We present the results of an extensive analysis of the star formation rates determined from the NICMOS deep images of the northern Hubble Deep Field. We use SED template fitting photometric techniques to determine both the redshift and the extinction for each galaxy in our field. Measurement of the individual extinctions provides a correction for star formation hidden by dust obscuration. We determine star formation rates for each galaxy based on the 1500 angstrom UV flux and add the rates in

more »

... dshift bins of width 1.0 centered on integer redshift values. We find a rise in the star formation rate from a redshift of 1 to 2 then a falloff from a redshift of 2 to 3. However, within the formal limits of the error bars this could also be interpreted as a constant star formation rate from a redshift of 1 to 3. The star formation rate from a redshift of 3 to 5 is roughly constant followed by a possible drop in the rate at a redshift of 6. The measured star formation rate at a redshift of 6 is approximately equal to the present day star formation rate determined in other work. The high star formation rate measured at a redshift of 2 is due to the presence of two possible ULIRGs in the field. If real, this represents a much higher density of ULIRGS than measured locally. We also develop a new method to correct for faint galaxies or faint parts of galaxies missed by our sensitivity limit, based on the assumption that the star formation intensity distribution function is independent of redshift. We measure the 1.6 µm surface brightness due to discrete sources and predict the 850 µm brightness of all of our galaxies based on the determined extinction. We find that the far infrared fluxes predicted in this manner are consistent with the lack of detections of 850 µm sources in the deep NICMOS HDF, the measured 850 µm sky brightness due to discrete sources and the ratio of optical-UV sky brightness to far infrared sky brightness. From this we infer that we are observing a population of sources that contributes significantly to the total star formation rate and these sources are not overwhelmed by the contribution from sources such as the extremely super luminous galaxies represented by the SCUBA detections. We have estimated the errors in the star formation rate due to a variety of sources including photometric errors, the near-degeneracy between reddening and intrinsic spectral energy distribution as well as the effects of sampling errors and large scale structure. We have tried throughout to give as realistic and conservative an estimate of the errors in our analysis as possible. The data set for this work comes from the NICMOS and WFPC2 observations of the northern HDF. The NICMOS images used in this analysis are the IRAF-generated F110W and F160W images described in . Object Selection and Photometry Szalay, Connolly, and Szokoly 1999 describe a method for combining images from several wavebands which gives appropriate weights to the various images. We have followed the Szalay et al. 1999 procedure for selecting galaxies, a subset of which comprise the galaxies discussed in this paper. The F160W and F110W images were trimmed from their original 512 x 512 0.1 drizzled pixel sizes to 481 x 486, eliminating the outer regions of the frames where the dither pattern caused these regions to be much noisier than the inner regions. The four WFPC2 images were then transformed to the same pixel coordinate system as the NICMOS images by measuring a large number of compact objects and using the IRAF tasks GEOMAP and GEOTRAN. The transformation has an accuracy of about 0.02 , with the uncertainties probably dominated by intrinsic differences in the centroids for the different bandpasses. Next, we convolved the transformed WFPC2 images with a gaussian whose width was chosen such that the radial profile of the central star NICMOS 145.0) in the images closely matched the radial profile in the F160W image. While the agreement of these resultant radial profiles is not perfect, this procedure should be adequate for the fluxes used to carry out the template fitting described in this paper. We next laid down a grid of points where we placed bright artificial point sources and then used the program SExtractor (Bertin and Arnouts 1996) to determine the background -8and local σ at these grid points. We then fit these grids with 2D Chebychev polynomials to define the background and σ at every point in the 6 frames. 1 Each of the 6 frames is then normalized to have zero background and unit variance. This must be done locally, especially for the NICMOS frames, since the S/N varies substantially across the image and there are also small residual background variations in the reduced images described above. From these 6 frames we produced a single (weighted) χ 2 map and selected a threshold in χ 2 for identifying the pixels in the map that lie above that threshold. Since we are interested in extending our analysis to quite high redshifts, we gave the F300W and F450W images zero weight, since except for very blue low redshift objects these two bands contribute little to the final selection of objects and at high redshifts simply increase the noise. In fact, there is very little difference in the maps produced with weighting equally the F606W, F814W, F110W and F160W images and assigning zero weight to the F606W image. In the following we use the maps formed from just the F814W, F110W and F160W images. We selected a threshold of the Szalay R parameter, R = √ χ 2 = 2.3, as the significance level for "real" pixels and required 3 contiguous pixels (ie for the SExtractor parameter DETECT MINAREA) for selecting a preliminary list of objects. With these parameters, SExtractor detected 365 objects. We then measured fluxes with SExtractor in an 0.6 diameter aperture at the centroid found by SExtractor from the χ 2 image. Apertures larger than this admit too much sky 1 Because the frames were drizzled and the WFPC2 frames were further transformed and convolved, there is some short scale correlation in the pixel-to-pixel noise: nevertheless, the distribution of the pixel values is very accurately gaussian, and it is the variance of this distribution that we have measured.