From a clever trick to a place to stand
On the previous rung you learned a liberating trick: you can differentiate a rough function by handing the derivative over to a smooth probe. To define u', you never differentiate u directly; you decree that integration by parts holds, sliding the d/dx off u and onto a test function phi that can always take it. That is the distributional derivative, and it let you differentiate kinks, jumps, even spikes. But a trick that can be done is not yet a place to live. This rung asks the next, structural question: if a function has a derivative only in this weak averaged sense, which functions are we actually allowed to put through the machine, and what room do they all sit in together?
Here is the discipline that distinguishes this rung from the last. Distributions are gloriously permissive — the Dirac delta and worse all count, and they have no honest size. But to solve a PDE we will want to measure things: how big is u, how big is its gradient, does a sequence of approximations actually converge to a limit? That needs not just a derivative but a norm — a length you can compute and compare. The Sobolev space is the answer: we keep only the rough functions whose weak derivatives are not merely defined but small enough to measure, and we equip that collection with a tape measure. We are trading the wild freedom of distributions for a smaller, sturdier world we can do analysis in.
The weak derivative, made into a measurement
Let us state the weak derivative cleanly, because the whole rung rests on it. We say a function v is the weak derivative u_x of u when integration by parts holds against every test function, with no boundary leftover: the integral of u times phi_x equals minus the integral of v times phi, for all smooth phi vanishing near the edge. The single minus sign is the whole content — it is exactly the boundary term you would discard if u were smooth, now promoted to the definition. If a classical derivative exists, this v agrees with it; if not, this v may still exist and be a perfectly ordinary function. The point is that v lives in the same honest world of integrable functions, not out in distribution-land.
A tiny example fixes it. Take the tent u(x) = 1 - |x| on the interval (-1, 1), which has a corner at the origin. Classically u' does not exist at x = 0, so old-style calculus stalls. But weakly, u_x is simply the step function that is +1 for x < 0 and -1 for x > 0 — an ordinary, square-integrable function. Crucially, no Dirac delta appears: because u is continuous (only its slope jumps, not its value), one weak derivative stays a function. The delta only arrives if you differentiate that step a second time. So the tent has one weak derivative inside our measurable world, but not two — and that distinction is precisely what the Sobolev spaces are built to record.
Building the rooms: L^2, H^1, and H^k
Now we furnish the spaces. Start with the ground floor, L^2: all functions whose square has a finite integral — finite energy, in physical language. Its size is the L^2 norm, the square root of the integral of u^2, and it is what lets us say two functions are close. The first Sobolev space H^1 is the floor above it: those u that are themselves in L^2 and whose first weak derivative is also in L^2. So membership in H^1 is a deal — you may be rough, but both you and your one weak slope must have finite energy. The tent u = 1 - |x| passes: u is bounded and its weak derivative is the bounded step, both square-integrable.
L^2 : integral of u^2 < infinity (finite energy)
H^1 : u in L^2 AND u_x in L^2
H^k : u and all weak derivatives up to order k are in L^2
norm: ||u||_{H^1}^2 = integral of ( u^2 + u_x^2 ) ( value AND slope counted )
||u||_{H^k}^2 = integral of ( u^2 + u_x^2 + ... + (k-th deriv)^2 )The general space H^k just iterates the deal: u and all its weak derivatives up to order k must sit in L^2. The Sobolev norm then bundles them — you square u, square each weak derivative, add, integrate, take the root. The deep payoff is that H^k is complete: any sequence whose terms get arbitrarily close in this norm actually converges to a limit that is still in H^k. That word, complete, is not decoration. It is the property that lets us hunt for a solution as a limit of approximations and be guaranteed the limit exists and is the kind of object we wanted. A space where Cauchy sequences land — an L^2-based one, hence a Hilbert space with an inner product — is the natural arena for the existence theorems this rung is marching toward.
The room with the boundary built in: H^1_0
One more room is special enough to name on its own. Many boundary-value problems pin the solution to zero on the edge — a drum clamped at its rim, a temperature held at the wall. We want a space that has "= 0 on the boundary" baked into its very membership, so we never have to impose it separately. That space is H^1_0. The cleanest way to build it: start from the smooth functions that are switched completely off near the boundary, then take the completion of that collection in the H^1 norm. Every function you obtain inherits, in a precise limiting sense, the property of vanishing at the edge.
Why go to this trouble instead of just writing "u = 0 on the boundary"? Because a typical H^1 function is too rough to have a value on the boundary in the naive sense — the edge is a measure-zero sliver, and changing u there changes no integral. Making sense of "the values on the boundary" of a rough H^1 function is a genuine theorem, the trace theorem, which the next guide proves. For now, H^1_0 sidesteps that subtlety: it is the honest home of functions that are interior to the domain, the natural place to look for the solution of a clamped, zero-on-the-edge problem. It will be the trial space in which we phrase the weak formulation of such a problem.
Why these rooms are exactly the right ones
Step back and see the strategy of the whole rung in one breath. A boundary-value problem like Laplacian u = f, classically, demands that u be twice differentiable everywhere — a steep entry fee that many honest physical solutions cannot pay. The Sobolev idea lowers the fee. We will rewrite the equation in its weak form by multiplying by a test function and integrating by parts once, which moves one derivative off u and onto the probe. The cost drops from "twice differentiable" to "one weak derivative in L^2" — that is, to merely belonging to H^1. A weak solution is any u in the right Sobolev space that satisfies this integrated identity, and the spaces of this guide are exactly the rooms where such a u is allowed to be sought.
Why does this particular trade pay off so handsomely? Because once the problem lives in a complete, inner-product space like H^1 or H^1_0, the heavy machinery of functional analysis becomes available. The integrated equation turns into a clean statement about a bilinear form paired with a linear functional, and a single abstract theorem — Lax-Milgram, in guide 4 — delivers existence and uniqueness of the weak solution in one stroke. None of that is reachable in the classical space of twice-differentiable functions, which is too cramped and not complete. The whole point of choosing these rooms is that they are where the existence theorems can actually run.
And here is the honest worry you should already be feeling, raised so you can see how the rung answers it. A weak solution that merely lives in H^1 might be a rough beast — is it the real solution, the smooth one physics wanted, or a cheap impostor we let in by lowering standards? That tension is real and it is the climax of this rung. The resolution is regularity: theorems showing that for nice data and nice domains, a weak solution is secretly far smoother than its membership card admits, often smooth enough to be a classical solution after all. So we descend to H^1 to find a solution easily, then climb back up to recover its smoothness. The map for the rung is now set: this guide built the rooms, guide 2 measures them (embeddings, traces, Poincare), guide 3 writes the weak formulation, guide 4 proves existence, and guide 5 earns the classical solution back.