Functional Analysis
Sobolev Spaces
Цели урока
- Define weak derivatives via integration by parts and understand W^{k,p} Sobolev spaces
- Apply Sobolev embedding theorems to determine regularity: subcritical (L^q) and supercritical (Holder)
- Use the Lax-Milgram theorem to establish existence and uniqueness of weak solutions
- Understand the trace theorem and fractional Sobolev spaces H^s
Предварительные знания
- L^p spaces and Holder inequality
- Banach spaces and dual spaces
- Basic distribution theory
Fluid simulation software like Ansys solves PDEs on meshes with discontinuous coefficients. Classical C^2 solutions do not exist. What is the correct mathematical setting - and how does it guarantee the numerical method converges?
- Ansys Fluent: FEM for Navier-Stokes on 10^8-cell meshes uses H^1 weak formulation; Lax-Milgram guarantees the discrete system has a unique solution
- Physics-Informed Neural Networks (PINNs): loss function is a discretized H^1 residual; Sobolev embedding guarantees uniform convergence
- Kernel methods in ML: Matern RKHS = H^{nu + d/2}; regularization = Sobolev norm penalty; Sobolev embedding explains continuity of sample paths
- Quantum mechanics: wave functions in H^1(R^3) guarantee finite kinetic energy; trace theorem defines boundary scattering amplitudes
From Dirichlet Principle to Modern FEM
Riemann used the Dirichlet principle (minimize integral |nabla u|^2) to prove existence of harmonic functions, but Weierstrass showed the minimizer need not exist in C^2. The resolution required a larger function space. Sobolev defined the spaces W^{k,p} in 1938 and proved embedding theorems. Lax and Milgram published their theorem in 1954. Cea (1964) and Strang-Fix (1973) built the error analysis for finite elements. Today FEM software processes 10^9 degrees of freedom using these exact mathematical foundations.
Sobolev Spaces W^{k,p} and Weak Derivatives
Ansys Fluent uses Sobolev spaces H^1 for finite element computations: simulating turbulence behind a Boeing 737 airfoil at Reynolds number 10^7 requires weak solutions of Navier-Stokes in H^1(Omega). Classical C^2 regularity is unavailable for rough boundaries and discontinuous coefficients - Sobolev spaces are the correct setting.
Absolute value function in Sobolev spaces
Example of a non-smooth function with a weak derivative
The function u(x) = |x| on (-1,1) is Lipschitz but not C^1. Its weak derivative is sign(x): the piecewise constant function +1 for x > 0 and -1 for x < 0. This is in L^2(-1,1), so u is in W^{1,2}(-1,1) = H^1(-1,1). The second weak derivative does not exist in L^2: delta(0) is a distribution but not a function.
The Poincare inequality ||u||_{L^2} <= C*|u|_{H^1} on H^1_0(Omega) is the key coercivity estimate for elliptic PDE. It says: on functions with zero boundary conditions, the L^2 norm is controlled by the gradient. This is what makes the Lax-Milgram theorem applicable to the weak Poisson problem.
How does a weak derivative differ from a classical derivative?
Weak derivative: v = D^alpha u means integral u * D^alpha phi = (-1)^|alpha| * integral v * phi for all test phi. No pointwise differentiability assumed.
Sobolev Embedding and Trace Theorems
The central question: when does membership in W^{k,p}(Omega) force better regularity - continuity, Holder continuity, or membership in a larger L^q? Sobolev embedding theorems answer this precisely, and they are the mathematical foundation for proving convergence of finite element methods.
| Space | Dimension n | Condition | Embedding |
|---|---|---|---|
| H^1 = W^{1,2} | n=1 | 2*1 > 1 | H^1 subset C^{0,1/2}: Holder-1/2 |
| H^1 = W^{1,2} | n=2 | 2*1 = 2 (critical) | H^1 subset L^q for all q < inf |
| H^1 = W^{1,2} | n=3 | 2*1 < 3 | H^1 subset L^6 (q=2n/(n-2)=6) |
| H^2 = W^{2,2} | n=3 | 2*2 = 4 > 3 | H^2 subset C^{0,1/2}: Holder-1/2 |
The critical exponent p* = np/(n-kp) is the sharp Sobolev exponent. For n=3, p=2, k=1: p* = 6. Functions in H^1(R^3) automatically lie in L^6 - this is used in nonlinear Schrodinger equations (cubic nonlinearity is critical in 3D) and in the analysis of the Navier-Stokes equations.
