# Computational mathematics Computational mathematics is the branch of [[Applied_mathematics|applied mathematics]] that treats computation itself as a mathematical object: it designs and proves things about the [[Algorithm|algorithms]] by which continuous and discrete problems get solved on finite machines. It sits at the junction of [[Mathematics]], [[Computer_science|computer science]], and the sciences that consume its solvers, and its theorems are about convergence rates, error bounds, and [[Algorithmic_efficiency|algorithmic efficiency]] rather than existence alone. The field matters because the distance between "solvable in principle" and "solved by Thursday" is where modern science lives: a [[Partial_differential_equation|partial differential equation]] with no closed form becomes usable the moment a stable discretization, a fast [[Linear_algebra|linear-algebra]] kernel, and an honest error estimate exist — which is why every serious [[Simulation|simulation]], [[Mathematical_model|mathematical model]], and [[Weather_forecasting|weather forecast]] runs on this discipline's output. ## Finite precision: the arithmetic underneath everything Real numbers do not fit in hardware, so computational mathematics begins with rounded arithmetic. The IEEE 754 standard (1985) fixes the dominant format: binary64 carries a 53-bit significand in [[Binary_number|binary]], giving unit roundoff u = 2⁻⁵³ ≈ 1.1×10⁻¹⁶. Every operation may inject a relative error of that size, and the field's first duty is tracking how such errors amplify. Two ideas carry the load. *Conditioning* is a property of the problem: for a linear system the condition number κ(A) = ‖A‖·‖A⁻¹‖ bounds how much input perturbations grow, independent of method. *Stability* is a property of the algorithm: James Wilkinson's backward-error analysis (1960s) showed that a good solver returns the exact answer to a nearby problem. Confusing the two is the classic failure mode; [[Accuracy_and_precision|accuracy and precision]] are different budgets, and [[Correctness_(computer_science)|correctness]] claims for numerical code mean nothing without both. Catastrophic cancellation — subtracting nearly equal quantities — can erase all significant digits in one step, a fact no amount of hardware [[Robustness_(computer_science)|robustness]] repairs. ## The kernels: linear algebra, transforms, quadrature A short list of kernels does most of the world's numerical work. Dense Gaussian elimination factors an n×n system in about ⅔·n³ floating-point operations; Strassen showed in 1969 that O(n^2.81) is possible, and sparse solvers exploit [[Graph_theory|graph-theoretic]] orderings of the nonzero pattern (an [[Adjacency_matrix|adjacency-matrix]] view of the matrix) to skip the zeros entirely. The [[Fast_Fourier_transform|fast Fourier transform]] (Cooley–Tukey, 1965; anticipated by Gauss ca. 1805) computes a length-N discrete transform in O(N log N), which is why [[Convolution|convolution]], [[Signal_processing|signal processing]], and spectral solvers are cheap; the [[Discrete_cosine_transform|discrete cosine transform]] variant sits inside [[Image_compression|image compression]]. For integrals, [[Numerical_integration|numerical integration]] by n-point Gaussian quadrature is exact for polynomials through degree 2n−1 — a precision theorem, not a heuristic. These kernels are packaged once and reused everywhere; the BLAS/LAPACK lineage catalogued under [[List_of_numerical_libraries|numerical libraries]] and [[Mathematical_software|mathematical software]] is the field's shared infrastructure. ## Differential equations and the simulation stack Most simulation is the numerical solution of differential equations. For an [[Ordinary_differential_equation|ordinary differential equation]], the Runge–Kutta family (Runge 1895, Kutta 1901) trades function evaluations for order: classical RK4 has local error O(h⁵). Stiff systems — fast decaying modes coupled to slow dynamics, ubiquitous in [[Chemical_process_modeling|chemical process modeling]] — force implicit methods, a discovery about problems, not codes. For a [[Partial_differential_equation|PDE]], the Courant–Friedrichs–Lewy condition (1928) caps the stable time step of explicit schemes at roughly Δt ≤ Δx/c: the numerical scheme must not outrun the physics. On this base sits the stack the vault knows well: [[Fluid_dynamics|fluid dynamics]] codes, [[Molecular_dynamics|molecular dynamics]], [[Process_simulation|process simulation]], and every [[Atmospheric_model|atmospheric model]]. The first computer forecast ran on ENIAC in 1950 under [[John_von_Neumann|John von Neumann]]'s program, redeeming Richardson's 1922 hand-computed attempt; a decade later [[Edward_Norton_Lorenz|Edward Lorenz]] found, inside such a model, the sensitive dependence that became [[Chaos_theory|chaos theory]] — a numerical experiment producing fundamental mathematics, and a permanent caveat on [[Predictability|predictability]]. ## Randomness as a numerical method When dimension defeats grids, computational mathematics reaches for chance. The [[Monte_Carlo_method|Monte Carlo method]] (Metropolis and Ulam, 1949, with von Neumann's random-number machinery) estimates an [[Expected_value|expected value]] by averaging N samples; the error shrinks as σ/√N *regardless of dimension*, which is the entire point — a 100-dimensional integral costs no more per digit than a 3-dimensional one. That bound comes straight from [[Probability_theory|probability theory]], and its practical use leans on [[Statistics|statistics]] for variance reduction. The same logic now drives stochastic gradient descent in [[Machine_learning|machine learning]] and sampling in [[Bayesian_network|Bayesian networks]], with [[Graphics_processing_unit|graphics processing units]] supplying the parallel arithmetic. Randomized algorithms are not approximations of ideal ones; they are algorithms with theorems of their own. ## Optimization, symbols, and the discrete side The field is broader than floating point. [[Mathematical_optimization|Mathematical optimization]] became computational the day [[George_Dantzig|George Dantzig]] published the simplex method (1947); [[Richard_Bellman|Richard Bellman]]'s [[Dynamic_programming|dynamic programming]] (1950s) did the same for sequential decisions, and [[Combinatorial_optimization|combinatorial optimization]] carries the discrete load for [[Operations_research|operations research]]. [[Computer_algebra|Computer algebra]] manipulates symbols exactly — Buchberger's Gröbner-basis algorithm (1965) decides membership in polynomial ideals — while [[Computational_geometry|computational geometry]] handles meshes and arrangements, and [[Cryptography|cryptography]] consumes fast arithmetic in finite fields ([[Field_(mathematics)|fields]] of prime order). The costs here are governed by [[Theory_of_computation|theory of computation]]: knowing a problem is NP-hard redirects effort from exact solvers toward relaxations and heuristics. ## What the vault buys from it Downstream, the dependencies are concrete. [[Control_theory|Control theory]] ships Riccati and [[Kalman_filter|Kalman-filter]] solvers; [[System_identification|system identification]] and [[Time_series|time-series]] estimation are numerical linear algebra in disguise; [[Reliability_engineering|reliability engineering]] runs Monte Carlo over failure models; [[Quantum_computing|quantum computing]] promises new asymptotics for old kernels. Tool inventories — [[Comparison_of_optimization_software|optimization software]], [[List_of_computational_physics_software|computational physics codes]] — are the visible surface of the guarantee culture underneath: every trustworthy number in a [[Complex_system|complex-system]] study was produced by an algorithm someone proved something about. **On the spine:** [[Applied_mathematics]] · [[Algorithm]] · [[Numerical_integration]] · [[Linear_algebra]] · [[Monte_Carlo_method]]. ## Wikipedia : Wikitube **Strict pair:** [Wikipedia](https://en.wikipedia.org/wiki/Computational_mathematics) : [Wikitube](https://en.wikitube.io/wiki/Computational_mathematics) ## Previous hub tags Hubs: `Systems`. Portals: [[PORTAL_Graph_theory]], [[PORTAL_Decision_theory]], [[PORTAL_Information_theory]], [[PORTAL_Control_theory]], [[PORTAL_Operations_research]]. --- *Repopulated 2026-08-12 · redlink fill · 0 deletions.*