. Because the IRAS satellite scanned most of the sky 2 or 3 times, usually at a di †erent angle, the contamination in Fourier space occurs at di †erent wavenumbers for each scan. Thus, the contaminated wavenumbers from one scanning angle can be replaced with wavenumbers from the other scanning angles. The steps involved are (1) identifying the contaminated wavenumbers and (2) an optimal replacement strategy that allows for the lack of full coverage for many of the scans.
Because the stripes are radial in Fourier space, we parameterize the Fourier domain in polar coordinates for (k r , k h ) the purposes of this discussion, although it is discretely sampled in and for calculations. k x k y For each ISSA plate there are 3 HCONs (hours-con-Ðrmed scans), referred to as HCON-1, HCON-2, and HCON-3. The composite map generated from these is called HCON-0, but when we refer to "" the HCONs,ÏÏ we mean HCONs 1, 2, and 3. Many of these individual HCONs contain data with striping in only one direction. In general, however, a single HCON image may contain data taken in multiple scan directions and signiÐcant regions of missing data. In many cases HCON-3 contains little or no data. For each HCON map there is a corresponding coverage map, which was used by IPAC to compute a weighted average of the three HCONs for each plate. (These coverage maps were kindly provided to us by S. Wheelock). The Explanatory Supplement also reports the (Wheelock 1994) use of a local destriper, but it appears to have had negligible e †ect in low-Ñux regions of the sky.

Criterion for Bad W avenumbers
With each plate we begin by generating a composite map from the three HCONs, using the three coverage maps for weights. This is equivalent to the HCON-0 composite maps released with the ISSA, except that no local destriper has been applied, and asteroids and other artifacts have not been removed.
Because our method is based on Fourier transforms where bright point sources may overwhelm many modes, we perform a crude point-source removal before any analysis. This consists of selecting objects or regions that are more than 0.7 MJy sr~1 above the median-Ðltered image. Several tests of this method indicate that it masks all signiÐcant point sources but not the striping. These sources are then replaced with the median of an annulus around them. Because these sources are only removed when identifying contaminated modes, the details of this process are unimportant.
This composite HCON-0 image (without point sources) is smoothed with a 6@ FWHM Gaussian to produce a baseline image that is used to Ðll in missing pixels in each of the three HCONs. This smoothing is enough to suppress most of the striping, and the replacement of missing pixels allows us to Fourier transform completely Ðlled square images. Before transforming, we subtract the mean value from each image, apodize the edges, and embed the 500 ] 500 image in a 1024 ] 1024 array (which is equivalent to demanding isolated boundary conditions). The transform for a typical plate is shown in Figure 4b.
It is necessary to deÐne a criterion with which to select the contaminated wavenumbers. We denote the fast Fourier transform (FFT) of HCON-n as and a particular element F n of k-space as We bin the Fourier domain into 90 F n (k r , k h ). azimuthal bins, each 2¡ wide in A background power as k h . a function of is determined by Ðrst Ðnding the median at k h each h and then median Ðltering the result. We deÐne a power ratio as the ratio of the power in each bin to the c h k h background power. By normalizing in this way, we expect to be near unity, except in angular bins that are contamic h 1 10 100 deg -1 1 P k (destripe)/P k (raw) 536 SCHLEGEL, FINKBEINER, & DAVIS Vol. 500 nated by stripes. We somewhat arbitrarily choose c h \ 1.2 to be "" good ÏÏ power and to be "" bad.ÏÏ Between c h [ 1.6 these two cuts, the information is deemed acceptable for retention, but poor enough to avoid replacing other "" bad ÏÏ power with it. In this way, "" good ÏÏ power is used where available, and we avoid corrupting real structures by not discarding power of ratio less than 1.6.

T he Destriping Algorithm
Each HCON is destriped separately before they are combined to an averaged map. Because of the incomplete coverage in some HCONs, it would not be optimal to combine the images in Fourier space. Instead, for each HCON-n we create a stripe image which is removed in real space. In S n , this way we may utilize the good data in each HCON, even if only a fraction of the plate is covered.
A composite of "" good ÏÏ power, is generated by F good , averaging all power at each with Any mode (k r , k h ) c h \ 1.2. with no "" good ÏÏ power is set to zero. (Modes with wavelengths larger than 1¡ are not changed to avoid discreteness problems.) A bad pixel mask is produced for each HCON, indicating which regions of each are contaminated. The F n di †erence between and for bad modes only is F n F good inverse transformed to obtain the stripes corresponding S n , to HCON-n. This stripe map is subtracted from each raw HCON to produce individual destriped HCONs. Note that only stripes have been removed, and the destriped images retain all of the original point sources and other real structures. These destriped images are then deglitched and (°3.2) averaged, weighted by the coverage maps, to obtain the Ðnal HCON-4 images.
shows the one-dimensional power spectrum of Figure 5 four plates before and after the destriping process. The destriping has reduced the high spatial-frequency variance by nearly a factor of 2 in some high Galactic latitude areas. In cases where the point-source subtraction fails, nothing is done to the plate. This occurs only at low Galactic latitude, where the striping artifacts are overwhelmed by cirrus emission. We have not found stripes to be important for the generation of a reddening map in these areas, so we have done nothing about them.

Deglitching the ISSA Maps
In addition to striping, the ISSA images contain other artifacts that must be removed. These anomalous features include transient sources, such as asteroids, and detector glitches that should have been removed from the timeordered data stream. Images with two or more HCONs allow identiÐcation of these artifacts by looking for discrepant pixels. We generically term this process "" deglitching.ÏÏ IPAC chose to deglitch the ISSA images by visual inspection. This postproduction step removed anomalies from individual HCONs before combining them to produce HCON-0 images. Thus, HCON-0 di †ers from a coveragemap weighted average of the HCONs wherever anomalies were removed. Rather than rely upon visual inspection, we implement an automated deglitching algorithm.
The Ðrst task is to identify discrepant points. We compare each destriped HCON to a local median, Ðltered on a 10@ scale. Because some ISSA pixels have better coverage than others, they have lower noise. Therefore, we set the threshold for a deviant pixel at 2/N1@2 MJy sr~1, where N is the number of times IRAS observed that pixel. The pixels that deviate from the median-Ðltered background by more than this threshold are Ñagged, as are their immediate neighbors. Pixels Ñagged in all available HCONs are deemed to be "" conÐrmed sources.ÏÏ These include real point sources, knots in the cirrus, and also pixels near regions of no data. In all these regions, asteroids cannot be di †erentiated from features that should remain in the map. However, pixels that are Ñagged in at least one, but not all, HCONs are Ñagged as "" glitches.ÏÏ These pixels and their neighbors are replaced with data from the other HCONs. The reader should note that cirrus features are not modiÐed by this algorithm.
Our deglitching algorithm a †ects less than 0.1% of the sky. Three planets were scanned by IRAS and not adequately removed from the ISSA plates. These require special attention, as Saturn in particular is enormously bright compared to the background Ñux. Neptune and Uranus moved sufficiently between scans that they can be removed in one HCON while retaining coverage in the other HCONs. Because Neptune appears close to the Galactic plane, its surface brightness is comparable to the background level of D100 MJy sr~1. We remove it in each HCON with a circle of radius Uranus is only incompletely removed from the 7@ .5. ISSA plates, still causing hysteresis in the in-scan direction. We remove areas of 220@ in the in-scan direction and of 94@ in the cross-scan. Saturn is problematic, since it appears only in HCON-1 and HCON-2, and it moves very little between those scans. Because of this, we excise a circle of radius 2¡ in all HCONs. This area, centered at (l, b) \ is treated like the exclusion strip and (326¡ .28, ]51¡ .66), Ðlled with DIRBE data (see The planet removal is°3.3). implemented before the destriping algorithm and the deglitching procedure, afterward.

