Fourier Transform 13 | Fourier Series Converges in L²

Q: Why does the argument focus on L² convergence instead of pointwise convergence?

The result is formulated in terms of the L² norm, ||f−f_n||₂, which measures average squared error via an integral over [−π,π]. The transcript emphasizes that the conclusion is not pointwise convergence; it is convergence in the integral sense. This matters because L² functions may be irregular, so pointwise behavior can fail even when the L² error goes to zero.

Q: How does density of step functions in L² get proved?

First, continuous functions are dense in L²: for any ε>0, an L² function f can be approximated by a continuous 2π-periodic G with ||f−G||₂<ε. Then uniform continuity of G on the compact interval [−π,π] lets one choose a partition into finitely many subintervals of length <δ so that G changes by less than ε across each subinterval. A step function H is built by taking the supremum of G on each subinterval. This construction ensures |G(x)−H(x)|<ε for all but finitely many boundary points, so ||G−H||₂ is small; the triangle inequality then gives ||f−H||₂ small.

Q: What role does orthogonality play in the final convergence step?

Fourier partial sums correspond to orthogonal projections onto the span of the relevant exponentials in L². Because of this, the error components are orthogonal, enabling a generalized Pythagorean theorem: the squared norm of the combined error equals the sum of squared norms of orthogonal parts. In practice, this yields an inequality that bounds ||f−f_n||₂ using ||f−H||₂ and ||H−(Fourier partial sum of H)||₂, where the second term goes to zero.

Q: Why is the approximation by a step function H enough to prove convergence for an arbitrary f?

Once ||f−H||₂ can be made arbitrarily small, the remaining task is to control the Fourier error for H. The transcript notes that Fourier convergence (and Parseval’s identity) has already been proved for step functions. Therefore, ||H−f_n(H)||₂→0. Combining this with the orthogonality-based inequality and the triangle inequality forces ||f−f_n(f)||₂→0 for the original f.

Q: How does Parseval’s identity connect to L² convergence?

The transcript states that L² convergence of Fourier series is equivalent to Parseval’s identity on L². Parseval’s identity relates the L² norm squared of f to the sum of squares of its Fourier coefficients. Since Fourier coefficients encode the orthogonal projection structure, proving convergence in L² for all L² functions completes the extension of Parseval’s identity to the entire space.

TL;DR

Fourier series for 2π-periodic functions in L² converge to the function in the L² norm, meaning ||f−f_n||₂→0 as n→∞.

Briefing Cornell Notes

Briefing

Fourier series for 2π-periodic, square-integrable functions converge in the L² sense: as n→∞, the L² norm of the difference between a function f and its nth Fourier partial sum f_n goes to zero. This is the key result because it upgrades earlier convergence facts from special classes of functions to the full space L², and it is equivalent to Parseval’s identity holding for every L² function—meaning the energy (norm squared) of f can be recovered from its Fourier coefficients.

The proof strategy starts by reframing the goal in Hilbert-space terms. Using the standard inner product on L² over [−π,π], the complex exponentials form an orthonormal system. For each f in L², the Fourier coefficients define an orthogonal projection onto the span of the first 2n+1 exponentials, so the error f−f_n is orthogonal to the projection space. That orthogonality is what later produces a “generalized Pythagorean theorem” inequality that controls the L² error.

The main technical hurdle is extending convergence from step functions (where Parseval’s identity was already established) to arbitrary L² functions. The argument relies on density: continuous functions are dense in L², and then step functions are dense in L² as well. Concretely, given any ε>0 and any f∈L², one first picks a continuous 2π-periodic function G with ||f−G||₂<ε. Because G is continuous on the compact interval [−π,π], it is uniformly continuous. That uniform continuity allows choosing a partition of [−π,π] into finitely many subintervals of length less than some δ so that G varies by less than ε across each subinterval.

From this partition, a step function H is constructed by taking, on each subinterval, the supremum (or equivalently a maximum on a closed version) of G. On almost every point (all but finitely many partition boundaries), the difference |G(x)−H(x)| is bounded by ε. Integrating the squared error then yields ||G−H||₂ bounded by a constant multiple of ε (the transcript gives √(2π·ε)). With the triangle inequality, ||f−H||₂ can be made arbitrarily small, proving step functions are dense in L².

Finally, the Fourier convergence for general f follows from a two-step approximation. Choose H close to f in L². Since Fourier partial sums converge to H in L² for step functions, ||H−f_n(H)||₂→0. For the remaining gap, orthogonality of the Fourier projection gives an inequality: the L² error ||f−f_n(f)||₂ is controlled by ||f−H||₂ plus the vanishing term ||H−f_n(H)||₂. Taking n→∞ and using that ε was arbitrary forces the limit error to be zero. The result is L² convergence (not pointwise), completing the extension of Parseval’s identity to all of L².

