In zonation, the basis vectors (A in Eq. 1) are piecewise constant functions. For the pilot point method the ith basis vector is CM(x - xi) where xi is the location of the ith pilot point and CM(x - xi) is the covariance of model parameters at x with model parameters at xi. Other types of basis vectors A that have been used for history matching include spline function, wavelet function, leading eigenvectors of the covariance matrix of model variables CM and discrete cosine transform. For the pilot point method, the values of the rock property fields at the pilot points are the parameters to be determined through history matching. Values of the properties at other locations are found by kriging interpolation. One key issue for implementation of the pilot point method is the determination of the proper number of pilot points and the optimal locations. In the first application to groundwater hydrology, de Marsily et al. [41] located the pilot points moreor- less uniformly but also attempted to follow zones of large transmissivity contrasts. LaVenue and Pickens [93] located pilot points in regions of highest sensitivity. One or two pilot points were added at each iteration, the transmissivity was adjusted, and then if the residual was still large, more pilot points were added. Transmissivity at the pilot points were constrained to lie within prescribed bounds. Wen et al. [177] located the pilot points at randomly selected locations in such a way that there was approximately one pilot point per correlation length. This was done iteratively, and the locations were changed at each iteration. Bissell et al. [21] used sensitivity information computed using the direct method to compute optimal locations for placement of pilot points. In a more recent publication, Wen et al. [178] used a genetic algorithm to search for optimal pilot point locations, as well as the associated optimal permeability perturbations at the pilot locations. Because the basis vectors are rows of the prior covariance matrix, it has been argued that regularization is not necessary with the pilot points method [184], but McLaughlin and Townley [116] point out that the lack of a regularization term could result in instability. The practical advantage of a Tykhonov-type regularization on constraining pilot point calibration was illustrated by Fienen et al. [53]. Liu and Oliver [107] showed that pilot point methods tend to result in anomalous distributions of extreme values in history matching problems, but that the problem is reduced when regularization is used. Spline bases have also been used in history matching to represent either the property field to be estimated or the change to the property field [44, 114]. Lee et al. [97] used bicubic splines to parameterize estimates of permeability and porosity distributions from well pressure data in a single-phase, two-dimensional reservoir. Lee and Seinfeld [99] extended the methodology to the problem of estimation of permeability in a twophase, two-dimensional reservoir from well pressure data. The bicubic spline representation of the change to the parameter field is less likely to result in visual artifacts that are sometimes seen in pilot point and zonation parameterizations [see 116, Fig. 3] or [see 131, Fig. 9.2]. On the other hand, the spline parameterization is flexible enough to be used for piecewise smooth representation of reservoir properties [37]. For random fields with a known covariance, a basis formed from the spectral (eigenvalue/eigenvector) decomposition of the prior model covariance can be efficient for history matching. Reynolds et al. [138] showed that use of the KarhunenLoeve reparameterization significantly decreased the computational effort required to generate realizations of the reservoir model conditioned to multiwell pressure data. Similar approaches to reparameterization have been used by Sarma et al. [147] and Zhang et al. [190]. Romary [141] discussed a method for determining the number of terms to be kept in a truncated KarhunenLoeve expansion. The discrete cosine transform has been used as a parameterization by Jafarpour and McLaughlin [86] and Jafarpour et al. [85]