Combining ISSA with DIRBE
The ISSA and DIRBE data are combined in such a way as to retain the small-scale information from IRAS and the large-scale calibration of DIRBE. The ISSA HCON-4 maps are processed to match the color response and zero point of the zodiacal-subtracted DIRBE 100 km maps on scales larger than 1¡. Regions with no IRAS data are Ðlled with DIRBE-resolution data. The maps are Ðrst projected into two 4096 ] 4096 pixel polar projections, one for the north and another for the south.
The ISSA maps are Ðrst multiplied by a constant value C to correct the IRAS gain approximately to that of DIRBE. The relative gain factors between DIRBE and IRAS are difficult to assess exactly, owing to the drifts in IRAS zero points. The Explanatory presents a DIRBE Supplement preliminary linear transformation between IRAS and DIRBE data based upon carefully selected regions (in an obsolete version of the supplement, but reprinted in et al. Table IV.D.1). They Ðnd the IRAS Wheelock 1994, brightness levels to be too high by 38% at 100 km. We Ðnd the IRAS calibration to be too high by this value for the cirrus, yet it is consistent with DIRBE for bright point sources. We choose to recalibrate the IRAS data globally with a compromise value of C \ 0.87. We also convolve the ISSA images with a Gaussian, FWHM \ 3@ .2 W G(3@ .2), bringing the e †ective IRAS smoothing to 6@ .1.
In order to correct the IRAS zero point to that of DIRBE on scales greater than 1¡, we construct a di †erence map between the two on these scales. This di †erence map S is added to the destriped IRAS map to yield a Ðnal map I des containing the small-scale information from IRAS with the large-scale calibration of DIRBE (after zodiacal correction) : The di †erence map must be taken between the IRAS and DIRBE data Ðltered to the same point-spread function. The procedure is complicated by the fact that neither the IRAS nor DIRBE beams are Gaussian. The DIRBE point-spread function is approximately 42@ ] 42@ with power-law tails. Fortunately, for most of the sky, repeated scans from di †erent directions result in an averaged DIRBE response approximated by a circular top hat. Therefore, the ISSA map is convolved with a circular top hat of radius 21@, while still in the pixel plates, and is then repro-W (21@), 1@ .5 jected to a polar projection, introducing the same distortions that exist in the DIRBE projections. At this point, both maps are further smoothed by a FWHM \ 40@ Gaussian to obtain a Ðnal map with The di †er-FWHM \ 1¡ .00. ence map is where DQ represents the 100 km DIRBE map after zodiacal correction The rms of the di †erence map is 0.9 (eq. [3]). MJy sr~1 at o b o [ 30¡, which is representative of the zeropoint drifts in the IRAS/ISSA maps.
Due to the di †ering color responses, the residual of the smoothed ISSA and DIRBE maps has a few high and low regions near very bright sources. Other than NGC 253, these are all in the Galactic plane, the LMC, or the SMC. We mask these regions in the di †erence map to the median of an annulus around them (see This treatment is Table 1). not meant to be strictly correct and introduces substantial absolute errors relative to the rest of the map. However, on top of the sources the fractional errors are not signiÐcant, and this procedure eliminates halos and recovers the correct zero point nearby the sources.
Regions of missing ISSA data must be Ðlled with DIRBE data in a way that treats the boundaries of the missing ISSA data properly. Both steps of the ISSA smoothing process are done in the following way : A mask, set to 1 for good pixels and 0 for missing data, is smoothed in the same way as the data. The missing data is zero Ðlled so that it does not contribute to the smoothed result. After smoothing, mask pixels with values less than 0.5 are discarded, and the remaining pixels are properly weighted by dividing by the value in the smoothed mask. The mask generated by the second ISSA smoothing is applied to the DIRBE data set as well, so that in both maps pixels near a boundary only contain information from the "" good ÏÏ side of the boundary. Point sources just barely on the "" bad ÏÏ side should contribute equal tails to the "" good ÏÏ side, eliminating potential problems with the zero calibration.
The column density of dust that radiates at 100È240 km is recovered with where we have replaced the DIRBE 100 km Ñux with the corrected IRAS 100 km Ñux in The tem-equation (14). perature correction map X is necessarily only at the DIRBE resolution of D1¡, although the reprocessed IRAS map has a Ðnal resolution of 6@ .1.

REMOVING POINT-SOURCE AND EXTRAGALACTIC OBJECTS
For the purposes of an extinction map, only the Galactic infrared cirrus is of interest. Our dust maps are meant to trace the di †use emission from dust as well as from localized clumps, such as Bok globules, which are typically on the scale of a few arcminutes. The stars, planetary nebulae, and other point (unresolved) sources in the ISSA must be eliminated. We also remove veriÐed extragalactic objects down to a given Ñux limit. The resulting maps should be free of all contaminating sources and galaxies at o b o [ 5¡ and most such sources in select regions at lower latitudes.

