Reduced-Dimensionality Inverse Scattering Using Basis Functions Andrew E. Yagle Dept. of EECS, The University of Michigan Ann Arbor, MI Presentation Overview • • • • Problem Statement Basis Function Representation Matrix Problem Formulation Reformulation as Overdetermined Multiparameter Eigenvalue Problem • Use of Left Null Matrix • 1-D Illustrative Numerical Example Inverse Problem Statement • GIVEN: 1 monopole point source antenna 1 frequency, moving platform (e.g., plane) • Unknown scatterer V(x); compact support • Unknown Green’s function G(x,y) • Response at x to source at same x: u(x) • GOAL: Reconstruct V(x) from u(x) Inverse Scattering Formulation G(x,x’) V(x’) Inverse Problem Statement u ( x) G( x, y )V ( y )G( y, x)dy Reciprocity: G(x,y)=G(y,x); x and y in 3 u ( x) G ( x, y )V ( y )dy 2 Assume: Born (single-scatter) approximation Basis Function Representation Assume: Unknown linear combinations of known basis functions, as follows: N G ( x, y ) g i i ( x, y ) 2 i 1 M V ( y) v j j ( y) j 1 uk u ( x) k ( x)dx Basis Function Representation • Should not be separable in receiver x and source y locations (precludes deconvolution) • [Don’t confuse this with separable in (x,y,z)] • Need not be orthonormal, complete, or biorthogonal to each other • Sample observations spatially: uk=u(x-xk) [special case: impulse basis functions] Basis Function Representation • Selections of all of these basis functions are problem-dependent • Multilayered media: Green’s function=sum of several terms with unknown reflections • Multipole, wavelet, Fourier representations • Need (NM) independent observations u(x) [either samples or coefficient dimensionality] Matrix Problem Formulation Method-of-Moments (MoM) linear system: Insert expansions into integral equation: N M uk g i v j Ai , j ,k i 1 j 1 Where : Ai , j ,k i ( x, y ) j ( y ) k ( x)dx dy Matrix Problem Formulation Rewrite as huge (NM)X(NM) linear system u1 A111 ANM 1 g1v1 u NM A11( NM ) ANM ( NM ) g N vM Matrix Problem Formulation In principle: Could solve this, and then: g1v1 g1vM g1 v1 vM g N v1 g N vM g N BUT: Far too large to be practical! Reformulation as Overdetermined Multiparameter Eigenvalue Problem Define: N matrices Ai, each (NM)XN, as: Ai11 Ai Ai1( NM ) AiM 1 AiM ( NM ) Reformulation as Overdetermined Multiparameter Eigenvalue Problem Rewrite: Previous (NM)X(NM) system as: u1 g1v1 A A 1 N u NM g N vM Reformulation as Overdetermined Multiparameter Eigenvalue Problem Rewrite: Multiparam eigenvalue problem: u1 g1v1 ( g A ... g A ) 1 1 N N u NM g N vM Reformulation as Overdetermined Multiparameter Eigenvalue Problem 1. Heavily overdetermined (NM>>N+M) 2. Actually (NM) simultaneous polynomial equations in (N+M) unknowns gi and vj 3. But solution not easy (see below) 4. Make use of (NM) data points as follows: Use of Left Null Matrix • Apply recent procedure for multichannel blind deconvolution (both 1-D and 2-D): • “Tall” matrices (#rows>#columns) have left nullspaces; basis can be computed • [null vectors]X[matrix of unknowns]=[0] • This becomes linear system in unknowns • Now adapt this to the present problem: Use of Left Null Matrix • • • • • • There is a “reclining” matrix [B] so that: [B][A1|A2|…|AN]=[0 0…0] Ai is (NM)XM as defined previously [A1|A2|…|AN] is thus (NM)X(N-1)M B is MX(NM) where M=NM-(N-1)M B can be PRECOMPUTED from Ai! Use of Left Null Matrix Premult: Huge linear system by known B: u1 v1 B Y g N BAN vM u NM BUT: MXM linear system, not (NM)X(NM)! Use of Left Null Matrix • Instead of the huge (NM)X(NM) linear system, have small MXM linear system! • Precompute the left null vector B from known basis-function-derived A matrix: Off-line computation; do for many bases • Solve system directly for vi coefficients: Can incorporate