Earth Sciences / Numerical Analysis and Inversion
/ Topic 5 of 5 · 5/5 Numerical Analysis and Inversion Short Notes One-page CSIR NET geophysics revision: quadrature, FDM and FEM stability, RK4, pattern recognition, ill-posedness, least squares, SVD and global optimisation.
7 min read Updated 16 Nov 2026
Share
Copy link
Print · PDF
Save Share this topic
csirnetnotes.com/earth-sciences/numerical-analysis-and-inversion/quick-revision
Copy link Ctrl C Pre-written message Numerical Analysis and Inversion Short Notes — notes, formulas and practice, free. csirnetnotes.com/earth-sciences/numerical-analysis-and-inversion/quick-revision
On this page · 7 sections▾ Part C in one table 1. Differentiation, integration, FDM, FEM 2. Initial value problems and pattern recognition 3. Ill-posed problems and least squares 4. Optimisation, SVD and global methods Numbers worth recognising Where to go next Every formula and trap from Numerical Differentiation, Integration, FDM and FEM , Initial Value Problems and Pattern Recognition , Well- and Ill-Posed Problems and Least Squares and Gradient and Direct Search, Generalized Inversion, SVD and Global Optimization , on one page.
Part C in one table Per the NTA information bulletin for December 2026, Earth Sciences (702):
Given Attempt Correct Wrong Maximum 80 25 +4 −1.32 100
Only the first 25 answered Part C questions are evaluated. Marking for the whole paper is on the Earth Sciences overview .
1. Differentiation, integration, FDM, FEM Formula Expression Error Forward difference \( \frac{f_{i+1} - f_i}{h} \) \( O(h) \) Central, first derivative \( \frac{f_{i+1} - f_{i-1}}{2h} \) \( -\frac{h^2}{6}f''' \) Central, second derivative \( \frac{f_{i+1} - 2f_i + f_{i-1}}{h^2} \) \( -\frac{h^2}{12}f^{(4)} \) Richardson \( \frac{2^pD(h) - D(2h)}{2^p - 1} \) removes the \( h^p \) term
$$\boxed{T = h\left[\tfrac{1}{2}f_0 + f_1 + \cdots + f_{n-1} + \tfrac{1}{2}f_n\right]} \qquad \boxed{S = \frac{h}{3}\left[f_0 + 4(\text{odd}) + 2(\text{even interior}) + f_n\right]}$$
Rule Composite error Degree of precision Trapezoidal \( -\frac{(b-a)h^2}{12}f'' \) 1 Simpson 1/3 (even panels) \( -\frac{(b-a)h^4}{180}f^{(4)} \) 3 Simpson 3/8 \( -\frac{(b-a)h^4}{80}f^{(4)} \) 3 Gauss–Legendre, \( n \) points \( 2n - 1 \)
Two-point Gauss on \( [-1, 1] \): nodes \( \pm\frac{1}{\sqrt 3} \), weights 1. Three-point: \( 0, \pm\sqrt{3/5} \), weights \( \frac{8}{9}, \frac{5}{9}, \frac{5}{9} \). Other intervals: multiply by \( \frac{b-a}{2} \). Halving \( h \): trapezoid error ÷ 4, Simpson ÷ 16. Romberg step on the trapezoid = Simpson. Differentiation amplifies noise (high-pass); integration smooths it. $$\boxed{r = \frac{\kappa\Delta t}{\Delta x^2} \le \frac{1}{2}\ \text{(explicit heat)}} \qquad \boxed{C = \frac{v_{\max}\Delta t}{\Delta x} \le 1\ \text{(1-D wave; } 1/\sqrt{2}\text{ on a square 2-D grid)}}$$
Scheme (heat equation) Order Stability FTCS explicit \( O(\Delta t, \Delta x^2) \) \( r \le \frac{1}{2} \) BTCS implicit \( O(\Delta t, \Delta x^2) \) unconditional Crank–Nicolson \( O(\Delta t^2, \Delta x^2) \) unconditional
Laplace 5-point stencil: node = mean of four neighbours. Lax: consistency + stability ⇔ convergence (well-posed linear IVP).FEM: weak form, shape functions, Galerkin (test = basis), element matrices, assembly, BCs, solve \( K\mathbf{u} = \mathbf{F} \). 1-D linear element stiffness \( \frac{1}{h}\begin{pmatrix}1 & -1\\ -1 & 1\end{pmatrix} \). \( K \) sparse, symmetric, positive definite (self-adjoint problems). Neumann = natural, Dirichlet = essential.⚠ Trap · Traps
Simpson on an odd number of panels; the slowest velocity in the CFL check; “implicit means accurate”; forgetting \( \frac{b-a}{2} \) in Gauss quadrature; stability confused with numerical dispersion.
