Computational Instability and Filtering
CSIR NET meteorology: CFL condition, von Neumann analysis, time schemes, leapfrog computational mode, Robert–Asselin filter, aliasing and filtered equations.
- 25 min
On this page · 15 sections
- Prerequisites
- The Hook: A Forecast of 145 hPa in Six Hours
- From Equations to Numbers
- The Two Test Equations
- Space Differencing and the CFL Condition
- Computational Dispersion: Phase Errors of Centred Differences
- Nonlinear Computational Instability and Aliasing
- Meteorological Filtering: Changing the Equations
- Memory Tricks
- Don’t Confuse
- Common Mistakes
- Practice
- Quick Revision Table
- Key Takeaways
- What’s Next
Prerequisites
- Finite differences, truncation error and the CFL idea in a geophysical setting: Numerical Differentiation, Integration, FDM and FEM.
- Euler and Runge–Kutta time stepping and the notion of a stability region: Initial Value Problems and Pattern Recognition, and the mathematicians’ version in Picard, Euler and Runge–Kutta Methods.
- Aliasing of sampled signals, which reappears here as nonlinear instability: Sampling Theorem, FFT, Z-Transform and Convolution.
- Hydrostatic balance and the scale height, used in the filtering argument: Composition, Thermal Structure, Hydrostatics and the Standard Atmosphere.
- Chapter map: Numerical Weather Prediction.
The Hook: A Forecast of 145 hPa in Six Hours
From Equations to Numbers
A numerical model replaces continuous fields by values on a grid (or by a finite set of spectral coefficients) and replaces derivatives by differences. For a field \( u(x, t) \) on a grid \( x_j = j\Delta x \), \( t_n = n\Delta t \), write \( u_j^n \).
Three properties decide whether a scheme is any use:
| Property | Meaning | How you check it |
|---|---|---|
| Consistency | The difference equation tends to the differential equation as \( \Delta x, \Delta t \to 0 \) | Taylor expansion: truncation error \( \to 0 \) |
| Stability | The numerical solution stays bounded for a fixed final time as the grid is refined (no spurious growth) | von Neumann (Fourier) analysis, energy method |
| Convergence | The numerical solution tends to the true solution | Lax equivalence theorem: for a well-posed linear problem, consistency + stability ⇔ convergence |
A scheme can be consistent and still useless if it is unstable. That is the whole subject of computational instability: errors (round-off, truncation, initial errors) are amplified by the scheme itself rather than by the physics.
The von Neumann method
For a linear equation with constant coefficients, try a single Fourier mode
$$u_j^n = \lambda^n e^{ikj\Delta x},$$substitute it into the difference equation and solve for the amplification factor \( \lambda \) (for a two-level scheme) or factors (for a three-level scheme). The scheme is stable if
$$\lvert \lambda \rvert \le 1 \quad \text{for all resolvable } k$$(strictly, \( \lvert\lambda\rvert \le 1 + O(\Delta t) \) is allowed when the true solution itself grows, but for the advection and oscillation equations the strict condition is the one used). If \( \lvert\lambda\rvert = 1 \) the scheme is neutral; if \( \lvert\lambda\rvert < 1 \) it is damping (dissipative); the phase of \( \lambda \) compared with the true phase gives the phase error.
The Two Test Equations
Meteorological models are judged on two simple equations that carry the essential behaviour of the full set.
The oscillation equation (time differencing alone):
$$\frac{du}{dt} = i\omega u, \qquad u(t) = u_0 e^{i\omega t}.$$Inertia–gravity oscillations, and each Fourier mode of advection, obey it. Define \( p = \omega\Delta t \). The exact solution has \( \lvert\lambda\rvert = 1 \) and phase change \( p \) per step.
The linear advection equation (space and time together):
$$\frac{\partial u}{\partial t} + c\frac{\partial u}{\partial x} = 0, \qquad u = f(x - ct).$$Define the Courant number \( \mu = c\Delta t/\Delta x \) (some texts write \( C \) or \( \alpha \)).
