Differential equations and well-posed evolution

An ordinary differential equation follows a state as one variable, usually time, changes. For a first-order equation, initial data select a trajectory from a family of solutions. A standard local existence-and-uniqueness theorem assumes the vector field is continuous in time and locally Lipschitz in the state. Local means that the theorem provides a solution for a sufficiently short time near the starting point. It does not exclude later blow-up or guarantee boundedness forever.

Uniqueness can fail when the Lipschitz condition fails. Starting at zero, the equation y'=2√|y| admits the zero solution and solutions that wait until an arbitrary nonnegative time before following a squared ramp. Conversely, uniqueness does not imply numerical ease. The equation y'=−1000y has a unique smooth decaying solution, yet an explicit time integrator can produce exploding oscillations if its step is too large. Logical solvability and practical computability are different questions.

A partial differential equation relates changes across several coordinates. The heat equation diffuses information and damps high-frequency modes. The wave equation propagates disturbances with a finite characteristic speed in the ideal continuum model. The Poisson equation instead constrains a spatial profile from sources and boundary data. These examples motivate the broad parabolic, hyperbolic, and elliptic classifications, though nonlinear systems and variable coefficients can require a more careful local analysis than a single label suggests.

Initial and boundary conditions must match the problem. A vibrating string needs initial displacement and velocity, plus endpoint behavior. A diffusion problem generally needs one initial profile and compatible boundary information. On a periodic interval, Fourier modes make this visible: the wave amplitude obeys a second-order oscillator equation, while the heat amplitude obeys a first-order decay equation. The number and kind of data follow from the evolution law rather than an arbitrary preference of the solver.

Forward Euler approximates a time derivative by a forward difference. Substituting a smooth exact solution shows the local truncation error shrinks as the step shrinks, but this consistency statement does not control accumulated error. For the decaying test equation y'=−λy, Euler multiplies the state by 1−λh at every step. The method damps errors only when the multiplier has magnitude below one. A stable physical law can therefore have an unstable discrete approximation.

For PDEs, refining a spatial mesh often introduces faster modes and tighter explicit time-step restrictions. Implicit methods can improve stability but require solving algebraic systems and do not erase accuracy requirements. Some problems also develop shocks or lack classical derivatives, so the proper target is a weak solution with admissibility conditions. Always ask which equation, which solution concept, which norm of error, and which discretization are being validated. Agreement between two plots at one resolution answers only a small part of that checklist.