Strategy
Our goal is to replace contaminating sources with the most likely value of the underlying 100 km emission. The most obvious solution is to subtract the Ñux multiplied by the PSF for each source. However, the shape of the PSF is neither Gaussian nor circularly symmetric, and its shape depends upon the IRAS scan directions at each position on the sky. Because of these uncertainties and because many extragalactic objects are not point sources anyway, we replace sources with a median value from the surrounding sky. For point sources, we replace pixels within a radius of either (for Jy) or (for Jy). For 5@ .25 f 100 \ 10 7@ .5 f 100 º 10 extended sources, we replace pixels within the measured radius plus The median is always taken from a sur-1@ .5. rounding annulus with the same area as its interior. In total, 1.2% of the ISSA pixels at o b o [ 5¡ are Ñagged as sources and replaced. The LMC, SMC, and M31, and sources in their vicinity are not removed. At latitudes o b o \ 5¡ we remove galaxies and stars only in those parts of the sky deemed unconfused by the PSCZ galaxy survey (W. Saunders 1997, private communication).
Sources are not identiÐed from the IRAS 100 km maps because of the confusion with cirrus. Instead, sources are primarily identiÐed from the 60 km band of the Point IRAS Source where confusion with cirrus is less of Catalog (PSC), a problem, and the spatial resolution is somewhat better. completed a redshift survey of IRAS galaxies to (1995) a Ñux limit of 1.2 Jy at 60 km over most of the sky. The PSCZ redshift survey (W. Saunders 1997, private communication) covers a somewhat smaller area of sky to a lower Ñux limit of 0.6 Jy.

Extragalactic Sources
Many nearby galaxies are resolved in the ISSA maps and cannot be treated as point sources.
et have studied Rice al. those galaxies with blue-light isophotal diameters greater than 8@ and report improved total Ñux densities from the IRAS data. The 70 large galaxies with a total Ñux at 60 km exceeding 0.6 Jy are removed from our maps as extended objects (excepting the LMC, SMC, M31, and M32, due to its proximity to M31). These objects are removed using a radius equal to their major axis, isophotal diameter at 1 2 B \ 25 mag arcsec~2, plus (as tabulated by Vaucou-1@ . 5 de leurs The radius of NGC 253 is increased to 36@ to 1991). remove IRAS hysteresis artifacts near this extremely bright source.
The IRAS 1.2 Jy Galaxy Survey identiÐes most infraredemitting galaxies at o b o [ 5¡ down to a Ñux limit of 1.2 Jy at 60 km et al. Fluxes for extended sources were (Fisher 1995). systematically underestimated by the PSC and were improved with a process called ADDSCANing (Strauss, Davis, & Huchra The galaxy survey employed a 1990). color selection, which very efficiently discrimif 60 2 [ f 12 f 25 , nates between galaxies and stars. Each candidate source was classiÐed based upon visual inspection of POSS and ESO plates, or upon inspection of CCD images if the plates were inconclusive. Planetary nebulae were identiÐed spectroscopically. The survey identiÐes 5320 galaxies with f 60 [ 1.2 Jy. In addition, 444 point sources made the Ñux and color cuts : 98 extragalactic H II regions, 210 stars, and 136 planetary nebulae. All these galaxies and point sources are removed from our maps, unless they coincide with the large et galaxies already removed, or with the LMC, (Rice al.) SMC, or M31. We ignore the sources classiÐed as "" cirrus or dark cloud,ÏÏ "" unobserved,ÏÏ or "" empty ÏÏ Ðelds, and two reÑection nebulae.
The PSCZ redshift survey covers nearly the same region of sky as the 1.2 Jy survey, but to a Ñux limit of 0.6 Jy at 60 km. This survey employed an additional color cut, which retains approximately 97% of galaxies. f 100 /f 60 \ 4, A list of PSCZ identiÐcations was provided by W. Saunders (1997, private communication) before publication. This list includes 15,285 galaxies, many of which are duplicates from the 1.2 Jy survey. These galaxies, plus an additional 455 objects measured as unresolved by the ADDSCAN process, are removed from the dust maps.
The 1.2 Jy and PSCZ surveys did not have complete sky coverage. Two strips in ecliptic longitude (the exclusion strip) were not observed by the IRAS satellite, amounting to 4% of the sky. Other regions, primarily at low Galactic latitude, are too confused by cirrus emission to allow reliable source identiÐcations. These galaxy surveys divided the sky into approximately 1 deg2 "" lune ÏÏ bins, from which masks were constructed. The 1.2 Jy survey masks those bins at o b o \ 5¡, in the exclusion strip, or with more than 16 sources per deg2. These "" confused ÏÏ bins comprise 3.5% of the sky at o b o [ 5¡ and are primarily located in star-forming regions such as Orion-Taurus, Ophiuchus, and the Magellanic Clouds. The PSCZ survey masks a larger fraction of the sky at high latitudes (8.5%), but includes 1.2% of the sky at low latitudes that is not confused.
We identify and remove a small number of galaxies by hand to make our dust maps as free as possible from contamination at o b o [ 5¡. The Point Source Catalog was searched for objects brighter than 1.2 Jy that were masked from both the 1.2 Jy and PSCZ surveys. 195 satisÐed the PSCZ color criteria for galaxies. Upon visual inspection of ISSA plates and the Digitized Sky Survey, 68 were identiÐed as point sources or galaxies and removed from the dust maps.
The many faint galaxies that remain in our maps are not signiÐcant. As the median galaxy color is our f 100 /f 60 \ 2.0, adopted Ñux cut at 60 km roughly corresponds to f 100 [ 1.2 Jy. Galaxies just below this limit will have a peak Ñux density of D0.12 MJy sr~1 at 100 km, corresponding to A(B) B 0.01 mag. Thus, the Ðnal extinction maps will have some small-scale contamination from extragalactic objects at that level. This contamination is very nearly uniformly distributed. In addition, very faint galaxies must contribute at some level to a uniform extragalactic Ñux. This contribution cannot be easily extrapolated from known galaxies, since the measured number counts must Ñatten N P f 60 1.4 at low Ñux levels to prevent the total number and total Ñux from diverging. However, any uniform contamination is irrelevant, as it is degenerate with the constant term in the zodiacal contamination that has already been removed (°2.2.2). The BH maps (1978,1982). derive the column density of dust from H I 21 cm Ñux. At northern declinations (d[[23¡) the BH maps are modulated by local H IÈtoÈdust ratios on scales of 13 deg2, derived from the Shane-Wirtanen galaxy counts. The DIRBE/IRAS dust maps, on the other hand, directly measure the dust, and no corrections need be made for optically thick H I emission, for ionization, or for the formation of molecular hydrogen. Furthermore, the DIRBE/ IRAS maps are of uniform quality over the full sky, whereas the BH maps exclude all regions at o b o \ 10¡ (17.4% of the sky), as well as a region of 1080 deg2 (2.6% of the sky) lacking 21 cm data. This latter region is near the south Galactic pole, approximately bounded by [135¡ \ l \ 21¡ and b \ [62¡.