Time schemes on the oscillation equation
| Scheme | Formula | \( \lvert\lambda\rvert \) | Verdict |
|---|---|---|---|
| Forward (Euler) | \( u^{n+1} = u^n + i p\, u^n \) | \( \sqrt{1 + p^2} \) | Always unstable (amplifies every step) |
| Backward (implicit) | \( u^{n+1} = u^n + i p\, u^{n+1} \) | \( 1/\sqrt{1 + p^2} \) | Unconditionally stable, damping |
| Trapezoidal (Crank–Nicolson) | \( u^{n+1} = u^n + \tfrac{ip}{2}(u^n + u^{n+1}) \) | 1 | Unconditionally stable, neutral, second order |
| Matsuno (Euler-backward) | predictor forward, corrector backward | \( \sqrt{1 - p^2 + p^4} \) | Stable for \( \lvert p\rvert \le 1 \); damps high frequencies most |
| Heun | predictor forward, corrector trapezoidal | \( \sqrt{1 + p^4/4} \) | Weakly unstable (slow growth) |
| Leapfrog (centred) | \( u^{n+1} = u^{n-1} + 2ip\, u^n \) | 1 (both roots) | Neutral for \( \lvert p\rvert \le 1 \); has a computational mode |
Two schemes from this table dominate meteorology. Leapfrog is cheap (one function evaluation per step), second order and neutral, and it has been the workhorse of grid-point models. Matsuno’s selective damping of high frequencies makes it useful for dynamic initialization (page 4) and as an occasional “restart” step.
Why leapfrog has two solutions
Leapfrog is a three-level scheme: it needs \( u^{n-1} \) and \( u^n \) to give \( u^{n+1} \). Substituting \( u^n = \lambda^n u^0 \) gives a quadratic:
$$\lambda^2 - 2ip\lambda - 1 = 0 \quad\Rightarrow\quad \lambda_{1,2} = ip \pm \sqrt{1 - p^2}.$$- For \( \lvert p\rvert \le 1 \), both roots have \( \lvert\lambda\rvert = 1 \): neutral.
- For \( \lvert p\rvert > 1 \), the roots are purely imaginary and one has \( \lvert\lambda\rvert > 1 \): unstable.
The physical mode \( \lambda_1 = \sqrt{1-p^2} + ip \) tends to 1 as \( \Delta t \to 0 \) and approximates the true solution. Its phase per step is \( \theta = \arcsin p \), which exceeds \( p \): leapfrog makes oscillations run slightly fast (for \( p = 0.5 \), by about 4.7%).
The computational mode \( \lambda_2 = -\sqrt{1-p^2} + ip \) tends to \( -1 \) as \( \Delta t \to 0 \). It flips sign every step. It has no physical meaning: it exists only because the scheme needs two starting values while the equation needs one. In a linear problem it stays at whatever amplitude the start-up gave it. In a nonlinear model it is fed by the physics and the two interleaved sequences (odd and even steps) drift apart, a problem called time splitting (or the separation of solutions).
The Robert–Asselin time filter
The standard cure is a weak three-point smoothing in time applied after each leapfrog step (Robert’s idea, analysed by Asselin):
$$\bar{u}^{n} = u^{n} + \nu\left(\bar{u}^{n-1} - 2u^{n} + u^{n+1}\right).$$The bracket is a second difference in time. It is large for a signal that flips sign every step (the computational mode) and small for a slowly varying signal (the physical mode), so the filter damps the computational mode strongly and the physical mode only slightly. The coefficient \( \nu \) is small. The price is that the scheme’s accuracy drops from second to first order in time, and the physical mode is weakly damped. Refinements of the filter that limit this damage exist in the literature; the exam needs only the idea.
Other cures: start each leapfrog sequence with a forward or Matsuno step, and restart periodically; or use a two-level scheme (no computational mode at all).
Space Differencing and the CFL Condition
Schemes for the advection equation
| Scheme | Difference equation | Stability | Character |
|---|---|---|---|
| Forward time, centred space (FTCS) | \( u_j^{n+1} = u_j^n - \tfrac{\mu}{2}(u_{j+1}^n - u_{j-1}^n) \) | Unconditionally unstable | Do not use |
| Upstream (upwind), \( c > 0 \) | \( u_j^{n+1} = u_j^n - \mu(u_j^n - u_{j-1}^n) \) | \( 0 \le \mu \le 1 \) | First order, strongly damping |
| Leapfrog, centred space | \( u_j^{n+1} = u_j^{n-1} - \mu(u_{j+1}^n - u_{j-1}^n) \) | \( \lvert\mu\rvert \le 1 \) | Second order, neutral, computational mode |
| Lax–Wendroff | two-step, second order | \( \lvert\mu\rvert \le 1 \) | Second order, mildly damping for short waves |
| Implicit (centred in time) | trapezoidal in time | unconditional | Needs a matrix solve each step |
The upstream scheme’s damping comes from its truncation error, which acts like a diffusion term with coefficient \( \tfrac{c\Delta x}{2}(1 - \mu) \). That is why it is stable and why it smears sharp features.
