Boundary and Initial Conditions

An ordinary differential equation like y' = 2y has a whole family of solutions, y = Ce^{2x} — one for every constant C. Give it a single extra fact, an initial value such as y(0) = 5, and the family collapses to exactly one curve.

A PDE has exactly the same problem, only bigger. The equation alone is solved by a flat bar, a cooling bar, a bar heated at one end, and endlessly more — because a PDE's general solution doesn't just contain arbitrary numbers, it contains whole arbitrary functions. Pinning one solution down takes correspondingly more information: not a single number, but data specified across an entire boundary and, for time-dependent problems, across an entire initial time-slice.

Two kinds of "extra information"

An initial condition gives the state of the system at the start, t = 0, everywhere in space at once: u(x, 0) = f(x) for every x in the domain. Think of it as a snapshot — a photograph of the whole rod's temperature the instant the clock starts.
A boundary condition gives the behaviour at the edges of the spatial domain, for all time: for a rod occupying [0, L], that means specifying something about u at x = 0 and x = L for every t, not just once.

Both are needed together. The initial condition says nothing about what happens at the ends after t = 0; the boundary conditions say nothing about how the temperature starts out in the middle. Neither one alone comes close to determining the whole future of the rod.

Worked example: a rod with both pieces of data

Take the heat equation u_t = \alpha u_{xx} on a rod occupying [0, L]. Suppose we're told:

Initial condition: at the moment we start the clock, the rod's temperature profile is u(x, 0) = f(x) — perhaps warmer in the middle, cooler at the ends, described by some given function f.
Boundary conditions: both ends are clamped at freezing, u(0, t) = 0 and u(L, t) = 0, for every t > 0 — say each end sits in an ice bath the whole time.

With only the initial profile f(x), we'd have no idea whether the ends are insulated, clamped, or heated — the rod could evolve in infinitely many different ways. With only the boundary values, we'd know the ends stay at zero forever but have no idea what's happening in the middle at the start. It's the combination of both pieces of data — the snapshot at t=0 plus the rule at the edges for all time — that determines a single, unique temperature u(x, t) for every point on the rod, at every future moment. Working out exactly what that solution looks like, using a technique called separation of variables, is the heart of solving the heat equation on an interval.

The flavours of boundary condition

On the spatial boundary, three kinds cover almost everything. Picture again a rod occupying [0, L]:

Dirichlet — fix the value: u(0, t) = T_0. The end is held at a known temperature, exactly as if it were clamped into an ice bath or a heating block that keeps it pinned at T_0 no matter what the rest of the rod is doing.
Neumann — fix the derivative, i.e. the flux: u_x(0, t) = q. This says how fast heat is flowing through that end, not what temperature it's at. The special case q = 0 means the end is perfectly insulated — wrapped in something that lets no heat escape, so whatever temperature builds up there is whatever it ends up being.
Robin — fix a combination of the two: a\,u(0, t) + b\,u_x(0, t) = g. This models convective cooling: an end exposed to open air loses heat at a rate proportional to how much hotter it is than the surrounding room, rather than being clamped to a fixed temperature or perfectly sealed.

Dirichlet and Neumann are the two extremes — completely fixed value, or completely fixed flow — and Robin sits between them, blending both.

How many conditions, and of what kind

The type of the equation decides what data it needs:

Parabolic (heat): boundary conditions on the spatial edges plus one initial condition u(x, 0) = f(x) — the equation is first order in time.
Hyperbolic (wave): boundary conditions plus two initial conditions, u(x, 0) and u_t(x, 0) — second order in time needs both the initial shape and the initial velocity.
Elliptic (Laplace): no time at all — just boundary values around the entire edge of the region. It is a pure boundary-value problem.

Dirichlet fixes u; Neumann fixes u_n (the normal derivative); Robin fixes a combination.
Parabolic: BCs + one IC. Hyperbolic: BCs + two ICs (position and velocity). Elliptic: BCs only, around the whole boundary.

Supplying data is not automatically safe — the wrong amount or the wrong kind can wreck the problem in two different ways:

Too little data leaves the problem underdetermined: many different solutions all satisfy what little you've specified, and there's no way to know which one is "the" answer. Give a parabolic rod problem an initial profile but no boundary conditions at all, and the ends are free to do almost anything.
Too much, or conflicting, data leaves the problem overdetermined: the conditions fight each other and no function can satisfy all of them simultaneously. Demand that a rod's end simultaneously sit at a fixed temperature and have a separately fixed, incompatible flux, and generally nothing fits.

Exactly how much data — and which flavour — turns a PDE problem into one with a single, sensible solution depends directly on the equation's classification: an elliptic equation wants boundary data only, a parabolic one wants boundary data plus one initial slice, a hyperbolic one wants boundary data plus two. Getting this balance exactly right, so a problem has one solution that also depends continuously on the data, is the subject of well-posedness.

Dirichlet and Neumann conditions aren't just two equally abstract options on a menu — they model two genuinely different, very ordinary physical setups that any kitchen could produce.

Dunk the end of a metal rod into a large ice bath, and no matter how much heat arrives from the rod, the huge bath barely warms up: the end is effectively pinned at 0°, forever. That's a Dirichlet condition — the value is fixed.

Now wrap that same end in thick insulating foam instead. Heat can't get in or out through it at all, so the flux through that end is fixed at 0 — but the temperature right at the tip is free to drift to whatever the rest of the rod pushes it to. That's a Neumann condition with zero flux. Same rod, same equation, two completely different pieces of kitchen equipment on the end — and two completely different kinds of boundary condition.