Stars
We limit the discussion of tests of our reddening map to two extragalactic reddening measurements. These tests provide conÐrmation that our dust maps are indeed suitable estimates of reddening, as good as or better than those of BH. These tests also normalize the amplitude of reddening (in magnitudes) per unit of 100 km Ñux (in MJy sr~1). Our reddening estimates can be written simply as where we seek the calibration coefficient p. DT represents the point sourceÈsubtracted IRAS-resolution 100 km map, corrected to a reference temperature of 18.2 K using the DIRBE temperature map (eq. [20]). The reddening of external galaxies allows a straightforward calibration of our maps. For example, consider the sample of brightest cluster ellipticals of & Lauer Postman for which B[R colors are provided for 106 gal-(1995), axies. These objects are at modest redshift (z \ 0.05) and are well distributed over the sky. We use the k-corrected colors provided and apply our dust maps to estimate the reddening toward each galaxy, assuming Appendix For a good normalization of the dust map, one expects B). no correlation between intrinsic B[R for the galaxies and E(B[V ). Because the distribution of the intrinsic B[R of the BCG galaxies is not Gaussian, it is best to use a nonparametric statistical procedure to test the correlation. We measured the Spearman rank correlation coefficient between extinction-corrected B [R and E(B[V ). For calibration constants p \ 0.0118 or p [ 0.0196, the chances that a random distribution would have a correlation as large as observed is less than 5%. Thus, our 95% conÐdence limits for the normalization of the maps is p \ 0.016^0.004. As conÐrmation of the procedure, we applied the same statistic to these galaxies using the BH reddening estimates : EBH represents the BH reddening in B[V , limiting the analysis to the 99 galaxies with BH estimates. For a scaling factor in the range 0.66 \ q \ 1.05, the BH reddening applied to the BCG galaxies leaves no signiÐcant correlation between intrinsic B[R and foreground reddening, indicating that the BH normalization (q \ 1) is acceptable. A sharper test of the reddening calibration is made possible by using an auxiliary correlate to reduce the variance of the intrinsic color distribution. For example, the Mg 2 index described by et al. is expected to correl-Faber (1989) ate closely with the intrinsic B[V color for elliptical galaxies. For both increased metallicity and increased stellar population age, the line strength increases, and the colors become redder. Use of the index is ideal, since it is Mg 2 available for large samples and is not a †ected by reddening. We use the sample of 472 elliptical galaxies presented by Faber et al., which has very broad sky coverage, including many galaxies in rather dusty directions. Of this sample, 389 galaxies have photoelectric colors in ¹30A apertures, measurements, and estimates of BH reddening. Mg 2 -index Five of these galaxies are reported by D. Burstein (1997, private communication) to have suspect photometry and have been removed from our list (NGC 83, NGC 545, NGC 708, NGC 1603, and NGC 7617). The B[V colors are listed in this catalog as being k-corrected and reddeningcorrected in a manner that deviates slightly from BH reddenings given in the published full-sky BH map. In one region of the sky (230¡ \ l \ 310¡, [20¡ \ b \ ]15¡) it was determined that the BH map is not reliable, and reddenings were assigned based upon deviations in the Mgcolor relation et al.
We have added the (Burstein 1987). tabulated reddening corrections back to the colors to obtain the raw (only k-corrected) colors before proceeding with our analyses. We should note that the unreliable region lacks a dust-to-gas ratio estimate in the BH map construction.
Dave Burstein kindly sent us an updated version of the Faber et al. regression, plotted against Mg 2 foreground reddening, for (a) the BH maps and (b) the DIRBE/IRAS maps. Pluses represent galaxies at southern declinations where the BH maps lack dust-to-gas ratio information, and asterisks are those lacking any BH values. residuals is evident for both BH and DIRBE/IRAS corrections, in the sense that the highest reddening values appear to be overestimated. However, this trend is not statistically signiÐcant for the DIRBE/IRAS corrections. Note that the galaxies lacking BH estimates (asterisks) were not used in the Ðts, but still have reasonable color residuals using DIRBE/IRAS reddening corrections.
The larger size and reduced intrinsic color scatter of the et sample leads to a much tighter constraint on p, Faber al. one that is consistent with the value derived from the BCG sample. The 10% precision of the calibration is remarkably tight and is consistent with early estimates of IRAS reddening calibrations from et al. Rowan-Robinson (1991). (However, we note that IRAS-only estimates of the reddening severely su †er from the poorly constrained zero point, zodiacal contamination, and unaccounted variations in the dust temperature.) The accuracy of the reddening maps can be estimated from the residuals in the Mg-color relation. We assume that the errors in the reddening estimates are fractional, p D~I increasing with increasing reddenings such that We further assume that the intrinsic dispersion in the Mgcolor relation is mag (including measurement p BV \ 0.0257 errors), which is the measured dispersion for a subsample of galaxies in clean regions of the sky. A total error for the color of each galaxy is the quadrature sum of and p D~I p BV . Using this model for the errors, one can determine the accuracy of the reddenings by increasing f until the s2 of the Ðt equals the number of degrees of freedom. For the BH reddening maps we Ðnd f \ 0.29, and for the DIRBE/IRAS reddening maps we Ðnd f \ 0.16. This demonstrates that the DIRBE/IRAS reddening estimates have an accuracy of 16%, which is nearly a factor of 2 improvement over the BH estimates. In regions of low reddening, mag, E(B[V ) [ 0.1 this data set indicates that DIRBE/IRAS and BH reddening estimates may be equally good.
We had originally attempted to use counts in cells of the APM galaxy survey (Maddox et al. as a nor-1990a(Maddox et al. as a nor- , 1990b malization of the dust map. Extinction as measured by the DIRBE/IRAS column-density maps can be calibrated by studying the statistical covariance between the APM and DIRBE/IRAS dust maps : dusty regions have increased dust emission and diminished galaxy counts. However, we encountered a problem similar to that seen by Heiles (1976) when he attempted to compare reddening measures to the counts of Shane-Wirtanen galaxies. Using the APM counts we Ðnd a normalization (p) that is approximately twice that described above. The extra sensitivity of the galaxy counts to dust is almost certainly a result of the catalog of galaxies being surface brightness limited as well as magnitude limited ; increased foreground dust not only diminishes the total Ñux of the galaxy but diminishes the size of the isophotal aperture and increases the likelihood that the galaxy will be either classiÐed as a star or not be counted at all. We shall provide details of this analysis in a separate paper et al. (Finkbeiner 1997).