a priori information • Sufficient statistic: M-point Y=B[u] Use of Left Null Matrix: Stochastic Formulation • Usually have a priori pdfs for coefficients • Compute MAP (Maximum A posteriori Probability) estimator instead of the ML (Maximum Likelihood) estimator • If noise and a priori information pdfs are Gaussian, get least-squares solution • Otherwise, use iterative algorithm (EM) 1-D Illustrative Numerical Example • 1-D problem; entirely discrete space-time • u(i)=response at i to impulsive source at i • G(i,j)=response at i to impulse at j • u(i)= G(i,j)V(j)G(j,i)= G^2(i,j)V(j) • GOAL: Reconstruct V(j) from u(i) 1-D Illustrative Numerical Example • BASIS FUNCTION EXPANSIONS: • G^2(i,j)=g1/(i-j)^2+g2/(i+j)^2 [N=2] • Toeplitz-plus-Hankel structure (not exploited here, but not uncommon) • Symmetric: G(i,j)=G(j,i) (reciprocity) • V(j)=v1(j-1)+v2(j-2) [M=2] • 2-point support for scatterer 1-D Illustrative Numerical Example • • • • • BASIS FUNCTIONS: Green’s function: 1(i,,j)=1/(i-j)^2; 2(i,j)=1/(i+j)^2 j(n)=(n-j) (scatterer support: [1,2]) k(n)=(n+2-j) (sampled observations) OBSERVATIONS: I I=3 I=4 I=5 I=6 U(I) 5.445 1.796 0.962 0.617 1-D Illustrative Numerical Example “Huge” Linear System of Equations 1 22 5.445 1 1.796 2 3 0.962 1 2 0.617 4 1 52 1 2 1 1 2 2 1 2 3 1 2 4 1 2 4 1 2 5 1 2 6 1 2 7 1 2 5 g v 1 1 1 2 g v 6 1 2 1 g 2 v1 2 7 g v 2 2 1 2 8 1-D Illustrative Numerical Example “Huge” Linear System of Equations Solving this and arranging into matrix: g1v1 g v 2 1 g1v2 3 4 1 3 4 g 2 v2 6 8 2 SOLUTION: V(j)=3(j-1)+4(j-2) [to an unknowable scale factor] 1-D Illustrative Numerical Example “Tiny” Linear System of Equations 1 / 2 1 / 32 BA1 B 2 1 / 4 1 / 52 2 1/1 2 0 1/ 2 2 1 / 3 0 2 1 / 4 2 0 0 .070 .645 .501 .573 B .000 .119 .677 .726 1-D Illustrative Numerical Example “Tiny” Linear System of Equations 2 u ( 3 ) 1 / 4 u ( 4) 1 / 5 2 Y g2 B B 2 u (5) 1 / 6 2 u (6) 1 / 7 1/ 5 2 1 / 6 v1 2 1 / 7 v2 2 1/ 8 2 57.38 4.166 4.048 v1 11.54 0.792 0.849 v 2 1-D Illustrative Numerical Example “Tiny” Linear System of Equations • POINT: Solving “tiny” 2X2 linear system instead of solving “huge” 4X4 linear system • Sufficient statistic Y=B[u]: 4-vector to 2-vector • Null matrix B precomputed from basis functions ahead of time, off-line. Reformulation as Overdetermined Multiparameter Eigenvalue Problem Recall: This form of large linear system: u1 g1v1 ( g A ... g A ) 1 1 N N u NM g N vM Reformulation as Overdetermined Multiparameter Eigenvalue Problem Left-multiply by matrix C where C[u]=[0]: [0]=[g1A1+…+gnAn][v] so that we have: g1A1+…+gnAn is rank-deficient; and: Vec[g1A1+…+gnAn] is linear combination {vec[A1]…vec[An]}=known basis set. Reformulation as Overdetermined Multiparameter Eigenvalue Problem [g1A1+…+gnAn] can be computed iteratively using Lift-and-Project method: 1. Project [g1A1+…+gnAn] rank-deficient using SVD and setting smallest SV to 0; 2. Project vec[g1A1+…+gnAn] onto span{vec[A1]…vec[An]} Reformulation as Overdetermined Multiparameter Eigenvalue Problem Both of these are (Frobenius matrix) norm-reducing operations. By Composite Mapping Theorem, this is guaranteed to converge (maybe to 0!) Problem: Takes long time to converge. CONCLUSION • Solve inverse scattering problem in Born approximation with coincident point source and receiver on moving platform • Using precomputed null vectors, reduce (NM)X(NM) system to MXM system; M=#coefficients representing scatterer; N=#coefficients representing Green’s • Sufficient statistic reduce data dimension FUTURE WORK • Should need much less data: N+M<<NM • Apply the algorithms we are presently developing to solve non-overdetermined multiparameter eigenvalue problem • Sample data for well-conditioned problem: adaptively choose the vehicle trajectory