Regularisation and Stability in Calibration

Hard·25 min read
CalibrationRegularisationTikhonovInverse ProblemsNumerical Stability

Quick Quiz

1. With J=UΣVJ=U\Sigma V^\top, the least-squares step is δθLS=jujδσ^σjvj\delta\theta_{\mathrm{LS}}=\sum_j \frac{u_j^\top\delta\hat\sigma}{\sigma_j}v_j. Why do near-zero singular values σj\sigma_j destabilise calibration?
2. Tikhonov regularisation has SVD filter factors fj(λ)=σj2/(σj2+λ)f_j(\lambda)=\sigma_j^2/(\sigma_j^2+\lambda). What happens to directions with large singular values (σjλ\sigma_j\gg\sqrt\lambda)?
3. Morozov's discrepancy principle sets r(θλ)ϵN\|r(\theta_\lambda)\|\approx\epsilon\sqrt N. With N=50N=50 instruments and half-spread ϵ=0.002\epsilon=0.002 vol points, the target residual norm is:
4. Adding a Tikhonov penalty λθθprev2\lambda\|\theta-\theta_{\mathrm{prev}}\|^2 (with θprev\theta_{\mathrm{prev}} = yesterday's calibration) removes calibration bias on days when the market moves sharply.
5. On an L-curve (residual norm vs regularisation norm), what characterises the 'horizontal arm' at small λ\lambda?
6. Penalising surface curvature λ1kkσloc2\lambda_1\|\partial_{kk}\sigma_{\mathrm{loc}}\|^2 in local-vol calibration. What is the main failure when λ1\lambda_1 is too large?