PROCESSING SUMMARY
We have extensively processed the available far-IR data sets to generate a uniform-quality column density map of the dust that radiates from 100 to 240 km. The major steps in the processing are summarized below : 1. The annual average DIRBE maps have been regenerated such that the same zodiacal dust column is observed at all wavelengths. Data with solar elongation e \ 80¡ have been deleted.
2. A zodiacal light model has been constructed from the DIRBE 25 km map. The model parameters were constrained by forcing the dust to trace the gas (as measured by the Leiden-Dwingeloo H I survey) at high Galactic latitudes.
3. Striping artifacts in the IRAS/ISSA 100 km maps were removed using a Fourier-space Ðltering algorithm.
4. Asteroids and non-Gaussian noise were removed from the IRAS/ISSA maps using a deglitching algorithm that compared multiple scans.
5. The IRAS and DIRBE 100 km maps were combined, preserving the DIRBE zero point and calibration.
6. Stars and galaxies were removed to a limiting Ñux of Jy over most of the sky at o b o [ 5¡ and over parts f 60 \ 0.6 of the sky at lower latitudes.
7. The color temperature of the dust was derived from the DIRBE 100 and 240 km maps. A temperature correction map is used to convert the 100 km cirrus map to a map proportional to dust column.
8. The temperature-corrected dust map is calibrated to the reddening of external elliptical galaxies. The use of the B[V versus correlation allows us to set the cali-Mg 2 bration constant to 10% precision.
We have conÐrmation of the superiority of the new procedure by a test of the scatter of residual B[V colors after regression against line strength. The scatter of the Mg 2 regression is dominated by intrinsic scatter among the galaxies. We measure a reduction in the variance that is consistent with a fractional error in the reddening estimates of 16% for the new maps versus 29% for the BH estimates. Blitz, & Mundy to (Magnani, 1985), the right of center. The small-scale structure in the plot is at the resolution limits of the IRAS data. Higher resolution data would perhaps show even Ðner scale structure.
The full-sky dust map is shown as with a tunately, no high-quality H I data exists to verify that the lowest density holes are also minima in H I emission. A tabulation of dust properties for our low-density holes and the Galactic poles appears in A large region of the Table 5. southern Galactic sky (b \ [40¡, 160¡ \ l \ 320¡) appears to have very low dust emission and should therefore have the least foreground contamination for CMBR studies or large-scale structure analyses of redshift surveys. Projects in the northern hemisphere, such as the Sloan Digital Sky Survey and the north Galactic strip of the Two Degree Field (2dF) survey, may be more compromised by Galactic extinction. In particular, the north Galactic strip of the 2dF  is signiÐcantly more dusty than their southern region. However, the corrections derived from the new dust maps should prove adequate for most analyses.
In we show the power spectra of intensity Ñuc- Figure 9 tuations of the dust map calculated separately in eight independent patches of sky at high Galactic latitude. For each hemisphere we have extracted four quadrants in Galactic latitude bounded by o b o [ 45¡. The power spectra of these high-latitude patches have no preferred scale, but are reasonably well described as power laws with P(k) P k~2.5. The amplitude of the power is di †erent in each patch, which is indicative of extreme phase coherence in the cirrus structure of the Galaxy. To assume that the Ñuctuations can be approximated as random phase power is a completely inadequate description of this infrared cirrus and could lead to misleading estimates of Galactic foregrounds in CMBR experiments.
FIG. 9.ÈPower spectrum of dust. The four solid curves represent four quadrants in the north Galactic sky at b [ 45¡, while the four dashed curves represent four quadrants of the south Galactic sky.

Di †erences between the Dust and H I Maps
As seen in the correlation between the dust Figure 10, and the H I emission is remarkably tight for low Ñux levels, but has substantial scatter at higher Ñux levels. Because the 240 km map is on the Rayleigh-Jeans portion of the dust FIG. 10.ÈH I correlation with (a) DIRBE 240 km (corrected for zodiacal contamination), (b) DIRBE 100 km (also corrected), and (c) our derived dust column density. This plot demonstrates that the gas to dust relationship deteriorates at high Ñux levels. emission curve, one might expect it to correlate better with the H I than would the 100 km map. We see the opposite e †ect : there is more scatter in the 240 kmÈH I correlation. We attribute this partly to the considerably larger noise of the 240 km map. But some of this increased scatter is undoubtedly generated by cool molecular cloud regions, in which much of the hydrogen is in molecular form, and the 100 km is exponentially reduced because of the lower temperature, while the 240 km emission is suppressed only linearly with temperature. Furthermore, the dust-H I scatter is nonnegligible even in the regions of low Ñux, where the gas is expected to be predominantly neutral. This scatter may indicate either Ñuctuations in the ionization fraction of the gas, extensive regions of H`, or variations in the dust-to-gas ratio within neutral zones.
A ratio map of temperature-corrected dust to the Leiden-Dwingeloo H I map is shown as The Leiden- Figure 11. Dwingeloo map includes H I gas from the full velocity range of the survey km s~1) converted to ([450 ¹ v LSR ¹ ]400 dust column via the ratio 0.0122 (MJy sr~1)/(K km s~1) found in Note that there are Ñuctuations in the Table 2. ratio that are spatially coherent, indicating they are not instrumental noise but rather real features on the sky. Most of the high-latitude sky is a neutral grey, while regions of high dust-to-gas ratio appear as white. The most prominent of these features at high latitudes must be di †use interstellar molecular clouds as well as regions of saturated H I emission, such as Orion (l B 180¡, b B [20¡) and Ophiuchus (l B 0¡, b B 15¡). The Polaris Flare (l B 150¡, b B 40¡) is very conspicuous. Comparison of with Figure 11 Figure 3 shows that these regions of high dust-to-gas ratio often coincide with the regions of lower temperature, which is as expected, since the regions are optically thick in the UV and therefore shielded from the full ionizing Ñux of the Galaxy. Although searches for CO have failed to detect an abundance of molecular gas above the Galactic disk (e.g., Magnani, & Thaddeus it is very likely Hartmann, 1997), that molecular hydrogen is very abundant in these cooler regions & Neufeld (Spaans 1997). Regions of low dust-to-gas ratio appear as black. These regions are primarily distant high-velocity H I clouds in which the dust, if present, is not as well illuminated by a UV radiation Ðeld, and is therefore too cold to emit at 100 km. The Magellanic Stream is partially present in the south, while numerous high-velocity clouds are very prominent at high latitude in the north.
A general gradient of dust-to-gas ratio from Galactic center to anticenter is perhaps present in the northern sky, but if present in the southern sky, the gradient is obscured by the Orion Nebula. Even in the gray high-latitude regions, the dust-to-gas ratio map is not uniform and exhibits Ñuctuations of^15% amplitude that are coherent over scales of 10¡. These modulations might indicate real gas to dust variation, or they may hint at some unresolved instrumental problems. For example, three parallel bands in the north are residuals of the imperfect zodiacal light removal, but their total modulation is only 15% peak to trough.
is a full-sky map of the DIRBE/IRAS Figure 12 reddening estimate minus the BH estimate for the region o b o [ 10¡. Recall that the BH maps are largely H I maps, with the zero point adjusted and with smooth variations in dust-to-gas ratio computed on the basis of galaxy counts. As expected, apart from an o †set, the BH reddening maps are very close to the new reddening estimates over most high-latitude regions of the sky, rarely di †ering by more than 0.02 mag (aside from a global zero-point di †erence). But there are large systematic di †erences at low latitude and toward molecular clouds such as Orion and Ophiuchus. These modulations largely reÑect the direct di †erences in the dust-to-gas ratio maps These coherent modiÐ- (Fig. 11). cations of the reddening estimates are important for full-sky analyses of the galaxy distribution, particularly for studies of large-scale Ñow Ðelds.

