Farnborough Hampshire UK = = regularisation parameter I = identity matrix = regularised weighting vector The regularised solution is then: ABSTRACT M0 u0 g (3)Expressions are presented which quantify the A major problem encountered in relation to this effects of mis-estimation of the support of the reconstruction method is the ill-conditioning of object function on the regularised minimum-norm equation (3) which is caused by the compact nature solution to the discrete bandlimited extrapolation of the operator of equation (1) or, more specifiproblem.Computational results based upon these cally, by the properties of matrix M0 [2] whose expressions are used to demonstrate the variation condition number (ratio of largest to smallest in the sensitivity of the regularised minimum-norm eigenvalues) becomes very large as the order of solution as a function of the regularisation paraincreases.meter and space-bandwidth product. Much attention has recently been given to stabilising (3) either by application of Miller-Tikhonov regularising constraints or by the use of the truncated singular value decomposition [3,4]. INTRODUCTION When the Miller-Tikhonov approach is used, equation The objective of bandlimited extrapolation is the (3) becomes: selection -according to some criterion of optimality -of a unique solution from the many (4) possible solutions to the first order Fredholm integral equation: g(t) = J f(uj)e327tdw; t e[t1,t2 ... t] (1) where g(t) is a signal sampled at t = t1,t2 tN (with a sampling interval L\T); t1c[-T,Tj whose Fourier transform is compactly supported on [-,D], where """ denotes Fourier transform. Most recent work aimed at inverting this underdetermined system of equations has sought to __ restore a unique mapping from the data space IR1 to fR°(t) eTuR0 the solution space C of smooth continuous functions I -(5) on [-,'] by minimising an energy cost function to ; HI > the constraints of (1) [i]. The effects which noise on the data vector g have Given the data vector = [g(t1), g(t2) ... g(t)] on R0 have been extensively studied [5]. Little nd defining the matrix M0 and vect.r ea H (i,j) effort has been directed, however, towards invest=s(21T%(tj_tj))/1T(ti-tj) and eT = [ii ... igating the sensitivity of the regularised solution e tN] respectively, this minimum energy solution R to mis-estimation of the support [-, Q] of