Issue No.05 - September/October (1999 vol.1)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/5992.790592
When phenomena occur at the threshold of detection capabilities, event rates are low and observers usually record data as discrete counts. Relatively large fluctuations of background noise might mask a signal of interest. (Of course, one investigator's noise might be another's signal, as the discovery of cosmic black-body radiation by Arno Penzias and Robert Wilson illustrates.) In an ideal observation, background counts would be uniform and we could subtract them out, leaving just the signal. However, photon and nuclear radiation counts, for example, are subject to Poisson statistics with inevitable fluctuations in both background and signal counts. Most methods for analyzing such data lose information because background counts are subtracted before the remaining counts are analyzed. Therefore, these counts have larger relative fluctuations, and the subtraction often attributes negative numbers of counts to the signal--a physical impossibility.
William J. Thompson, "Don't Subtract the Background", Computing in Science & Engineering, vol.1, no. 5, pp. 84-88, September/October 1999, doi:10.1109/5992.790592