Levenberg-Marquardt for Model Calibration

Hard·25 min read
CalibrationNumerical OptimisationHeston ModelImplied Volatility

Quick Quiz

1. In the LM update (JJ+λD)δθ=Jr(J^\top J+\lambda D)\,\delta\theta=-J^\top r, how does the step behave as λ\lambda\to\infty?
2. Marquardt's scaling uses D=diag(JJ)D=\mathrm{diag}(J^\top J) instead of D=ID=I. Its key advantage is that it:
3. The gain ratio ρ=F(θ)F(θ+δθ)L(θ)L(θ+δθ)\rho=\dfrac{F(\theta)-F(\theta+\delta\theta)}{\mathcal{L}(\theta)-\mathcal{L}(\theta+\delta\theta)} decides step acceptance. What does ρ<0\rho<0 indicate?
4. Satisfying the first-order condition Jr<ε\|J^\top r\|_\infty<\varepsilon in LM guarantees the calibrated parameters are the global minimum of the objective.
5. A central-difference Jacobian has truncation error O(h2)O(h^2) plus round-off from finite precision. What determines the optimal step size hh?
6. In Heston calibration, κ\kappa and νˉ\bar\nu (long-run variance) are often poorly identified. The formal characterisation is: