JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Bounded operators: linear maps when matrices run out

A linear map between infinite-dimensional spaces has no matrix, and continuity stops being automatic. This guide defines bounded operators, the operator norm, and the adjoint — and uses Riesz representation to recover the transpose you knew from Vol I.

No matrix, and continuity is no longer free

In Vol I every linear transformation between finite-dimensional spaces was a matrix once you chose bases, and every such map was automatically continuous. Both facts evaporate in infinite dimensions: there is no finite matrix, and a linear map can be wildly discontinuous, sending a convergent input sequence to a divergent output.

An operator T is a bounded operator when there is a constant C with ||Tx|| <= C ||x|| for all x. "Bounded" here does not mean its values stay in a box — it means it does not stretch lengths by more than a fixed factor. The decisive theorem: for linear maps, bounded is equivalent to continuous. So bounded operators are exactly the linear maps that respect limits.

Worked examples: shift, multiplication, differentiation

Three operators on sequence/function spaces

1) RIGHT SHIFT on ell^2:
     S(x_1, x_2, x_3, ...) = (0, x_1, x_2, ...)
   ||Sx|| = ||x||  exactly  ->  bounded, ||S|| = 1  (an isometry).
   Note: S is injective but NOT surjective (nothing maps to e_1).
   In finite dim, injective => surjective. Here it FAILS.

2) MULTIPLICATION on L^2[0,1]:   (M_g f)(t) = g(t) f(t)
   if |g(t)| <= K everywhere, ||M_g f|| <= K ||f||
   ->  bounded, ||M_g|| = sup |g| = ess-sup of g.

3) DIFFERENTIATION D f = f'  on L^2:
   take f_n(t) = sin(n t):   ||f_n|| stays ~ constant,
   but D f_n = n cos(n t):   ||D f_n|| grows like n -> infinity.
   No constant C bounds it  ->  D is UNBOUNDED (discontinuous).
Shift and multiplication are bounded; differentiation is the archetypal unbounded operator.

The shift operator already breaks a load-bearing Vol I intuition: in finite dimensions an injective operator is automatically surjective (the rank-nullity theorem guarantees it). The right shift is injective yet misses e_1 entirely — left versus right inverses come apart, and one-sided invertibility becomes a real phenomenon. This single example will return to haunt the spectrum in Guide 5.

Riesz representation and the adjoint

How do we recover the transpose without a matrix? Through the Riesz representation theorem: in a Hilbert space, *every* bounded linear functional f(x) is just an inner product f(x) = <x, v> against one fixed vector v, and ||f|| = ||v||. Functionals and vectors are the same data. This is the engine that turns abstract dual statements back into concrete geometry.

Riesz lets us define the adjoint T* of a bounded operator by the rule <Tx, y> = <x, T* y> for all x, y. For a matrix this is exactly the conjugate transpose, so T* generalizes A^T (or A^* over C). An operator is self-adjoint when T = T* — the infinite-dimensional twin of a symmetric matrix — and self-adjointness is what will make the spectral theorem work in Guide 4.

Adjoint via  <Tx, y> = <x, T*y>

Right shift S on ell^2:   S(x_1, x_2, ...) = (0, x_1, x_2, ...)
  <Sx, y> = x_1 y_2 + x_2 y_3 + ...
          = <x, (y_2, y_3, y_4, ...)>
  =>  S*(y_1, y_2, y_3, ...) = (y_2, y_3, ...)  = the LEFT shift.

So (right shift)* = left shift. Check the asymmetry:
  S* S = I     (left-then-right undoes -> identity)
  S S* != I    (S S* kills the first coordinate)
Left and right inverses differ -> a purely infinite-dim effect.
The adjoint of the right shift is the left shift, and S*S = I while SS* != I.