Phil B writes:
My day job does include parameter and state estimation using Least Squares and Kalman filtering.
I have replicated several of the non-ice borehole temperature reconstructions and I’ll “share” my observations. The inversion problem boils down to finding the solution to the inconsistent set of linear equations Ax~b, where A is a skinny matrix (mxn) whose columns are generated from the solution of the heat conduction equation. x is a nx1 solution vector where the first two elements are a slope and intercept and the remaining elements are the temperature reconstruction. The b mx1 vector is the borehole temperature profile. The solution is calculated using the pseudo inverse based on the singular value decomposition (svd) where A = U*S*V’ and where U (mxm) and V (nxn) are complete orthonormal basis sets. The A matrix is ill-conditioned with ratio of max to min singular value on the order of 1e6 to 1e7 on the boreholes I replicated. In the older literature only singular values of greater than 0.3 are used in the pseudo inverse, with recent literature using a ridge regression that optimizes the norm of the residual versus the norm of the solution. If you keep singular values that are less than 0.3, the reconstruction is physical unreasonable, i.e. pulses on the order of 20-40 deg K. For 500 year reconstruction and 100 year steps, the 3 smallest singular values out of the seven total aren’t used in the psuedoinverse.
So what is the problem??? The ill-conditioning of A!! For instance if A is rank deficient i.e. rank = n-1, then one has a single null vector z such that A*z = 0_mx1 and an infinite number of solutions for x. For our A, there are 3 “almost” null vectors which are the last three columns of V. So A*v(5), A*v(6), A*v(7)~0_mx1. Let x_est be the pseudo inverse solution using only singular values greater than 0.3. Let’s create a new solution x_new = x_est + 2*v(7). The individual residual changes are on the order of millikelvins. x_new “looks” substantial different then x_est but the values are reasonable. The point is that there are many x_estimates and many reasonable temperature reconstructions that have residuals that are almost identical, with differences that are less then the temperature sensor noise level.
Summarizing, the columns of the ill-conditioned A matrix are created using solutions to the heat conducting equation. x_est is one of many possible temperature histories.