2. Initial value problems and pattern recognition Method Update Global order \( f \) evaluations Euler \( y_n + hf_n \) 1 1 Heun (modified Euler) \( y_n + \frac{h}{2}[f(t_n, y_n) + f(t_{n+1}, y_n + hf_n)] \) 2 2 RK4 \( y_n + \frac{h}{6}(k_1 + 2k_2 + 2k_3 + k_4) \) 4 4 Backward Euler \( y_n + hf(t_{n+1}, y_{n+1}) \) 1 implicit solve
Local error = one power of \( h \) higher than global. Halve \( h \): global error ÷ \( 2^p \). One step of \( y' = y \), \( h = 0.1 \): Euler 1.1, Heun 1.105, RK4 1.1051708 (exact 1.1051709). Higher-order ODE → first-order system. BVP → IVP by shooting . Multistep (Adams–Bashforth / Adams–Moulton, predictor–corrector): cheap per step, not self-starting. $$\boxed{\text{Euler stable: } \lvert 1 + h\lambda \rvert \le 1 \Rightarrow h \le \frac{2}{\lvert\lambda\rvert}\ (\lambda < 0)}$$
Heun: \( -2 \le h\lambda \le 0 \); RK4: about \( -2.79 \le h\lambda \le 0 \); backward Euler: A-stable. Stiff = widely separated time scales; use implicit methods.
Pattern recognition Key point Supervised labelled training data: nearest mean, k-NN, Bayes / discriminant, perceptron / ANN, SVM Unsupervised no labels: k-means, hierarchical clustering, self-organising maps Distances standardise features; Mahalanobis \( (\mathbf{x} - \boldsymbol\mu)^{\mathsf T}\Sigma^{-1}(\mathbf{x} - \boldsymbol\mu) \) k-means assign, update, repeat; converges to a local minimum Scores accuracy \( \frac{TP + TN}{N} \), precision \( \frac{TP}{TP + FP} \), recall \( \frac{TP}{TP + FN} \) Overfitting test on held-out data; cross-validate
⚠ Trap · Traps
Wrong argument in \( k_2 \); calling Euler second order because its local error is \( O(h^2) \); A-stable taken to mean accurate; clustering unscaled logs; quoting training accuracy; a single perceptron cannot separate non-linearly-separable classes.
3. Ill-posed problems and least squares Hadamard: existence, uniqueness, stability. Lose any one and the problem is ill-posed.
Ill-posed example Fails Potential-field sources (equivalent sources) uniqueness Downward continuation, factor \( e^{k\Delta z} \), \( k = 2\pi/\lambda \) stability Deconvolution near spectral zeros stability Noisy over-determined data exact existence
$$\boxed{\mathbf{m}_{\text{LS}} = (G^{\mathsf T}G)^{-1}G^{\mathsf T}\mathbf{d}} \qquad \boxed{\mathbf{m}_{\text{MN}} = G^{\mathsf T}(GG^{\mathsf T})^{-1}\mathbf{d}} \qquad \boxed{\mathbf{m}_{\text{damped}} = (G^{\mathsf T}G + \varepsilon^2I)^{-1}G^{\mathsf T}\mathbf{d}}$$
Weighted: \( (G^{\mathsf T}WG)^{-1}G^{\mathsf T}W\mathbf{d} \), \( W = C_d^{-1} \) (weights \( 1/\sigma_i^2 \)). Covariance: \( C_m = \sigma_d^2(G^{\mathsf T}G)^{-1} \). Line \( d = a + bx \): \( \begin{pmatrix} N & \sum x \\ \sum x & \sum x^2\end{pmatrix}\begin{pmatrix}a\\ b\end{pmatrix} = \begin{pmatrix}\sum d\\ \sum xd\end{pmatrix} \); residuals sum to zero. Condition number \( \kappa = \sigma_{\max}/\sigma_{\min} \) multiplies the relative data error. L2 → mean (Gaussian errors); L1 → median (robust). Good fit: \( \chi^2 = \sum r_i^2/\sigma_i^2 \approx N - M \). Damping: stability for bias; choose \( \varepsilon \) by L-curve or discrepancy principle. Non-linear: linearise with the Jacobian, solve, update, repeat. ⚠ Trap · Traps
More data does not remove potential-field non-uniqueness; count independent equations (rank), not rows; \( G^{\mathsf T}G \) versus \( GG^{\mathsf T} \); fitting noise to zero misfit; treating damping as free.