When does W^{1,p}(Omega) in R^n embed in a space of continuous functions?
Morrey's theorem: W^{k,p}(Omega) embeds in C^{0,alpha} when kp > n, with alpha = k - n/p. For k=1, p > n: alpha = 1 - n/p. For n=1, p=2: alpha = 1/2 - Holder-1/2.
Lax-Milgram Theorem and Finite Element Method
The Lax-Milgram theorem is the abstract existence and uniqueness theorem for weak solutions of elliptic PDE. Combined with Sobolev embedding, it provides the complete mathematical foundation for finite element methods used in Ansys, Abaqus, COMSOL - software processing meshes with up to 10^8 elements.
What is coercivity of a bilinear form a(u,v) on a Hilbert space V?
Coercivity: a(u,u) >= alpha*||u||^2 for some alpha > 0. For -Delta with Dirichlet BC: a(u,u) = ||nabla u||^2_{L^2} >= (1/C_P) * ||u||^2_{H^1} by Poincare inequality.
Interpolation Inequalities and Regularity Theory
Between L^2 and H^k lie intermediate Sobolev spaces H^s for non-integer s, defined by interpolation. These fractional Sobolev spaces appear in the trace theorem (H^{1/2} on the boundary), in wave equations (H^{1/2} initial data), and in machine learning theory (RKHS are Sobolev spaces with fractional indices).
Kernel methods and RKHS as Sobolev spaces
Machine learning connection
Reproducing kernel Hilbert spaces used in Gaussian processes and SVMs are Sobolev spaces. The Matern kernel k(x,y) with parameter nu corresponds to the RKHS being H^{nu+d/2}(R^d). This connects regularization in ML (controlling function complexity) directly to Sobolev regularity. The embedding theorem H^s subset C^0 when s > d/2 explains why Gaussian processes with Matern-1/2 give continuous but non-differentiable sample paths.
What is the trace of a function in H^1(Omega) on the boundary dOmega?
Trace theorem: gamma: H^1(Omega) -> H^{1/2}(dOmega) is bounded and surjective. H^{1/2} is fractional Sobolev: between L^2 and H^1. The trace loses exactly 1/2 derivative.
Connections to other topics
Sobolev spaces are the common language of modern PDE analysis and numerical methods.
- Distributions and Weak Solutions — Sobolev spaces are completions inside the space of distributions under the Sobolev norm
- Distribution Theory — Weak derivatives in Sobolev spaces are distributional derivatives by definition
- Functional Analysis in PDEs — Sobolev embeddings give the regularity needed to solve elliptic and parabolic equations
Итоги
- Weak derivative: v = D^alpha u means integral identity with all test functions; requires only integrability
- W^{k,p}(Omega): functions with weak derivatives in L^p up to order k; Banach space with Sobolev norm
- Sobolev embedding: kp < n implies W^{k,p} subset L^{np/(n-kp)}; kp > n implies W^{k,p} subset C^{0,k-n/p}
- Lax-Milgram: coercive bounded bilinear form on Hilbert space => unique weak solution
- Trace: gamma: H^1 -> H^{1/2}(boundary); fractional spaces H^s via Fourier or interpolation
Вопросы для размышления
- Why does the Sobolev exponent p* = np/(n-kp) appear naturally in the embedding theorem, and what happens at the critical case kp = n?
- The Poincare inequality holds on H^1_0 but not on H^1. Why does the zero boundary condition matter for controlling the L^2 norm?
- In what sense are Sobolev spaces 'the right setting' for elliptic PDE - and why would demanding C^2 solutions be too restrictive?
Связанные уроки
- fa-22 — Distributions generalize weak derivatives in Sobolev spaces
- fa-20-fixed-point — Banach spaces are the ambient setting for Sobolev spaces
- fa-23 — Semigroups act on Sobolev spaces for PDE evolution