Leapfrog on the advection equation
Substitute \( u_j^n = \lambda^n e^{ikj\Delta x} \):
$$\lambda^2 + 2i\mu\sin(k\Delta x)\,\lambda - 1 = 0 \quad\Rightarrow\quad \lambda = -i\mu\sin(k\Delta x) \pm \sqrt{1 - \mu^2\sin^2(k\Delta x)}.$$This is the oscillation-equation result with \( p = -\mu\sin(k\Delta x) \). Stability needs \( \lvert\mu\sin(k\Delta x)\rvert \le 1 \) for every \( k \); the worst case \( \sin(k\Delta x) = 1 \) (the \( 4\Delta x \) wave) gives
$$\boxed{\mu = \frac{c\,\Delta t}{\Delta x} \le 1}$$This is the Courant–Friedrichs–Lewy (CFL) condition (1928). Its physical meaning: in one time step, information must not travel further than the scheme’s stencil can see. If the true domain of dependence lies outside the numerical one, no scheme can converge.
In two dimensions on a square grid with centred differences the limit tightens (for a wave moving diagonally) to about \( c\Delta t/\Delta x \le 1/\sqrt{2} \). Always use the fastest signal speed in the model.
Which speed sets the time step?
| Wave | Typical speed | Comment |
|---|---|---|
| Sound waves | about 300–340 m/s | Removed by the hydrostatic approximation (vertically propagating ones) or by anelastic equations |
| External gravity wave (Lamb-like external mode) | \( \sqrt{gH} \approx 300 \) m/s for \( H \approx 9 \)–10 km | Present in primitive-equation models; usually the binding limit |
| Internal gravity waves | tens of m/s, up to about 100 m/s for deep modes | |
| Jet-stream advection | up to about 100 m/s | Binding limit once gravity waves are treated implicitly |
| Rossby waves | a few to about 15 m/s relative to the ground | Carry the weather |
The meteorologically important motions move at 10–20 m/s, yet an explicit primitive-equation model must take steps short enough for 300 m/s waves. With \( \Delta x = 100 \) km, \( \Delta t \le 100\,000/300 \approx 333 \) s in 1-D. This mismatch drove three major developments:
- Semi-implicit schemes: treat the terms responsible for gravity waves (pressure gradient and divergence) implicitly, the rest (advection) explicitly. The gravity-wave limit disappears; the advective CFL (based on the jet speed) remains. Time steps grow by a factor of several. The cost is an elliptic (Helmholtz) equation to solve every step, which is cheap in spectral models.
- Semi-Lagrangian advection: follow fluid parcels back along trajectories to departure points and interpolate there. The advective CFL limit disappears; accuracy, not stability, now limits \( \Delta t \). Semi-implicit semi-Lagrangian models run with time steps several times longer than Eulerian ones at the same resolution.
- Filtered equations: remove the fast waves from the equations altogether (below).
The pole problem
On a latitude–longitude grid the east–west spacing \( a\cos\phi\,\Delta\lambda \) shrinks towards the poles, so the CFL limit near the pole forces a tiny time step for the whole globe. Remedies: Fourier filtering of the short zonal waves at high latitudes (a numerical filter again), reduced (thinned) grids with fewer points per latitude circle near the poles, spectral methods (the spherical-harmonic representation has uniform resolution), and quasi-uniform grids (icosahedral, cubed-sphere).
Computational Dispersion: Phase Errors of Centred Differences
Even a stable scheme can be inaccurate. Using the centred difference \( (u_{j+1} - u_{j-1})/2\Delta x \) on the advection equation (exact in time) gives a numerical phase speed
$$\boxed{c^{*} = c\,\frac{\sin(k\Delta x)}{k\Delta x}}$$| Wavelength \( L \) | \( k\Delta x = 2\pi\Delta x/L \) | \( c^{*}/c \) |
|---|---|---|
| \( 2\Delta x \) | \( \pi \) | 0 (stationary) |
| \( 3\Delta x \) | \( 2\pi/3 \) | 0.41 |
| \( 4\Delta x \) | \( \pi/2 \) | 0.64 |
| \( 8\Delta x \) | \( \pi/4 \) | 0.90 |
| \( \gg \Delta x \) | \( \to 0 \) | \( \to 1 \) |
Consequences:
- Short waves lag. A wave must span roughly 8 or more grid lengths to be advected with a phase error below about 10%. The \( 2\Delta x \) wave does not move at all.
