Avalanche photodiodes (APDs) are atlractive for use in large capacity long distance optical communication systems. They provide an advantage over p-i-n detectors because of their internal gains. Recent development of optical communication systems with transmission rates approaching seveml tens gigabits per second have stimulated research work in the area of widebandwidth APDs. Improved gain bandwidth products have been observed in APDs using either superlanice [1,2] or thin multiplication regions (typical < 500 nm) [3]. A gain-bandwidth product of -300 GHz has been achievable using an APD with a 4 0 0 nm thick multiplication layer [4]. Avalanche multiplication based on impact ionization is a noisy process intrinsically. In a conventional APD, the multiplication region is long. Camers can he ionized essentially anywhere within the avalanche region. This causes a large gain fluctuation; therefore, it becomes another source of noise in an APD in addition to the shot noise. However, in many State-Of-the-aTt devices the number of ionizing collisions per primilly carrier transit, N, is finite and perhaps even very small @=1-5). The smaller the multiplication region, the smaller the number impact ionization will happen and smaller the APD noise will be [5,6]. Indeed, Campbell et al. have observed S A M (separate absorption and multiplication) -APDs with lower multiplication noise than predicted by conventional continuum noise theory when they reduce the multiplication layer thickness h m 1500 nm to less than 500 m [7,81. One useful and important multiplication and noise local theory actually explained very well the thin APD noise behavior is described in the following. Using the method of recurrent generating functions, Van Vliet et al. @,IO1 developed a theory in which the number of ionizing collisions per carrier transit, N, could equal any number. If we express the APD noise as follows: where A is the area, 4 = eAJ.(O) is the primary injected photocurrent, F is the excess noise factor. Following Van Vliet's derivation, we have: (i:)= Z e [ e A J , ( O ) ] M : F = 2 e l , M : ( I + f ) (1) F = ( I + f ) = 1 + var(X,)/M: w h m M e = (g. -g 6 )/[(g, -1). gg ( gg / g , )' -g,(g, -(4) k = ( g r -I ) l ( g a -1 ) M, is the total gain of the N ionization stage and &, sg represent the electron and hole gain per ionization stage [9]. Fig. 1 (a), @) show the excess noise factor as a function of total gain for different N and k. We realize that the larger the N, the larger the excess noise. With a thin avalanche region, the ionij%g collisions p r carrier transit can be very small. It is not hard to understand h m the figures that the noise is smaller than what the conventional Mclntyre predicted when the N is finite and small. For the N=l case, the excess noise factor is always smaller than 2 and the highest gain can be obtained h m such a cay when the k-ratio is approaching to 1 16) ns 5hOWn in Fig. I@). Such an N=l device has been named as the " ballistic avalanche photodiode" and expected to be able to pmduce ulbd-low no...