Cornell Notes

For 2π-periodic functions in L², Fourier partial sums converge to the function in the L² norm. The proof hinges on two ideas: (1) step functions are dense in L², and (2) Fourier partial sums act as orthogonal projections in the Hilbert space L², giving a generalized Pythagorean control of errors. Density is established by approximating an L² function f by a continuous function G, then using uniform continuity of G on [−π,π] to build a step function H whose values track G on each small subinterval. Once ||f−H||₂ is small and Fourier convergence holds for step functions, orthogonality and the triangle inequality force ||f−f_n||₂→0. This is exactly the L² version of Parseval’s identity for all L² functions.

Why does the argument focus on L² convergence instead of pointwise convergence?

The result is formulated in terms of the L² norm, ||f−f_n||₂, which measures average squared error via an integral over [−π,π]. The transcript emphasizes that the conclusion is not pointwise convergence; it is convergence in the integral sense. This matters because L² functions may be irregular, so pointwise behavior can fail even when the L² error goes to zero.

How does density of step functions in L² get proved?

First, continuous functions are dense in L²: for any ε>0, an L² function f can be approximated by a continuous 2π-periodic G with ||f−G||₂<ε. Then uniform continuity of G on the compact interval [−π,π] lets one choose a partition into finitely many subintervals of length <δ so that G changes by less than ε across each subinterval. A step function H is built by taking the supremum of G on each subinterval. This construction ensures |G(x)−H(x)|<ε for all but finitely many boundary points, so ||G−H||₂ is small; the triangle inequality then gives ||f−H||₂ small.

What role does orthogonality play in the final convergence step?

Fourier partial sums correspond to orthogonal projections onto the span of the relevant exponentials in L². Because of this, the error components are orthogonal, enabling a generalized Pythagorean theorem: the squared norm of the combined error equals the sum of squared norms of orthogonal parts. In practice, this yields an inequality that bounds ||f−f_n||₂ using ||f−H||₂ and ||H−(Fourier partial sum of H)||₂, where the second term goes to zero.

Why is the approximation by a step function H enough to prove convergence for an arbitrary f?

Once ||f−H||₂ can be made arbitrarily small, the remaining task is to control the Fourier error for H. The transcript notes that Fourier convergence (and Parseval’s identity) has already been proved for step functions. Therefore, ||H−f_n(H)||₂→0. Combining this with the orthogonality-based inequality and the triangle inequality forces ||f−f_n(f)||₂→0 for the original f.

How does Parseval’s identity connect to L² convergence?

The transcript states that L² convergence of Fourier series is equivalent to Parseval’s identity on L². Parseval’s identity relates the L² norm squared of f to the sum of squares of its Fourier coefficients. Since Fourier coefficients encode the orthogonal projection structure, proving convergence in L² for all L² functions completes the extension of Parseval’s identity to the entire space.

Review Questions

What two density results are used to reduce the general L² case to step functions, and how is uniform continuity used in the step-function construction?
Where exactly does orthogonality enter the proof, and how does it produce a Pythagorean-type estimate for the L² error?
Why does the final argument yield L² convergence rather than pointwise convergence?

Key Points

1
Fourier series for 2π-periodic functions in L² converge to the function in the L² norm, meaning ||f−f_n||₂→0 as n→∞.
2
L² convergence is equivalent to Parseval’s identity holding for all L² functions, linking Fourier coefficients to the function’s energy.
3
Continuous 2π-periodic functions are dense in L², so any L² function can be approximated arbitrarily well in L² by a continuous one.
4
Uniform continuity of a continuous approximation on [−π,π] enables constructing a step function H whose values track G on each small subinterval.
5
A step function H is built by taking the supremum of G on each subinterval, ensuring |G−H| is small except at finitely many boundary points.
6
Fourier partial sums act as orthogonal projections in L², and orthogonality yields a generalized Pythagorean control of the approximation error.
7
Combining density (f≈H) with Fourier convergence for step functions (H≈f_n(H)) forces the error for f to vanish in the L² norm.

Highlights

The central theorem is L² convergence of Fourier series for every 2π-periodic square-integrable function, not pointwise convergence.

Step functions become the bridge: they are dense in L², built using uniform continuity of a continuous approximation.

Orthogonality of Fourier projections supplies a Pythagorean-type inequality that turns approximation closeness into convergence.

The result completes the extension of Parseval’s identity from special cases to the entire L² space.

Topics

Fourier Series
L² Convergence
Parseval Identity
Orthogonal Projections
Step Functions Density