- Computational dispersion. Phase speed depends on wavelength although the true equation is non-dispersive, so a sharp feature breaks into a train of trailing short waves.
- Wrong group velocity. The numerical group velocity is \( c\cos(k\Delta x) \), which is negative for \( k\Delta x > \pi/2 \): energy of waves shorter than \( 4\Delta x \) travels upstream.
- Fourth-order differences reduce the error for medium waves; staggered grids reduce it for gravity waves; spectral methods remove it for all resolved waves.
The shortest wave a grid can represent is \( 2\Delta x \), and in practice features smaller than several \( \Delta x \) are poorly forecast. “Effective resolution” is always coarser than grid spacing.
Staggered grids
Arakawa’s classification (grids A to E) places \( u \), \( v \) and \( h \) (or \( \Phi \)) at the same or different points. The choice changes how well geostrophic adjustment and gravity waves are represented. The unstaggered A grid handles the \( 2\Delta x \) wave badly (a checkerboard of height can sit with no pressure-gradient force). The C grid (velocities between mass points) represents gravity waves and adjustment well when the grid length is smaller than the Rossby radius of deformation, and is widely used in grid-point models.
Nonlinear Computational Instability and Aliasing
Linear stability is not enough. In the 1950s N. A. Phillips found that a long numerical integration of the (quasi-geostrophic) vorticity equations developed grid-scale noise and blew up even though its time step satisfied the linear CFL condition; he later analysed the effect (his 1959 paper) in a barotropic vorticity setting. The cause was the nonlinear advection term.
The aliasing mechanism
A product of two waves contains their sum wavenumber:
$$\sin(k_1 x)\sin(k_2 x) = \tfrac{1}{2}\left[\cos(k_1 - k_2)x - \cos(k_1 + k_2)x\right].$$On a grid, the shortest resolvable wave is \( 2\Delta x \), with wavenumber \( k_{\max} = \pi/\Delta x \). If \( k_1 + k_2 > k_{\max} \), the grid cannot represent the new wave. Sampled at the grid points it is indistinguishable from a longer wave with wavenumber
$$\boxed{k^{*} = \frac{2\pi}{\Delta x} - k}$$This is aliasing, the same effect as in signal processing. Energy that should cascade to scales below the grid is folded back into the shortest resolved scales, usually wavelengths of 2–4\( \Delta x \). Over many steps it accumulates there, the field becomes noisy, and the model blows up. Making \( \Delta t \) smaller does not help: nonlinear instability is not cured by reducing the time step.
Remedies
| Remedy | How it works |
|---|---|
| Conservative (Arakawa) Jacobian | A finite-difference form of the advection term \( J(\psi, \zeta) \) that conserves both kinetic energy and enstrophy (mean-square vorticity) in the discrete system, as the continuous equations do. Energy cannot pile up at the shortest scales, so the instability cannot occur. |
| Spectral transform with an alias-free grid | Compute products on a grid fine enough that quadratic products of resolved waves do not alias back (the “3/2 rule”: at least \( 3N + 1 \) points around a latitude circle for truncation \( N \)). |
| Smoothing and filtering | Periodic spatial filters (for example the Shapiro filter) that remove the \( 2\Delta x \) wave and damp short waves while leaving long ones almost untouched. |
| Horizontal diffusion | Add \( K\nabla^2 \) or, more selectively, hyperdiffusion \( -K\nabla^4 \) (damping rate proportional to \( k^4 \)), which removes energy at the shortest scales where it piles up. |
| Damping schemes | Upstream or Lax–Wendroff advection, which damp the shortest waves. |
Meteorological Filtering: Changing the Equations
The full compressible equations admit three families of waves:
| Wave | Restoring mechanism | Typical speed | Meteorological importance |
|---|---|---|---|
| Sound (acoustic) | Compressibility | about 300+ m/s | Negligible for weather |
| Gravity (inertia–gravity) | Buoyancy (stable stratification), plus rotation | up to about 300 m/s for the external mode | Real but carry little energy at synoptic scale; matter for adjustment and mesoscale |
| Rossby (planetary) | Variation of \( f \) with latitude (\( \beta \)), or of PV | about 10 m/s | The weather systems themselves |
A filtered model is one whose equations, through an approximation, no longer allow some of these waves. The faster waves then cannot limit the time step, and cannot be excited as noise by bad initial data.
