Sobolev Spaces as Banach Algebras

Why multiplication matters¶

So far we have used Sobolev spaces as linear function spaces: we add functions, take limits, and apply linear operators. But functions can also be multiplied, and a natural question arises: when is the pointwise product $uv$ again in the same Sobolev space? The answer turns out to connect the embedding theory we just developed to the structure of multiplication operators in spectral methods.

The algebra property from embeddings¶

An algebra over $\mathbb{R}$ (or $\mathbb{C}$) is a vector space $A$ that also carries a multiplication, a bilinear map $A \times A \to A$ that is compatible with the linear structure. You already know several:

A Banach algebra is a Banach space whose norm is submultiplicative: $\|uv\| \leq C\|u\|\|v\|$. This single inequality says that the multiplication is continuous as a bilinear map, or equivalently, that the space is closed under multiplication with quantitative control.

Not every function space is an algebra. The product of two $L^2$ functions is generally only in $L^1$ (by Cauchy--Schwarz, $\|fg\|_{L^1} \leq \|f\|_{L^2}\|g\|_{L^2}$, but $fg$ need not be in $L^2$). So $L^2$ with pointwise multiplication is not an algebra; multiplication takes you outside the space. The question for Sobolev spaces is: does the extra regularity fix this?

The key observation is that $C^0(\overline\Omega)$ is a Banach algebra (the sup norm is submultiplicative), so any Banach space that embeds continuously into $C^0$ inherits the algebra property. If $X \hookrightarrow C^0$ with $\|u\|_\infty \leq C\|u\|_X$, then

whenever the product rule and embedding estimate can be combined. The Sobolev embedding theorem (Theorem 2) gives $H^s \hookrightarrow C^0$ exactly when $s > d/2$. This is the threshold.

Multiplication in frequency space: why regularity implies bandedness¶

The Banach algebra property tells us that multiplication by $u \in H^s$ is a bounded operator on $H^s$. But what does this operator look like when we expand in a spectral basis? The answer reveals a striking structural property: regularity forces approximate bandedness.

To see this most cleanly, work on the torus $\mathbb{T} = [0, 2\pi)$ with Fourier basis $\{e_k(x) = e^{ikx}\}_{k \in \mathbb{Z}}$. Expand $u = \sum_k \hat{u}_k e_k$. The multiplication operator $M_u : v \mapsto uv$ has the matrix representation

Now the connection to regularity becomes immediate. If $u \in H^s$, then

In particular, the off-diagonal entries of $M_u$ decay:

The following picture makes this concrete.

Source

import numpy as np
import matplotlib.pyplot as plt
from matplotlib.colors import LogNorm

fig, axes = plt.subplots(1, 3, figsize=(15, 4.5))

N = 64  # number of Fourier modes

# --- Three functions with increasing regularity ---
k = np.arange(-N//2, N//2)

# 1. Step function (discontinuous, in H^{1/2-ε})
u_hat_step = np.zeros(N, dtype=complex)
for j in range(N):
    kj = k[j]
    if kj != 0:
        u_hat_step[j] = (1 - (-1.0)**kj) / (1j * np.pi * kj)

# 2. Hat function (Lipschitz, in H^1)
u_hat_hat = np.zeros(N, dtype=complex)
for j in range(N):
    kj = k[j]
    if kj != 0:
        u_hat_hat[j] = 2 * (1 - np.cos(np.pi * kj / 2)) / (np.pi * kj)**2
    else:
        u_hat_hat[j] = 0.5