T he Reddening at the Galactic Poles
The absolute zero point of dust column density in the Galaxy is both difficult to model from 21 cm or infrared emission and difficult to measure directly from stellar colors. The zero point of 21 cm maps are uncertain because of sidelobe contamination of the radio telescopes. Furthermore, the 21 cm maps completely miss any dust associated with ionized hydrogen. Although the DIRBE instrument chopped against a cold load to insure a stable zero point, the zero point of the combined DIRBE/IRAS dust maps is uncertain because of uncertain zodiacal contamination and a possible isotropic extragalactic background. Direct measure of the reddening is an equally difficult task. Extragalactic sources can not be used, because their intrinsic colors are not known. Studies of Galactic sources are limited to calibration of the dust column density between local and halo stars. Such comparisons may su †er from selection e †ects and do not sample Galactic dust beyond a few hundred parsecs.
The debate as to the reddening zero point has traditionally focused on the values at the Galactic poles. Although we Ðnd that the absolute minima in the dust column are not at the poles (see most studies Table 5), are near these two lines of sight. Some authors have claimed that the Galactic poles are essentially free of reddening, whereas others claim a reddening of order E(B[V ) B 0.02È0.05 mag, corresponding to an extinction A(B) B 0.1È0.2 mag. This debate was reviewed by Burstein & in 1982 and has yet to be resolved. Heiles Several large studies have been conducted to measure the colors of A and F stars near the Galactic poles. Color excesses from ubvby photometry for these stars with wellknown intrinsic colors are used to measure E(b A reanalysis of some of these data by Teerikorpi (1990) argues that these results are subject to statistical biases, suggesting that the average reddening reaches E(B[V ) \ 0.04 mag at 400 pc above the Galactic plane.
Our estimated reddening, averaged over 10¡ in diameter,  (1978,1982) 3.1 extinction law. We believe that a nonzero reddening at the poles is compelling for several reasons. First of all, we observe both 100 km and H I emission at the poles, not all of which could be scattered light or zodiacal emission, since it is structured and not entirely smooth at either pole. The fact that a linear dustÈtoÈH I regression Ðts so well through all the low-Ñux regions suggests that this is truly material in the ISM. The gas has low velocity, so it is certainly part of the Milky Way, and so, therefore, must be the dust. If the A and F star reddenings are not biased by selection e †ects, then this gas and dust must be more than several hundred parsecs above the Galactic plane. It is theoretically unreasonable for the dust at high Galactic latitude to have fully evaporated. Although the dust at the poles might have been preferentially ablated by the shocks that formed the Local Bubble, such ablation would generally be expected to act oppositely to the coagulation that occurs in dense molecular regions. The grain size distribution should tip further toward smaller grains, thereby lowering the value of but not eliminating the R V small grains that dominate UV and optical extinction.
For many analyses a change in the zero point of dust reddening is irrelevant. But for studies such as photometry of distant supernovae, such an e †ect is quite important, and its neglect will lead to systematic errors in the inferred q 0 . This arises because the extinction is not gray and more strongly a †ects the B-o rV-band spectra of nearby SNe than the R-orI-band spectra of distant SNe.
An o †set is readily apparent between the DIRBE/IRAS reddening maps and those of Burstein & Heiles. We measure a median di †erence of 0.020 mag in E(B[V ) between the maps at high Galactic latitudes (o b o [ 45¡), in the sense that the BH maps are systematically lower. For the reasons stated above, we believe the DIRBE/IRAS zero point to be more reliable. If one were to shift the DIRBE/ IRAS maps down by 0.020 mag, the lowest column density holes would have unphysical negative reddenings of E(B[V ) \[0.015 mag.