| Approximation | Waves removed | Waves kept | Where used |
|---|---|---|---|
| Hydrostatic \( \partial p/\partial z = -\rho g \) | Vertically propagating sound waves | Horizontally propagating external acoustic (Lamb) mode, gravity, Rossby | Primitive-equation models at grid lengths above about 10 km |
| Anelastic / Boussinesq (\( \nabla\cdot(\bar\rho\mathbf{V}) = 0 \)) | All sound waves | Gravity (including non-hydrostatic), Rossby | Cloud and mesoscale research models |
| Quasi-geostrophic (geostrophic wind in the advection and vorticity terms) | Sound and gravity waves | Rossby waves only | Early baroclinic models; theory |
| Non-divergent barotropic | Sound, gravity, baroclinic modes | Barotropic Rossby waves | The 1950 ENIAC forecasts |
Why quasi-geostrophy filters gravity waves: gravity waves involve divergence and imbalance between the wind and the pressure field. The QG system diagnoses the wind from the geopotential (geostrophic relation) and determines divergence diagnostically from the omega equation, so there is no prognostic equation for the divergent, unbalanced part of the flow. Only the slow, balanced (Rossby) evolution remains.
Filtered models made NWP possible on the computers of the 1950s. Their price is accuracy: the QG approximation is poor in the tropics (where \( f \to 0 \)), near fronts and over steep terrain, and fails to describe adjustment. Once semi-implicit methods removed the time-step penalty and initialization removed the noise problem, operational centres moved to primitive-equation models, which keep gravity waves but control them numerically. The model hierarchy is on the next topic page.
graph TD
A["Full compressible equations
sound + gravity + Rossby"] -->|hydrostatic approximation| B["Primitive equations
gravity + Rossby
(+ Lamb mode)"]
A -->|anelastic / Boussinesq| C["Non-hydrostatic, soundproof
gravity + Rossby"]
B -->|quasi-geostrophic approximation| D["QG baroclinic model
Rossby only"]
D -->|single level, non-divergent| E["Barotropic vorticity model
barotropic Rossby waves"]Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| Linear computational instability: cured by reducing \( \Delta t \) (CFL) | Nonlinear instability: aliasing; reducing \( \Delta t \) does not help | |
| Physical mode of leapfrog: \( \lambda \to 1 \) as \( \Delta t \to 0 \) | Computational mode: \( \lambda \to -1 \), sign flips each step | |
| Numerical filtering (Robert–Asselin, Shapiro, \( \nabla^4 \)) | Meteorological filtering (hydrostatic, anelastic, QG equations) | |
| Stability (bounded solution) | Accuracy (correct phase and amplitude) | |
| Semi-implicit: removes the gravity-wave time-step limit | Semi-Lagrangian: removes the advective CFL limit | |
| Hydrostatic models: no vertically propagating sound | They still have the horizontally propagating Lamb (external acoustic) mode | |
| Upstream scheme stable for \( 0 \le \mu \le 1 \) and damping | Leapfrog stable for \( \lvert\mu\rvert \le 1 \) and neutral |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper.
Level 1: Foundation
- State the CFL condition for the leapfrog, centred-space scheme applied to 1-D linear advection.
- Which filter is applied after each leapfrog step to control the computational mode?
- What is the shortest wavelength that a grid of spacing \( \Delta x \) can represent?
Answers
- \( \lvert c\rvert\Delta t/\Delta x \le 1 \), with \( c \) the fastest signal speed.
- The Robert–Asselin time filter, \( \bar{u}^{n} = u^{n} + \nu(\bar{u}^{n-1} - 2u^{n} + u^{n+1}) \).
- \( 2\Delta x \) (wavenumber \( \pi/\Delta x \)).