# 3. Gaussian (analytic)
sigma = 0.3
u_hat_gauss = np.exp(-sigma**2 * k**2 / 2)
u_hat_gauss = u_hat_gauss / np.max(np.abs(u_hat_gauss))

functions = [
    (u_hat_step, 'Step function\n(discontinuous)'),
    (u_hat_hat, 'Hat function\n($H^1$, Lipschitz)'),
    (u_hat_gauss, 'Gaussian\n(analytic)')
]

for ax, (u_hat, title) in zip(axes, functions):
    # Build Toeplitz multiplication matrix
    M = np.zeros((N, N), dtype=complex)
    # Shift to standard ordering for building Toeplitz matrix
    u_hat_shifted = np.fft.fftshift(u_hat)
    for j in range(N):
        for l in range(N):
            idx = (j - l) % N
            M[j, l] = u_hat_shifted[idx]

    absM = np.abs(M)
    absM = np.maximum(absM, 1e-16)  # floor for log scale

    im = ax.imshow(absM[:32, :32], norm=LogNorm(vmin=1e-10, vmax=absM.max()),
                   cmap='viridis', aspect='equal')
    ax.set_title(title, fontsize=11)
    ax.set_xlabel('Column $k$', fontsize=10)
    ax.set_ylabel('Row $j$', fontsize=10)
    ax.tick_params(labelsize=8)
    plt.colorbar(im, ax=ax, fraction=0.046, pad=0.04)

plt.suptitle(r'Multiplication matrices $|(M_u)_{jk}| = |\hat{u}_{j-k}|$: regularity $\to$ bandedness',
             fontsize=12, y=1.02)
plt.tight_layout()
plt.show()

The multiplication matrix in the Fourier basis for three functions of increasing regularity. Left: a step function (discontinuous) produces a full, slowly-decaying matrix. Center: a hat function ($H^1$) produces algebraic off-diagonal decay. Right: a Gaussian (analytic) produces exponential decay, so the matrix is effectively banded. The color scale is logarithmic.

The intuition: regularity is frequency localization¶

Why should smoothness produce banded multiplication matrices? The intuition comes from three equivalent ways to say the same thing:

1. Regularity = frequency concentration. A function in $H^s$ has Fourier coefficients decaying like $|k|^{-s-1/2}$. Higher regularity means the function’s energy is concentrated at low frequencies.

2. Multiplication = convolution in frequency. The matrix $M_u$ is Toeplitz with entries $\hat{u}_{j-k}$. Applying $M_u$ to a vector of Fourier coefficients is convolution with the sequence $(\hat{u}_k)$.

3. Convolving with something narrow produces something narrow. If the sequence $(\hat{u}_k)$ is concentrated near $k = 0$ (because $u$ is smooth), then convolution with it only couples mode $k$ to nearby modes $k \pm \Delta k$, which is exactly the statement that $M_u$ is approximately banded.

Putting these together: smooth functions act locally in frequency space. Multiplying by a smooth function is like a short-range interaction between Fourier modes. Multiplying by a rough function is a long-range interaction that couples all modes to all other modes.

Connection to spectral and collocation methods¶

This structure is exactly what makes spectral methods efficient for smooth problems.

In a pseudospectral (collocation) method, we want to compute the Fourier (or Chebyshev) coefficients of a product $uv$. The naive approach is to form $M_u$ and multiply, an $O(N^2)$ operation. But the collocation approach exploits a factorization:

where $F$ is the DFT matrix and $\operatorname{diag}(\mathbf{u})$ contains the values of $u$ at collocation points. Multiplication is diagonal in physical space, so we:

1. Transform $\hat{v} \to v$ at collocation points (inverse FFT, $O(N \log N)$),

2. Multiply pointwise: $w_j = u_j v_j$ ($O(N)$),

3. Transform back $w \to \hat{w}$ (forward FFT, $O(N \log N)$).

The total cost is $O(N \log N)$ instead of $O(N^2)$. But here is the subtle point: this procedure is exact only for trigonometric polynomials of degree $\leq N$. For general smooth functions, the product $uv$ may have Fourier modes up to degree $2N$, and the modes above $N$ get aliased back into the $[-N, N]$ range.

The Banach algebra property tells us this aliasing is harmless for smooth functions: since $|\hat{u}_k| \lesssim |k|^{-s-1/2}$, the aliased modes are exponentially small (for analytic $u$) or at least rapidly decaying (for Sobolev $u$). The error from aliasing is controlled by the off-diagonal decay of $M_u$, precisely the entries we truncate by working with a finite matrix.