Knowledge Bank

This paper proposes a method for reducing the effects of background (100% transmittance level) variations on line intensities measured by nonlinear least-squares fitting of transmittance spectra, without modelling the background. Let f be a low-pass filter function, much wider than the lines in the spectrum. The spectrum is BT. where B is a slowly varying background and T the transmittance. The convolution $f\ast (BT)$ is then a lower-resolution version of the spectrum, with essentially the same background. The ratio $BT/(f\ast (BT))$ is a distorted version of T with the background cancelled out. If T' is a calculated transmittance, then $T\prime /(f\ast T\prime )$ can be fit to the distorted spectrum, obtaining line intensities without modelling B. This method reduces peak (observed-calculated) errors due to an unmodelled background by more than a factor of 10. Careful selection of the filter function ? can yield error spectra which have almost zero correlation with the true spectrum, and therefore cause reduced errors in the retrieved intensities. Lipkus1 used this ratio, but with f a Legendre filter, to permit comparison between a sample spectrum containing a background and a background-free standard spectrum. Statham2 and Atakan et. al.3 used a similar method based on direct high-pass filtering of the spectrum. However, their filters produce error spectra which are highly correlated with the spectrum being fit.