Level 2: Part B Style
Practice Q1. An explicit primitive-equation model on a 1-D test grid with \( \Delta x = 50 \) km supports external gravity waves of speed 300 m/s. Using the leapfrog scheme, the largest stable time step is closest to:
(a) 83 s (b) 167 s (c) 333 s (d) 600 s
Answer: (b) 167 s
CFL: \( \Delta t \le \Delta x / c = 50\,000/300 = 166.7 \) s. Option (c) is the limit for \( \Delta x = 100 \) km; option (a) halves the correct value, as if a 2-D factor of one-half (not \( 1/\sqrt{2} \)) had been applied.
Practice Q2. The forward (Euler) scheme is applied to \( du/dt = i\omega u \) with \( \omega\Delta t = 0.1 \). After 100 steps the amplitude of the numerical solution, relative to its initial value, is about:
(a) 0.61 (b) 1.00 (c) 1.10 (d) 1.64
Answer: (d) about 1.64
\( \lvert\lambda\rvert = \sqrt{1 + p^2} = \sqrt{1.01} \approx 1.00499 \). After 100 steps: \( \lvert\lambda\rvert^{100} = 1.01^{50} \approx 1.645 \). The true amplitude stays 1 (option b). The growth looks slow per step, but it never stops: the forward scheme is unconditionally unstable for oscillations.
Practice Q3. With second-order centred space differences (time treated exactly), a wave of wavelength \( 4\Delta x \) is advected at a phase speed equal to what fraction of the true speed?
(a) 0 (b) 0.41 (c) 0.64 (d) 0.90
Answer: (c) about 0.64
\( k\Delta x = 2\pi\Delta x/(4\Delta x) = \pi/2 \), so \( c^{*}/c = \sin(\pi/2)/(\pi/2) = 2/\pi \approx 0.637 \). Option (a) is the \( 2\Delta x \) wave, (b) the \( 3\Delta x \) wave, (d) the \( 8\Delta x \) wave.
Practice Q4. On a grid with \( \Delta x = 100 \) km, a nonlinear product generates a wave of wavelength 150 km. The grid sees it as a wave of wavelength:
(a) 100 km (b) 200 km (c) 300 km (d) 450 km
Answer: (c) 300 km
The shortest resolvable wavelength is \( 2\Delta x = 200 \) km, so the 150 km wave aliases. \( k^{*} = 2\pi/\Delta x - k = 2\pi(1/100 - 1/150) = 2\pi/300 \) per km, so \( L^{*} = 300 \) km. Its energy is wrongly added to the 300 km wave.
Level 3: Part C Style
Practice Q5. The leapfrog scheme is applied to \( du/dt = i\omega u \) with \( p = \omega\Delta t = 0.5 \). Consider:
- Both roots of the amplification equation have modulus 1.
- The physical mode advances in phase by about 0.524 rad per step, so the numerical frequency is about 4.7% too high.
- The computational mode has \( \lambda \approx -0.87 + 0.5i \) and changes sign from step to step.
- If the Matsuno scheme were used instead with the same \( p \), the amplitude would fall by about 10% per step.
Which are correct?
(a) 1 and 2 only (b) 1, 2 and 3 only (c) 2, 3 and 4 only (d) 1, 2, 3 and 4
Answer: (d) 1, 2, 3 and 4
- True. \( \lambda = ip \pm\sqrt{1 - p^2} = \pm 0.866 + 0.5i \); \( \lvert\lambda\rvert^2 = 0.75 + 0.25 = 1 \).
- True. Phase \( \theta = \arcsin(0.5) = \pi/6 = 0.5236 \) rad against the true 0.5 rad; ratio 1.047.
- True. \( \lambda_2 = -0.866 + 0.5i \): its real part is negative and near \( -1 \), so the mode alternates sign (with a slow phase drift).
- True. Matsuno: \( \lvert\lambda\rvert = \sqrt{1 - p^2 + p^4} = \sqrt{1 - 0.25 + 0.0625} = \sqrt{0.8125} = 0.901 \), a loss of about 10% per step. This is the selective damping that makes Matsuno useful for dynamic initialization.
Practice Q6. A limited-area model with \( \Delta x = 25 \) km has external gravity waves of 300 m/s and a maximum jet speed of 100 m/s. In 1-D terms, which statement about time steps is correct?
