It has been proved that the classical TLS algorithm fails to construct a TLS solution of linear data fitting problems AX ≈ B that belong to the class F2. It will be shown how to modify this algorithm in order to reach a TLS solution. Such solution is not necessarily the minimum 2-norm or Frobenius norm one. A few ideas how to decrease its norm are briefly discussed.We are interested is solving a linear approximation problem by using the total least squares (TLS) minimization, i.e.,Any matrix X TLS satisfying (A + E)X TLS = B + G for the minimizer [G, E] is called the TLS solution. Analysis of such problems can be based on the (economic) singular value decomposition (SVD), see also [4,5]. Let z be the number of distinct singular values of [B, A]. Denote their multiplicities by m t , t = 1, . . . , z. Let σ n+1 be the kth largest singular value with the multiplicity m k = q + e so that σ n−q > σ n−q+1 = · · · = σ n+1 = · · · = σ n+e > σ n+e+1 . According to [3], consider the following notation of sub-matrices and sub-columns of V ∈ R (n+d)×(n+d) ,where V 11 ∈ R d×(n−q) , V 12 ∈ R d×(q+e) , V 13 ∈ R d×(d−e) ; V 1,t ∈ R d×mt , t = 1, . . . , z; and v 1,j ∈ R d , j = 1, . . . , n + d. Thus in particular V 12 = V 1,k = [v 1,n−q+1 , . . . , v 1,n+e ] is the sub-block corresponding to σ n+1 . Analysis in [3] divides problems (1) into several classes based on the properties of the blocks in (3). If rank([V 12 , V 13 ]) = d, then (1) belongs to the set F (corresponding to generic problems in [6]). Otherwise it belongs to the set S (nongeneric problems in [6]). The set F is futher divided into three mutually disjoint subsets, F = F 1 ∪ F 2 ∪ F 3 , where:• If rank(V 12 ) = e ∧ rank(V 13 ) = d − e, then (1) belongs to F 1 ; • if rank(V 12 ) > e ∧ rank(V 13 ) = d − e, then (1) belongs to F 2 ; • if rank(V 12 ) > e ∧ rank(V 13 ) < d − e, then (1) belongs to F 3 .The problem (1) has a TLS solution if and only if it belongs to F 1 ∪ F 2 , i.e. rank([V 12 , V 13 ]) = d ∧ rank(V 13 ) = d − e. The minimum Frobenius and 2-norm TLS solution of F 1 -problem takes the well-know closed-formwhere † denotes the Moore-Penrose pseudoinverse. However, this is not true for the F 2 -problems; see [3]. Note that for problems in F 3 and S the TLS solution does not exist.
Modification of the TLS algorithmThe problem (2) is typically solved by the classical TLS algorithm (see [6, pp. 87-88], [3, p. 767]). This algorithm seeks for the largest so that rank([V 1, , . . . , V 1,z ]) = d, and gives the output approximation in the form X OUT = −[V 2, , . . . , V 2,z ][V 1, , . . . , V 1,z ] † .