Measurement of the CIB
The search for the signature of an extragalactic background light (EBL) has a long history (Peebles 1971 ;Bond, Carr, & Hogan Carr, & Hogan & 1986 ;Bond, 1991 ;Zepf Silk et al. et al. 1996 ;Guiderdoni 1997 ;Franceschini 1997). Such a background is inevitable and can be estimated based upon a star formation history and the corresponding injection of energy and metals into the interstellar medium at various epochs. The predicted EBL can be expressed as where is the present baryon density and *Z is the overall o B metallicity produced at redshift et al. For z f (Bond 1986). and *Z \ 0.02, this yields ) B h2\0.025 lI l \ 37/(1 ] z f ) nW m~2 sr~1. A high background Ñux thus favors substantial metal production at low redshift. The emergent Ñux will appear either in the UV/visible/near-IR window if the star formation is not very dusty, or it will appear in the 100È300 km window if the majority of the light in star forming regions is reprocessed by dust.
In the discussion of above we described our regres-°2.3 sion of the dust maps on the H I data while removing the zodiacal foreground and a constant background term.
lists the derived background coefficients in MJy Table 3 B b sr~1 as a function of the limiting ecliptic latitude cut o b o.  (1996). only signiÐcant di †erence between our procedure and that of Boulanger et al. is that we Ðtted all terms of the linear regression simultaneously. In perhaps more convenient units, a surface brightness of 1.0 MJy sr~1 translates to 20, and 12 nW m~2 sr~1 at 100, 140, and 240 km, lI l \ 30, respectively. The derived Ñuxes from thus translate Table 3 to an upper limit of 15 nW m~2 sr~1 at 100 km, and detected Ñux of 32^13 nW m~2 sr~1 at 140 km, and 17^4nWm~2 sr~1 at 240 km, where we have included the estimated systematic error.
We suspect that this isotropic Ñux is either extragalactic or is some sort of foreground directly related to the DIRBE telescope. No Galactic source is likely to be isotropic. We detect no hint of a dust layer distributed like csc o b o that is uncorrelated with the H I distribution, as might be expected from a uniform disk of ionized gas. Since these results derive from comparison to the H I distribution, they are sensitive to unknown o †sets in the H IÈdust relation. We have assumed that the Galactic H IÈtoÈdust relationship has no zero o †set, which leads us to more reddening than predicted by the BH maps. If we force our mean high-latitude reddening to match the BH maps, then the inferred CIB would increase.
A possible instrumental explanation for the strong signature at 140 and 240 km is given by Although Hauser (1996). the entire DIRBE telescope was operated at 2 K, which should minimize radiation by the telescope mimicking an isotropic background, in these two channels there is a measurable radiative o †set induced by junction Ðeld e †ect transistors (JFETs), operating at 70 K, used to amplify the detector signals. Uncertainty in the correction for this e †ect is estimated to be 5 (2) nW m~2 sr~1 at 140 (240) km et al. which is considerably less than the (Fixsen 1997), signal that we measure.
Presuming that this Ñux is indeed a cosmic infrared background, the inferred value is somewhat above that expected from the integrated star formation and dust reprocessing history of high-redshift galaxies, even if they are shrouded in dust et al.  1996). Puget (1996) background detection from analysis of COBE/FIRAS data in the 400È1000 km region, and our background measurements appear to be on the high side of the extrapolation of their measurements.
The EBL derived from the optical Ñuxes in the Hubble Deep Field (HDF), with modest extrapolation of the galaxy counts, is approximately 7.5 nW m~2 sr~1 a (Madau 1997), factor of D2 smaller than that reported here. After masking detected sources in the HDF, has shown Vogeley (1997) that the remaining EBL, unless it is truly uniform, has a surface brightness that is at most a small fraction of the integrated light of the discrete sources. These results, together with recent observations of the UV continuum slope at high z et al. suggest that much early (Pettini 1997), star formation was very dust enshrouded, reprocessing most of the UV, optical, and near-IR photons to the far-IR. An integrated far-IR background Ñux of such a magnitude, if correct, is a very promising sign for high angular resolution studies of the far-infrared. Space missions such as FIRST and the Infrared Space Observatory, and groundbased submillimeter observations with SCUBA on the JCMT and with the MMA will have a tremendous opportunity to resolve this background or to demonstrate that it does not actually exist.
A more detailed analysis of the CIB by the DIRBE team is currently underway et al. et al. (Hauser 1998 ;Kelsall et al. et al. Instead of our 1998 ;Arendt 1998 ;Fixsen 1998). simple empirical model of the zodiacal foreground, they have solved for a detailed three-dimensional model of the interplanetary dust. They provide a much more complete analysis of the separation of Galactic emission from the infrared background, and their results will undoubtedly supersede the preliminary measurements reported here.

SUMMARY
We have constructed a full-sky map of the Galactic dust based upon its far-infrared emission. The IRAS experiment provides high angular resolution at 100 km, whereas the DIRBE experiment provides the absolute calibration necessary across several passbands to map the dust color temperature and to convert 100 km emission to dust column density. Point sources and extragalactic sources have been removed, leaving a map of the infrared cirrus. This dust map is normalized to E(B[V ) reddening using the colors of background galaxies. The Ðnal maps have a resolution of and are shown to predict reddening with an accuracy of 6@ .1 16%.
The new dust map leads to reddening estimates quite consistent with the Burstein-Heiles maps in most regions of the sky, with the new maps proving to be twice as reliable. These new maps are certainly to be preferred in regions of high extinction, where H IÈbased maps su †er from saturation of the 21 cm line or insensitivity to molecular hydrogen. Further tests are encouraged to determine the accuracy of our predicted reddenings and extinction. In particular, the accuracy of the maps at o b o \ 10¡ has yet to be established.
The maps will undoubtedly prove useful for analyses of current and future CMBR experiments, as well as a host of Galactic structure studies. For example, it will be of interest to determine whether the temperature variations observed at high latitude are consistent with molecular line observations, and whether better constraints on the distance to the high-latitude molecular clouds can be obtained. Molecular line observations of several of the regions of low dustÈtoÈ H I ratio (i.e., the high-velocity clouds) will give information on the metal abundance in these regions, which might lead to better constraints on their distance from the plane of the Milky Way.
We show that the dust correlates very well with the available H I maps over most of the sky. The ratio of dust to gas can be used to Ñag molecular clouds on one extreme and high-velocity H I clouds on the other. We note that the lowest regions of dust emission occur in the southern Galactic sky, in a region where there exist no high-quality H I maps. The south Galactic sky has less power in the dust and should be preferred for CMBR experiments and largescale Galaxy surveys.
We argue that & Heiles underestimate Burstein (1982) reddening by 0.020 mag in E(B[V ). Our predicted extinction at the Galactic poles is A(B) \ 0.065 mag (north) and A(B) \ 0.080 mag (south) on 10¡ scales. There do exist holes that have extinction as low as A(B) \ 0.02 mag ; the most prominent of these are listed in We Ðnd that the Table 5. lowest column density dust holes are in the southern Galactic sky, but it is not known if these are also the regions of lowest H I. If these selected regions are indeed low in H I, then they should become preferred directions for deep imaging and spectroscopy of extragalactic sources in the soft X-ray bands, where Galactic extinction is a severe problem.
In the process of generating these maps and removing the zodiacal foreground, we have detected what appears to be a signiÐcant far-IR background Ñux, approximately 32 and 17 nW m~2 sr~1 in the 140 and 240 km bands, respectively. This Ñux is surprisingly high, higher than the integrated light seen in the Hubble Deep Field. If our measurement is a detection of the CIB and not some artifact associated with the DIRBE instrument, this suggests that early star formation was heavily dust enshrouded in most of the universe.
The dust maps are publicly available, as described in Appendix C.
The COBE data sets were developed by the NASA Goddard Space Flight Center under the guidance of the COBE Science Working Group and were provided by the NSSDC. We thank Sherry Wheelock of IPAC for providing us with ISSA coverage maps. We thank Steve Maddox for supplying the APM maps, Will Saunders and the PSCZ team for providing positions and identiÐcations from the PSCZ survey, and Tod Lauer for providing the BCG data. We are especially indebted to Carl Heiles for stimulating discussions and invaluable guidance during the course of this project. Further discussions with Dave Burstein and Chris McKee were extremely useful. All the analysis described in this paper was performed in IDL, which increased our efficiency by an enormous factor. DPF was partially supported by an NSF Graduate Fellowship. This work was supported in part by NASA grant NAG 5-1360. The concern about remaining contamination from stars is that, although small on average, it may increase dramatically at low Galactic latitudes, where the star density is very high. We show that this is not a problem by extrapolating to the contribution from stars below our Ñux limit. In order to trace the star counts to low Ñux and low Galactic latitude, we restrict this analysis to a strict color cut that limits cirrus confusion : 0.1 \ f 60 /f 25 \ 0.3 ; 0.03 \ f 60 /f 12 \ 0.15 ; f 100 /f 60 \ 2. This strict color cut retains 70% of the stars, while efficiently excluding the cirrus. The color-color plane in Figure 13 compares it to the looser cut, The distribution of stars (satisfying this strict color cut) as a function of Ñux is With a slope in the range [1 \ m \ 0, the number of stars at faint Ñux levels diverges, but the Ñux from those sources converges. This allows us to normalize the Ñux from stars fainter than a given Ñux cut relative to the number of stars f cut , brighter than that cut : For a Ñux cut Jy, this ratio is 3.2 Jy~1. As the number of stars must converge, this ratio represents an upper limit to f cut \ 0.6 the Ñux from faint stars. Next, we trace the distribution of stars as a function of Galactic latitude. Again to minimize cirrus confusion, we limit ourselves to the strict color cut and brighter stars Jy). The distribution of stars roughly At lower latitudes, the PSC is losing stars primarily because of confusion with cirrus. At o b o \ 10¡ the cirrus has a surface density that is typically 17 times brighter than a 1.2 Jy point source in the ISSA 60 km maps.
We combine the Ñux and latitude distribution of stars to estimate the contamination from faint stars. Using a mean color for stars of we plot the estimated contribution from stars fainter than Jy in the 100 km maps f 100 /f 60 \ 0.54, f 60 \ 0.6 (Fig. 14). The contamination is approximated by the functional form For comparison, we overplot the total 100 km Ñux from the PSC stars explicitly removed from the maps. This demonstrates that the faint stars are expected to contaminate the 100 km maps at a level less than 0.01 MJy sr~1 for o b o [ 5¡. This is not signiÐcant, as it represents contamination in derived A(B) values of less than 10~3 mag.

