Construction of a ground-water model for a field area is not a straightforward process. Data are virtually never complete or detailed enough to allow substitution into the model equations and direct computation of the results of interest. Formal model calibration through optimization, statistical, and geostatistical methods is being applied to an increasing extent to deal with this problem and provide for quantitative evaluation and uncertainty analysis of the model. However, these approaches are hampered by two pervasive problems: 1) nonlinearity of the solution of the model equations with respect to some of the model (or hydrogeologic) input variables (termed in this report system characteristics) and 2) detailed and generally unknown spatial variability (heterogeneity) of some of the system characteristics such as log hydraulic conductivity, specific storage, recharge and discharge, and boundary conditions. A theory is developed in this report to address these problems. The theory allows construction and analysis of a ground-water model of flow (and, by extension, transport) in heterogeneous media using a small number of lumped or smoothed system characteristics (termed parameters). The theory fully addresses both nonlinearity and heterogeneity in such a way that the parameters are not assumed to be effective values. The ground-water flow system is assumed to be adequately characterized by a set of spatially and temporally distributed discrete values, ?, of the system characteristics. This set contains both small-scale variability that cannot be described in a model and large-scale variability that can. The spatial and temporal variability in ? are accounted for by imagining ? to be generated by a stochastic process wherein ? is normally distributed, although normality is not essential. Because ? has too large a dimension to be estimated using the data normally available, for modeling purposes ? is replaced by a smoothed or lumped approximation y?. (where y is a spatial and temporal interpolation matrix). Set y?. has the same form as the expected value of ?, y 'line' ? , where 'line' ? is the set of drift parameters of the stochastic process; ?. is a best-fit vector to ?. A model function f(?), such as a computed hydraulic head or flux, is assumed to accurately represent an actual field quantity, but the same function written using y?., f(y?.), contains error from lumping or smoothing of ? using y?.. Thus, the replacement of ? by y?. yields nonzero mean model errors of the form E(f(?)-f(y?.)) throughout the model and covariances between model errors at points throughout the model. These nonzero means and covariances are evaluated through third and fifth-order accuracy, respectively, using Taylor series expansions. They can have a significant effect on construction and interpretation of a model that is calibrated by estimating ?.. Vector ?.. is estimated as 'hat' ? using weighted nonlinear least squares techniques to fit a set of model functions f(y'hat' ?) to a. corresponding set of observations of f(?), Y. These observations are assumed to be corrupted by zero-mean, normally distributed observation errors, although, as for ?, normality is not essential. An analytical approximation of the nonlinear least squares solution is obtained using Taylor series expansions and perturbation techniques that assume model and observation errors to be small. This solution is used to evaluate biases and other results to second-order accuracy in the errors. The correct weight matrix to use in the analysis is shown to be the inverse of the second-moment matrix E(Y-f(y?.))(Y-f(y?.))', but the weight matrix is assumed to be arbitrary in most developments. The best diagonal approximation is the inverse of the matrix of diagonal elements of E(Y-f(y?.))(Y-f(y?.))', and a method of estimating this diagonal matrix when it is unknown is developed using a special objective function to compute 'hat' ?. When considered to be an estimate of f