What mathematical problem represents the system?
Equation, optimization, integral, simulation?
Numerical quality cannot exceed the relevance of the underlying model.
Side 99
Scientific Computing &
Numerical Methods
A study of solving scientific problems when exact analytic answers are unavailable or impractical. Numerical methods replace continuous or exact problems with controlled approximations, then track how discretization, rounding and algorithms shape the result.
model→discretize→approximate→compute→validate
06computing lenses
05error sources
05solver families
99Side
Scientific computing turns mathematical models into finite procedures.
Continuous quantities are sampled, infinite processes are truncated, and exact relationships are replaced by approximations that can be computed.
How is the continuous problem made finite?
Grid, timestep, basis?
Discretization defines what detail the computation can resolve.
What procedure computes the approximation?
Direct or iterative?
Different algorithms trade memory, speed, stability and accuracy.
How far can the result be from the target?
Bound and diagnose.
Numerical work is credible only when its error sources are understood.
Does refinement approach the intended solution?
Smaller step, better result?
Convergence links the finite approximation back to the mathematical problem.
Numerical error is not one thing.
Model error, discretization error, truncation and floating-point effects can each dominate under different conditions.
The equations omit reality.
Perfect computation of a bad model remains scientifically wrong.
Finite grids approximate continuous structure.
Coarse resolution can miss gradients, oscillations or boundaries.
An infinite process is cut short.
Series and iterative schemes stop after finite work.
Computers store finite precision.
Subtraction, accumulation and extreme scales can amplify floating-point limitations.
The problem itself may amplify input error.
An ill-conditioned problem can defeat even a stable algorithm.
The algorithm may amplify computational error.
Stable methods control perturbations introduced during computation.
Linear algebra is the workhorse beneath scientific computing.
Discretized models repeatedly generate systems of equations, eigenproblems and least-squares fits.
Direct solve
Factor the matrix.
LU, QR and related methods solve moderate systems predictably.
Iterative solve
Approach the solution progressively.
Large sparse systems often favor iterative methods.
Precondition
Reshape the problem for faster convergence.
A good preconditioner reduces difficult numerical geometry.
Least squares
Fit overdetermined systems.
Approximate solutions minimize residual error when exact consistency is impossible.
Eigenproblem
Find natural modes.
Stability, vibration and diffusion often reduce to dominant eigenstructure.
Nonlinear problems usually require iteration.
The algorithm proposes an approximation, measures the residual, and updates until a stopping criterion is met.
| Problem | Method family | Core idea |
|---|---|---|
| Root finding | Bisection / Newton | Narrow interval or follow local slope |
| Optimization | Gradient / Newton-type | Move toward lower objective value |
| Integration | Quadrature | Approximate area by weighted samples |
| ODE integration | Euler / Runge–Kutta | Advance state through time steps |
| PDE solution | Finite difference / volume / element | Replace continuous field with finite degrees of freedom |
Simulation creates a synthetic experiment on a model.
It is powerful because it reveals consequences of assumptions—but dangerous when output detail is mistaken for empirical truth.
Same inputs, same trajectory.
Useful for mechanism, sensitivity and parameter sweeps.
Random sampling propagates uncertainty.
Repeated draws approximate distributions or difficult integrals.
Local rules generate collective outcomes.
Useful when heterogeneity and interaction structure matter.
Map behavior across assumptions.
Exploring parameter space reveals robustness and regime transitions.
Which inputs control outputs?
Sensitivity analysis identifies parameters that deserve measurement or caution.
Do not hide unknown inputs behind precise outputs.
Simulation should propagate uncertainty rather than erase it.
A numerical result must be checked against both mathematics and reality.
Verification asks whether the equations were solved correctly; validation asks whether the right equations were solved for the intended purpose.
Sanity check
Confirm units, signs, scales and limiting cases.
Grid refinement
Change resolution and test whether the result converges.
Benchmark
Compare against a known analytic or trusted reference problem.
Cross-method
Use an independent algorithm where feasible.
Reality check
Compare model output with observations appropriate to the scientific claim.
Numerical Analysisapproximation, error and convergence
Numerical Linear Algebrascientific computing foundations
Scientific Computingalgorithms for mathematical models
Verification & Validationcomputation versus model adequacy