The Lax-Milgram Theorem: Existence From a Bounded, Coercive Form

The precise statement

Let H be a real Hilbert space with inner product ⟨·,·⟩ and norm ‖·‖. Let a: H × H → ℝ be bilinear (linear in each argument), and suppose there exist constants M, α > 0 such that:

• Boundedness (continuity): |a(u,v)| ≤ M‖u‖‖v‖ for all u, v ∈ H.
• Coercivity (ellipticity): a(u,u) ≥ α‖u‖² for all u ∈ H.

Then for every bounded linear functional f ∈ H* there exists a unique u ∈ H such that

a(u,v) = f(v) for all v ∈ H,

and moreover ‖u‖ ≤ (1/α)‖f‖_{H*}. Crucially, a is not assumed symmetric: a(u,v) ≠ a(v,u) is allowed. When a is symmetric it is an inner product equivalent to ⟨·,·⟩, and the statement collapses to the Riesz representation theorem. The complex version replaces coercivity with Re a(u,u) ≥ α‖u‖² and takes a sesquilinear.

The picture: an equivalent inner product tilted by a bounded operator

Think of a(u,v) as a distorted inner product. Boundedness says it never blows up faster than ‖u‖‖v‖; coercivity says a(u,u) stays comfortably above zero, so the form is uniformly "positive-definite" in the diagonal direction — no direction gets squeezed flat. Together these say a is an equivalent geometry on H, tilted and sheared but not degenerate.

By boundedness, a(u,·) is a continuous functional, so Riesz gives a bounded linear operator A: H → H with a(u,v) = ⟨Au,v⟩. Coercivity becomes ⟨Au,u⟩ ≥ α‖u‖². The equation a(u,v)=f(v) turns into Au = φ, where φ ∈ H is the Riesz vector of f. The whole theorem is then the single claim: A is invertible with a bounded inverse. Coercivity forces A to be bounded below (‖Au‖ ≥ α‖u‖), which is exactly what prevents A from crushing any direction to zero — the geometric reason a solution must exist and be unique.

Key idea of the proof: Riesz plus a contraction

Rewrite everything through Riesz. For each fixed u, v ↦ a(u,v) is bounded, so a(u,v) = ⟨Au,v⟩ for a unique Au; linearity in u makes A: H → H linear, and |a(u,v)|≤M‖u‖‖v‖ gives ‖A‖ ≤ M. Likewise f(v)=⟨φ,v⟩. The problem is Au = φ.

Now the trick: for a parameter ρ > 0 define the map T_ρ(u) = u − ρ(Au − φ). A fixed point of T_ρ solves Au = φ. Estimate ‖T_ρ u − T_ρ w‖². Writing z = u − w,

‖z − ρAz‖² = ‖z‖² − 2ρ⟨Az,z⟩ + ρ²‖Az‖² ≤ (1 − 2ρα + ρ²M²)‖z‖²,

using coercivity ⟨Az,z⟩ ≥ α‖z‖² and boundedness ‖Az‖ ≤ M‖z‖. Choosing 0 < ρ < 2α/M² makes the factor 1 − 2ρα + ρ²M² < 1, so T_ρ is a contraction. Completeness of H then lets the Banach fixed-point theorem deliver a unique u — existence and uniqueness at once. The bound ‖u‖ ≤ ‖f‖/α drops out of α‖u‖² ≤ a(u,u) = f(u) ≤ ‖f‖‖u‖.

Worked example: the weak form of a diffusion-advection problem

On a bounded domain Ω ⊂ ℝⁿ, solve −Δu + b·∇u + cu = f with u = 0 on ∂Ω, taking H = H₀¹(Ω), the Sobolev space with the norm ‖u‖ = ‖∇u‖_{L²}. Multiply by a test function v ∈ H₀¹, integrate, and use the divergence theorem to get the weak form a(u,v)=f(v) with

a(u,v) = ∫_Ω (∇u·∇v + (b·∇u)v + c u v) dx, f(v)=∫_Ω f v dx.

Boundedness follows from Cauchy-Schwarz and the Poincaré inequality (bounded b, c). Coercivity: a(u,u) = ‖∇u‖² + ∫(b·∇u)u + ∫c u². If b is divergence-free and c ≥ 0, the middle term vanishes (∫(b·∇u)u = −½∫(∇·b)u² = 0), giving a(u,u) ≥ ‖∇u‖² = ‖u‖², so α = 1. Note a is not symmetric because of the advection term b·∇u — yet Lax-Milgram still gives a unique weak solution with ‖u‖_{H₀¹} ≤ ‖f‖. Symmetric methods like energy minimization would not apply here.

Why the hypotheses matter — and what breaks

Drop coercivity. On H = ℝ², a(u,v) = u₁v₂ − u₂v₁ is bounded and nondegenerate but a(u,u) = 0 for all u — not coercive. The operator A is a rotation by 90°: invertible here, but the moment you weaken to a singular form (e.g. a(u,v)=u₁v₁ on ℝ²) you lose surjectivity and the bound ‖u‖ ≤ ‖f‖/α becomes meaningless (α = 0). Coercivity is exactly what makes A bounded below. In PDE terms, dropping it lets the operator have kernel — think of −Δ with pure Neumann conditions, where constants are in the kernel and solvability needs a compatibility condition (this is where the Fredholm alternative and Babuška's inf-sup take over).

Drop completeness. The contraction produces a Cauchy sequence; without completeness its limit need not lie in H, so the fixed point can fail to exist. Drop boundedness and A may be unbounded, the contraction estimate collapses, and even defining Au via Riesz fails. Babuška's generalization trades coercivity for an inf-sup condition inf_u sup_v a(u,v)/(‖u‖‖v‖) ≥ β > 0, allowing distinct trial and test spaces.

Applications and significance

Lax-Milgram is the existence theorem underneath modern elliptic PDE theory and numerical analysis. It certifies well-posedness of the weak (variational) formulation of second-order elliptic boundary-value problems — Dirichlet, Robin, and advection-diffusion — including nonsymmetric operators that no energy-minimization argument can reach. It is the theoretical foundation of the finite element method: restricting a to a finite-dimensional subspace V_h ⊂ H inherits boundedness and coercivity, so Lax-Milgram guarantees a unique discrete solution u_h, and Céa's lemma then bounds the error by ‖u − u_h‖ ≤ (M/α) inf_{v∈V_h}‖u − v‖ — the quasi-optimality estimate driving convergence rates. It also underlies existence for the stationary Stokes and linear elasticity systems, and the elliptic step in parabolic problems solved by Galerkin methods. In short, whenever a problem can be cast as "find u so that a(u,v)=f(v)," Lax-Milgram converts three inequalities into existence, uniqueness, and stability in one stroke.

Which representation theorem applies, by structure of the form a(·,·)

Property of a(·,·)	Symmetric + coercive	Nonsymmetric + coercive	Only inf-sup (Babuška)
Applicable theorem	Riesz / Lax-Milgram	Lax-Milgram	Babuška-Lax-Milgram
Boundedness needed	Yes (|a| ≤ M‖u‖‖v‖)	Yes	Yes
Coercivity a(u,u) ≥ α‖u‖²	Yes	Yes	Replaced by inf-sup β
Variational characterization	u minimizes ½a(v,v)−f(v)	No energy minimum	No energy minimum
Existence & uniqueness	Yes; ‖u‖ ≤ ‖f‖/α	Yes; ‖u‖ ≤ ‖f‖/α	Yes; ‖u‖ ≤ ‖f‖/β
Test/trial space	Same space H	Same space H	May differ (U, V)