In order to detect small-scale deformations during disease propagation while allowing large-scale deformation needed for inter-subject registration, we wish to model deformation at multiple scales and represent the deformation compactly at the relevant scales only. This paper presents the kernel bundle extension of the LDDMM framework that allowsmultiple kernels at multiple scales to be incorporated in the registration. We combine sparsity priors with the kernel bundle resulting in compact representations across scales, and we present the mathematical foundation of the framework with derivation of the KB-EPDiff evolution equations. Through examples, we illustrate the influence of the kernel scale and show that the method achieves the important property of sparsity across scales. In addition, we demonstrate on a dataset of annotated lung CT images how the kernel bundle frameworkwith a compact representation reaches the same accuracy as the standard method optimally tuned with respect to scale.