= Quadratic norm penalty
{title2=$\alpha\|x\|^2$}
For a <bounded linear operator> $A$ between <Hilbert spaces>, a quadratic penalty gives
$$
J_\alpha(x)=\|Ax-y\|^2+\alpha\|x\|^2,\qquad\alpha>0.
$$
Its <first variation> in direction $h$ is $2\langle A^*(Ax-y)+\alpha x,h\rangle$ in the real case; use the real part of the pairing in the complex case. Dividing stationarity by two gives $(A^*A+\alpha I)x=A^*y$. Thus the derivative of the displayed penalty itself is $2\alpha x$, while its contribution to the normalized <normal equation for a linear inverse problem> is $\alpha x$.
The positive quadratic term makes the objective <strongly convex> and coercive, yielding a unique minimizer. In a <singular system of a compact operator>, the data-to-solution coefficient is $\sigma/(\sigma^2+\alpha)$ instead of $1/\sigma$. Small <singular values> are suppressed, trading approximation bias against amplification of <measurement noise>. This is <Tikhonov regularization>, rather than exact minimization of the residual alone.
Back to article page