4. Optimisation, SVD and global methods $$\boxed{\text{Newton: } H\Delta\mathbf{m} = -\mathbf{g}} \qquad \boxed{\text{Gauss–Newton: } J^{\mathsf T}J\Delta\mathbf{m} = J^{\mathsf T}\mathbf{r}} \qquad \boxed{\text{Marquardt: } (J^{\mathsf T}J + \lambda I)\Delta\mathbf{m} = J^{\mathsf T}\mathbf{r}}$$
Method Remember Steepest descent \( -\mathbf{g} \) direction; zig-zags (successive gradients perpendicular with exact line search) Newton one step on a quadratic; quadratic convergence Marquardt step succeeds: \( \lambda \downarrow \) (towards GN); fails: \( \lambda \uparrow \) (towards steepest descent) Conjugate gradient at most \( M \) steps on a quadratic; LSQR for large sparse tomography Grid search \( n^M \) forward models; global on the grid Nelder–Mead simplex \( M + 1 \) vertices; reflect, expand, contract, shrink; no derivatives
$$\boxed{G = U\Lambda V^{\mathsf T}} \qquad \boxed{G^{\dagger} = V_p\Lambda_p^{-1}U_p^{\mathsf T}} \qquad \boxed{\mathbf{m} = \sum_{i=1}^{p}\frac{\mathbf{u}_i^{\mathsf T}\mathbf{d}}{\sigma_i}\mathbf{v}_i}$$
\( \sigma_i^2 \) = eigenvalues of \( G^{\mathsf T}G \); rank \( p \) = number of non-zero \( \sigma_i \). \( V_0 \): model null space (invisible models). \( U_0 \): data null space (unfittable data). \( R = V_pV_p^{\mathsf T} \) (model), \( N = U_pU_p^{\mathsf T} \) (data); both have trace \( p \); perfect = \( I \). \( C_m = \sigma_d^2V_p\Lambda_p^{-2}V_p^{\mathsf T} \): small \( \sigma_i \), big variance. Truncated SVD: filter 1 or 0. Damped: \( \frac{\sigma_i^2}{\sigma_i^2 + \varepsilon^2} \). Rank-one example \( G = \begin{pmatrix}3 & 0\\ 4 & 0\end{pmatrix} \): \( \sigma_1 = 5 \), \( G^{\dagger} = \begin{pmatrix}0.12 & 0.16\\ 0 & 0\end{pmatrix} \), \( R = \mathrm{diag}(1, 0) \). Global method Core rule Monte Carlo random models from prior ranges; keep acceptable ones; ensemble shows non-uniqueness Simulated annealing Metropolis: accept uphill with \( e^{-\Delta E/T} \); cool slowly Genetic algorithm population; selection, crossover, mutation; \( n \) bits give resolution \( \frac{\text{range}}{2^n - 1} \)
⚠ Trap · Traps
Small misfit is not uniqueness; Marquardt rule reversed; keeping singular values below the noise; “global methods guarantee the global minimum”; model versus data resolution; shapes of \( U \) (\( N \times N \)), \( V \) (\( M \times M \)), \( G^{\dagger} \) (\( M \times N \)).
Numbers worth recognising Quantity Value Central difference of \( e^x \) at 0, \( h = 0.1 \) 1.0016675 (error \( \approx h^2/6 \)) Explicit heat step, \( \kappa = 10^{-6} \) m²/s, \( \Delta x = 10 \) m \( \Delta t \le 5 \times 10^7 \) s, about 1.6 yr CFL, \( v = 3000 \) m/s, \( \Delta x = 15 \) m \( \Delta t \le 5 \) ms \( e^{\pi} \) (downward continuation, \( \lambda = 1 \) km, \( \Delta z = 0.5 \) km) about 23 \( e^{-2} \), \( e^{-0.5} \), \( e^{-4} \) 0.135, 0.607, 0.018 \( 50^4 \), \( 50^8 \) \( 6.25 \times 10^6 \), \( 3.9 \times 10^{13} \)
Where to go next Advertisement Advertisement