Many processes deal with piecewise input functions, which occur naturally as a result of digital commands, user interfaces requiring a confirmation action, or discrete-time sampling. Examples include the assembly of protein polymers and hourly adjustments to the infusion rate of IV fluids during treatment of burn victims. Estimation of the input is straightforward regression when the observer has access to the timing information. More work is needed if the input can change at unknown times. Successful recovery of the change timing is largely dependent on the choice of cost function minimized during parameter estimation. Optimal estimation of a piecewise input will often proceed by minimization of a cost function which includes an estimation error term (most commonly mean square error) and the number (cardinality) of input changes (number of commands). Because the cardinality (ℓ0 norm) is not convex, the ℓ2 norm (quadratic smoothing) and ℓ1 norm (total variation minimization) are often substituted because they permit the use of convex optimization algorithms. However, these penalize the magnitude of input changes and therefore bias the piecewise estimates. Another disadvantage is that global optimization methods must be run after the end of data collection. One approach to unbiasing the piecewise parameter fits would include application of total variation minimization to recover timing, followed by piecewise parameter fitting. Another method is presented herein: a dynamic programming approach which iteratively develops populations of candidate estimates of increasing length, pruning those proven to be dominated. Because the usage of input data is entirely causal, the algorithm recovers timing and parameter values online. A functional definition of the algorithm, which is an extension of Viterbi decoding and integrates the pruning concept from branch-and-bound, is presented. Modifications are introduced to improve handling of non-uniform sampling, non-uniform confidence, and burst errors. Performance tests using synthesized data sets as well as volume data from a research system recording fluid infusions show five-fold (piecewise-constant data) and 20-fold (piecewise-linear data) reduction in error compared to total variation minimization, along with improved sparsity and reduced sensitivity to the regularization parameter. Algorithmic complexity and delay are also considered.