(a) Explicit: about 250 s; semi-implicit: about 83 s (b) Explicit: about 83 s; semi-implicit: about 250 s; semi-implicit semi-Lagrangian: limited by accuracy rather than the CFL condition (c) Explicit and semi-implicit both about 83 s, because the gravity waves are still present (d) Explicit: about 83 s; semi-Lagrangian alone (gravity terms explicit): no stability limit at all
Answer: (b)
Explicit: the fastest wave sets the limit, \( \Delta t \le 25\,000/300 = 83.3 \) s. Semi-implicit: gravity-wave terms are implicit, so only advection limits it: \( \Delta t \le 25\,000/100 = 250 \) s, three times longer. Adding semi-Lagrangian advection removes the advective CFL limit too, leaving accuracy as the constraint.
(c) is wrong: the gravity waves remain in the model but no longer limit stability. (d) is wrong: semi-Lagrangian advection does nothing about explicitly treated gravity waves, which still impose the 83 s limit.
Practice Q7. A modeller runs a barotropic vorticity model with an ordinary centred finite-difference Jacobian. The run is stable for 20 days, then grid-scale noise appears and the model blows up. Halving the time step delays nothing. Which pair of statements explains the failure and a correct cure?
- The CFL condition was violated.
- Nonlinear interactions created waves shorter than \( 2\Delta x \), which aliased into the shortest resolved waves and accumulated there.
- Replacing the Jacobian with one that conserves energy and enstrophy would prevent the build-up.
- Using the forward time scheme instead of leapfrog would cure it.
(a) 1 and 4 (b) 2 and 3 (c) 1 and 3 (d) 2 and 4
Answer: (b) 2 and 3
Insensitivity to \( \Delta t \) and a slow build-up of grid-scale noise are the signature of nonlinear (aliasing) instability, not CFL violation (statement 1 would give rapid blow-up that a smaller step cures). An energy- and enstrophy-conserving (Arakawa) Jacobian prevents spurious accumulation of energy at the shortest scales. Statement 4 makes things worse: the forward scheme is itself unstable for advection.
Quick Revision Table
| Idea | Formula or fact |
|---|---|
| Lax equivalence | consistency + stability ⇔ convergence (linear, well posed) |
| von Neumann test | \( u_j^n = \lambda^n e^{ikj\Delta x} \); stable if \( \lvert\lambda\rvert \le 1 \) |
| Forward (oscillation) | \( \lvert\lambda\rvert = \sqrt{1+p^2} \): always unstable |
| Backward / trapezoidal | \( 1/\sqrt{1+p^2} \) damping / 1 neutral, both unconditional |
| Matsuno | \( \sqrt{1 - p^2 + p^4} \); stable \( \lvert p\rvert \le 1 \); damps high frequencies |
| Leapfrog | \( \lambda = ip \pm\sqrt{1-p^2} \); neutral \( \lvert p\rvert \le 1 \); computational mode \( \to -1 \) |
| Robert–Asselin | \( \bar{u}^{n} = u^{n} + \nu(\bar{u}^{n-1} - 2u^{n} + u^{n+1}) \) |
| CFL | \( c_{\max}\Delta t/\Delta x \le 1 \) (1-D); about \( 1/\sqrt2 \) on a 2-D square grid |
| FTCS advection | unconditionally unstable |
| Upstream | stable \( 0 \le \mu \le 1 \); diffusive |
| Phase speed, centred | \( c^{*}/c = \sin(k\Delta x)/(k\Delta x) \); \( 2\Delta x \) stationary |
| Aliasing | \( k^{*} = 2\pi/\Delta x - k \) for \( k > \pi/\Delta x \) |
| Nonlinear instability cures | Arakawa Jacobian, alias-free spectral grid, filters, \( \nabla^4 \) diffusion |
| Hydrostatic | filters vertical sound waves (Lamb mode stays) |
| Anelastic / Boussinesq | filters all sound waves |
| Quasi-geostrophic | filters sound and gravity waves; Rossby waves remain |
Key Takeaways
What’s Next
- Barotropic, Baroclinic and Primitive-Equation Models: the equations these numerical methods are applied to, from the 1950 barotropic forecast to modern PE models.
- Objective Analysis, Initialization, Data Assimilation and Satellite Data: the other half of Richardson’s failure, the imbalanced initial state.
- Numerical Weather Prediction: Quick Revision.
- The PDE classification behind all this: Classification of Second-Order PDEs.