The Resolvent Set and Resolvent Operator: Where (T−λI)⁻¹ Lives

Precise statement: the definition and its three guarantees

Let X be a complex Banach space and T ∈ B(X) a bounded linear operator (the theory extends verbatim to closed densely-defined T with X = domain considerations). For λ ∈ ℂ write T_λ := T − λI. Define:

• The resolvent set ρ(T) = {λ ∈ ℂ : T_λ : X → X is a bijection}.
• The resolvent operator R(λ,T) = (T − λI)⁻¹ for λ ∈ ρ(T).
• The spectrum σ(T) = ℂ \ ρ(T).

By the Bounded Inverse Theorem (a consequence of the Open Mapping Theorem, which needs X complete), whenever T_λ is a bijection its inverse is automatically bounded — so "bijective" already delivers "bounded inverse" for free. The three headline claims are: (i) ρ(T) is open; (ii) λ ↦ R(λ,T) is analytic (norm-holomorphic) on ρ(T); and (iii) σ(T) is a nonempty compact subset of {λ : |λ| ≤ ‖T‖}. The bound ‖R(λ,T)‖ ≤ 1/(|λ| − ‖T‖) holds for |λ| > ‖T‖.

The picture: a movie of inverses parametrized by ℂ

Think of R(·,T) as a single object: a function that eats a complex number λ and spits out an operator. As λ roams the plane, the operator R(λ,T) deforms smoothly — until λ hits the spectrum, where it blows up (‖R(λ,T)‖ → ∞ as λ → σ(T)). The spectrum is the "forbidden set" of poles and worse.

The cleanest mental model is the 1×1 case: X = ℂ, T = multiplication by a scalar a. Then T − λI is multiplication by (a − λ), invertible exactly when λ ≠ a, and R(λ,T) = 1/(a − λ). So σ(T) = {a} and R(λ,T) is literally the meromorphic function 1/(a − λ) with a simple pole at the eigenvalue. For an n×n matrix, R(λ,T) = adj(T − λI)/det(T − λI): a rational matrix-valued function whose poles are exactly the eigenvalues. Infinite-dimensionally the poles can thicken into curves, discs, or fractals — but the analytic-function-of-λ picture survives intact.

Key idea of the proof: Neumann series + the resolvent identity

Openness and analyticity. Fix λ₀ ∈ ρ(T), so R₀ := R(λ₀,T) exists and is bounded. For nearby λ write T − λI = (T − λ₀I) − (λ − λ₀)I = (T − λ₀I)[I − (λ − λ₀)R₀]. The bracket is invertible via the Neumann series ∑ₙ₌₀^∞ (λ − λ₀)ⁿ R₀ⁿ, which converges in operator norm whenever |λ − λ₀| · ‖R₀‖ < 1 (geometric domination, completeness of B(X) gives the limit). Hence

R(λ,T) = ∑ₙ₌₀^∞ (λ − λ₀)ⁿ R₀ⁿ⁺¹,

a convergent power series in λ centered at λ₀ — so ρ(T) is open and R(·,T) is analytic, with derivative dR/dλ = R(λ,T)². Nonemptiness. If σ(T) = ∅ then R(·,T) is entire; for |λ| > ‖T‖ the series ∑ Tⁿ/λⁿ⁺¹ shows ‖R(λ,T)‖ → 0 as |λ| → ∞, so R(·,T) is a bounded entire operator-valued function, forcing R ≡ 0 by Liouville's theorem — impossible since R(λ,T) is invertible. The two-point resolvent identity R(λ) − R(μ) = (λ − μ)R(λ)R(μ) glues everything together.

Worked example: the shift operators on ℓ²

Take X = ℓ²(ℕ) and the right shift S(x₁, x₂, …) = (0, x₁, x₂, …). It is an isometry, ‖S‖ = 1, so σ(S) ⊂ {|λ| ≤ 1}. S has no eigenvalues: Sx = λx forces x = 0. Yet for |λ| < 1 the operator S − λI is injective with non-dense range (its range misses vectors correlated with the geometric sequence (1, λ̄, λ̄², …), which spans the cokernel), so every such λ is in the residual spectrum. On the boundary |λ| = 1 we get continuous spectrum. Net result: σ(S) = the closed unit disc {|λ| ≤ 1} — a two-dimensional spectrum for an operator with zero eigenvalues.

Contrast the left shift S* (its adjoint): S*(x₁, x₂, …) = (x₂, x₃, …). Now every λ with |λ| < 1 is an eigenvalue, with eigenvector (1, λ, λ², …) ∈ ℓ². So σ(S*) is again the closed disc, but now the interior is pure point spectrum. Spectrum is stable under adjoints (σ(S*) = conjugate of σ(S)); its fine structure is not.

Why the hypotheses matter: completeness, boundedness, closedness

Completeness is essential. The claim "T_λ bijective ⟹ R(λ,T) bounded" is exactly the Bounded Inverse Theorem, which fails on incomplete X. On a non-Banach normed space one can have a bijective bounded T_λ whose inverse is unbounded, so λ would be in the resolvent set by the algebraic definition yet R(λ,T) is not a bounded operator — the whole analytic machinery collapses. Complex scalars are essential too. Over ℝ the rotation matrix [[0,−1],[1,0]] has empty (real) spectrum; nonemptiness of σ(T) genuinely needs ℂ so Liouville and complex power series apply.

For unbounded T (e.g. differential operators like d/dx), boundedness is dropped but one requires T to be closed; then ρ(T) is still open and R(·,T) analytic, but σ(T) need no longer be bounded — it can be all of ℂ, ∅, or a half-plane. This connects to semigroup theory (Hille–Yosida), where control of R(λ,T) for Re λ large generates e^{tT}.

Applications: holomorphic functional calculus and beyond

The analyticity of R(·,T) is what makes the holomorphic (Riesz–Dunford) functional calculus possible: for f holomorphic on a neighborhood of σ(T), define f(T) = (1/2πi) ∮_Γ f(λ) R(λ,T) dλ over a contour Γ enclosing σ(T). This is a Banach-algebra Cauchy integral formula; it yields e^{tT}, spectral projections (taking f = indicator-like functions around isolated spectral pieces), and the spectral mapping theorem σ(f(T)) = f(σ(T)). For self-adjoint or normal operators on Hilbert space it upgrades to the full Borel functional calculus and the spectral theorem, via Stone's formula recovering the spectral measure from boundary values of R(λ,T). Resolvent estimates ‖R(λ,T)‖ ≤ M/|Re λ − ω| drive the Hille–Yosida theorem for C₀-semigroups (evolution PDEs, Markov processes), while the resolvent's poles and residues encode eigenprojections in perturbation theory (Kato). In number theory and physics, resolvent traces Tr R(λ,T) build spectral zeta functions and heat kernels. The resolvent is, quietly, the bridge from operators to complex analysis.

How λ sits relative to T decides which part of the spectrum (or resolvent set) it lands in — via injectivity, surjectivity, and boundedness of the inverse.

Condition on T−λI	T−λI injective?	Range dense / all of X?	Where λ lives
Bijective with bounded inverse	Yes	Range = X	Resolvent set ρ(T)
Not injective (has nontrivial kernel)	No	—	Point spectrum σ_p(T) (eigenvalue)
Injective, range dense but ≠ X	Yes	Dense, not closed	Continuous spectrum σ_c(T)
Injective, range not dense	Yes	Range not dense	Residual spectrum σ_r(T)
Bounded self-adjoint on Hilbert H	—	—	σ(T) ⊂ ℝ, σ_r(T) = ∅