This thesis investigates the conjugate-gradient method and the Lanczos method for the solution of under-determined nonlinear least-squares problems regularized by a quadratic penalty term. Such problems often result from a maximum likelihood approach, and involve a set of m physical observations and n unknowns that are estimated by nonlinear regression. We suppose here that n is large compared to m. These problems are encountered for instance when three-dimensional fields are estimated from physical observations, as is the case in data assimilation in Earth system models. A widely used algorithm in this context is the Gauss-Newton (GN) method, known in the data assimilation community under the name of incremental four dimensional variational data assimilation. The GN method relies on the approximate solution of a sequence of linear least-squares problems in which the nonlinear least-squares cost function is approximated by a quadratic function in the neighbourhood of the current nonlinear iterate. However, it is well known that this simple variant of the Gauss-Newton algorithm does not ensure a monotonic decrease of the cost function and that convergence is not guaranteed. Removing this difficulty is typically achieved by using a line-search (Dennis and Schnabel, 1983) or trust-region (Conn, Gould and Toint, 2000) strategy, which ensures global convergence to first order critical points under mild assumptions. We consider the second of these approaches in this thesis. Moreover, taking into consideration the large-scale nature of the problem, we propose here to use a particular trust-region algorithm relying on the Steihaug-Toint truncated conjugate-gradient method for the approximate solution of the subproblem (Conn, Gould and Toint, 2000, pp. 133-139). Solving this subproblem in the n-dimensional space (by CG or Lanczos) is referred to as the primal approach. Alternatively, a significant reduction in the computational cost is possible by rewriting the quadratic approximation in the m-dimensional space associated with the observations. This is important for large-scale applications such as those solved daily in weather prediction systems. This approach, which performs the minimization in the m-dimensional space using CG or variants thereof, is referred to as the dual approach. The first proposed dual approach (Courtier, 1997), known as the Physical-space Statistical Analysis System (PSAS) in the data assimilation community starts by solving the corresponding dual cost function in m-dimensional space by a standard preconditioned CG (PCG), and then recovers the step in n-dimensional space through multiplication by an n by m matrix. Technically, the algorithm consists of recurrence formulas involving m-vectors instead of n-vectors. However, the use of PSAS can be unduly costly as it was noticed that the linear least-squares cost function does not monotonically decrease along the nonlinear iterations when applying standard termination. Another dual approach has been proposed by Gratton and Tshimanga (2009) and is known as the Restricted Preconditioned Conjugate Gradient (RPCG) method. It generates the same iterates in exact arithmetic as those generated by the primal approach, again using recursion formula involving m-vectors. The main interest of RPCG is that it results in significant reduction of both memory and computational costs while maintaining the desired convergence property, in contrast with the PSAS algorithm. The relation between these two dual approaches and the question of deriving efficient preconditioners (Gratton, Sartenaer and Tshimanga, 2011), essential when large-scale problems are considered, was not addressed in Gratton and Tshimanga (2009).