APPENDIX B EXTINCTION IN DIFFERENT BANDPASSES
We have shown that the DIRBE/IRAS dust maps faithfully trace the dust responsible for reddening of blue light We (°5). need some method of extrapolating the reddening in B[V to reddening and extinction in other passbands. Measuring such relative extinctions is a well-developed industry, albeit one still fraught with some controversy. In particular, the composition of interstellar dust is not well known, and nor is its variation within the Galaxy.
FIG. 13.ÈColor-color diagram for PSC sources. The diagonal line efficiently discriminates between galaxies (above the line) and stars (below the line). The square box is a strict color cut that retains 70% of stars. For clarity, only 1/10 of the stars are plotted.
FIG. 14.ÈContamination at 100 km from faint stars. The solid histogram represents the derived contamination at 100 km from stars with Ñuxes below our Ñux cut. The dotted histogram shows the Ñux from stars explicitly removed from the maps.
The selective extinction is variable across optical passbands. This extinction curve is usually parameterized in terms of the V -band extinction A(V ) and a measure of the relative extinction between B and V band, .  (Draine 1985 ;Guhathakurta 1989 ;Kim 1995). source of concern, as the extinction curves vary signiÐcantly in blue passbands for di †erent values of These variations can R V . be well represented by the same two-parameter function in any region of the Galaxy. We use the functional form of OÏDonnell in the optical and that of Clayton, & Mathis in the ultraviolet and infrared. The extinction curve is (1994) Cardelli, D . (B2) The system response, in terms of quantum efficiency, is represented by W (j), and S(j) is the photon luminosity of the source. As we have normalized the dust maps to the reddening of elliptical galaxies, we use an elliptical galaxy for the source. We average the normal elliptical galaxy SEDs from extrapolating the source as S(j) P j outside his range of Kennicutt (1992), spectral coverage. The above expression has been evaluated for a variety of passbands in the limit of low extinction (*m V ] 0), then rescaled to A(V ) \ 1.
The system response is the convolution of the atmosphere, telescope optics, Ðlter, and detector responses. Where possible, we have chosen the system responses that correspond to the commonly used standard stars for each Ðlter. A , R S has been widely used at Lick Observatory and is convolved with the Orbit 2K Q.E. and the atmosphere at Lick. The infrared JHKL@ bands are represented by the IRCAM3 Ðlter set at UKIRT, convolved with the atmosphere at Mauna Kea. The u@g@r@i@z@ Ðlter set for the Sloan Sky Survey represent the total system response at Apache Point as published by et al. Fukugita The total system response for some of the broadband Ðlters for WFPC2 were taken from the Space Telescope Science (1996). Institute web page. The photographic responses for the Second Generation Digitized Sky Survey are from Djorgovski, Weir, & Fayyad convolved with the atmosphere at Palomar. (1995), The passband used for the APM maps has an unusual deÐnition and is treated di †erently. The APM maps were b J calibrated with CCD photometry using the following deÐnition et al.
Combining equations and one can show the following to be rigorously true : (B2) (B3), The results for all passbands appear in The e †ective wavelength represents that wavelength on the extinction Table 6. j eff curve with the same extinction as the full passband. The Ðnal column normalizes the extinction to photoelectric measurements of E(B[V ). Assuming an dust model, the dust maps should be multiplied by the value in this column to R V \ 3.1 determine the extinction in a given passband. Extinctions in narrow passbands can be determined by evaluating the Cardelli et or extinction law at the corresponding wavelength. al. OÏDonnell APPENDIX C

DATA PRESENTATION
The 100 km map and the dust map are available electronically. Both maps consist of the destriped, deglitched ISSA data recalibrated to DIRBE and point-source subtracted. The 100 km map maps the emission in units of MJy sr~1. The dust map applies our temperature correction to convert 100 km emission to column density of dust, calibrated to E(B[V ) reddening in magnitudes. Reddenings and extinctions in other passbands are computed by multiplication with the numbers in Table 6. All where n \]1 for the NGP, and n \[1 for the SGP. Pixel numbers are zero indexed, with the center of the lower left pixel having position (x, y) \ (0, 0). These Lambert projections are minimally distorted at high Galactic latitudes, with the distortion approaching 40% at b \ 0¡. The pixel size of well samples the FWHM of (2@ .372)2 6@ .1. The caveats to using these maps to measure reddening or extinction can be summarized as follows : 1. Every e †ort has been made to remove both extragalactic sources and unresolved (Galactic) sources from the dust maps at o b o [ 5¡ and unconfused regions at lower latitudes. Some sources will remain, owing to confusion or unusual FIR colors. No sources fainter than 0.6 Jy at 60 km are removed, although these are not signiÐcant contaminants.
2. The IRAS satellite did not scan a strip amounting to 3% of the sky. In addition, we remove a circle of radius 2¡ centered at (l, that is contaminated by Saturn. These regions are replaced with DIRBE data and have no point b) \ (326¡ .28, ]51¡ .66) sources removed.  1989 OÏDonnell 1994.