Definitions¶
Matrices¶
Starting of, we have the general Tikhonov function
where \(\Lambda\) is the regularization matrix, and \(W\) is the weight matrix. We want to solve for \(x\) such that \(J\) is minimal:
which yields
Alternative derivation: defining the residue \(r = Ax - y\), we instead find
we can substitute this solution for \(x\) back into \(r\):
The simplest choice of \(\Lambda = \lambda I\), in which case this equation simplifies to
We introduce the following notation:
This notation is chosen because \(H_x\) is the Hessian of \(x\), and \(H_y = A A^T\) can be thought of as the Hessian of \(y\). \(R_y\) is so named because it is the regularized version of \(A\).
Multiple Datasets¶
In the previous section \(y\) was assumed to be a vector. (Technically, a \((N_y, 1)\)-matrix.) However, it is perfectly allowed to regularize multiple data sets at once by turning it into a \((N_y, N_{sets})\)-matrix, where \(N_{sets}\) is the number of data sets. The function \(J\) then becomes
Functionals¶
Things get truly interesting, and surprisingly simple, when we work with functionals instead. We start from
where \(A_i(t)\) is the kernel of integral, for example \(e^{- s_i t}\) for a Laplace transform. As always, there is some ambiguity/freedom in the shape of \(\Lambda\). Here it is written as a scalar function, but it could also be chosen as a constant, or as a function with index \(i\).
Repeating the same steps as above, we find that
which leads to
where \(M_ij = \int_{-\infty}^{\infty} \frac{A_j(t) A_i(t)}{\Lambda(t)^2} dt\).