The theory of localizing a beam source [I‐Tai Lu, J. Acoust. Soc. Am. Suppl. 1 85, S18 (1989)] is generalized to allow an arbitrarily stratified medium. Numerical results for forward modeling and data inversion of a beam source in an arctic channel are presented. When propagating in a waveguide, an initially collimated beam undergoes diffusion after successive reflections and refractions, and is converted eventually into the oscillatory pattern of one or more guided modes. However, the up‐going and down‐going plane‐wave constituents of a given mode do not have the same excitation strength as in the case of isotropic sources. To generalize an existing modal code for a beam‐type source, one needs to decompose each eigenmode into its up‐going and down‐going constituents and then superpose these constituents with proper spectral strength and phase. For a Gaussian beam source, this procedure can be simplified via the complex source technique. By the analytic continuation of the coordinate of a point source into a complex location, one is able to derive the Green's function in a waveguide due to a beam source with arbitrary width, direction, and location of its waist. Regarding source localization, although the source plane of a beam is not localized in the real space, it is localized in the complex space. (This can be viewed via the complex source point technique.) Therefore, the existing matched‐field processes are still useful after proper modifications. The complex source coordinates whose real parts represent the location of the beam waist, and whose imaginary parts represent the width and the direction of the beam, are being sought. If the source is a collimated beam, the procedure can be further simplified into two sequential steps: (1) beam parameter inversion, and (2) beam‐waist location inversion. The first step is to determine the direction and width of the beam source by matching the envelope of the mode spectra, which is the spectra of that source in unbounded space. The second step is to determine the beam‐waist location by conventional matched‐field processing. [Work supported by Joint Services Electronics Program and by the University's Foldes Fund.]