English中文日本語Français

The Basic Geometric Meaning of Frequency in Fourier Analysis (1)

Fourier geometry · Part I

Fourier atoms, directional variation, and the reciprocal relationship between spatial scale and frequency.

Home / Harmonic Analysis / The Geometry of Frequency

On this page

• Fourier atoms and frequency
• Directional smoothness and frequency localization
• Spatial scale versus frequency scale

When I first encountered the concept of frequency in harmonic analysis, I had only a rough understanding of it.

It wasn’t until I studied Littlewood–Paley theory that I began to appreciate the geometric structure underlying the notion of frequency. Even then, I hadn’t really thought deeply about what frequency meant from a geometric perspective.

Recently, I’ve been studying some of Seeger’s work on singular integrals. Certain techniques used in his proofs have pushed me to revisit these fundamental ideas and think more carefully about the geometry of frequency.

In fact, it is due to the need for my recent research that I’ve begun to explore these geometric ideas in greater depth.

The reason I chose the word “basic” in the title of my article is that the ideas themselves are simple, yet profound.

I’m planning a series of about three articles, through which I’ll gradually build up from these fundamental ideas to Seeger’s work on singular integrals.

Although the ideas can be quite deep, I’ll do my best to explain them in a clear, intuitive, and accessible way, so that even readers without a background in mathematics can follow along.

So, if you’re not a mathematics major but still have an interest in the subject, don’t worry! I’ll try to make these ideas approachable without requiring much advanced mathematical knowledge. Just bring your curiosity and enjoy the journey!

---

Fourier atoms and frequency

We begin with the simplest Fourier atom:

$$e^{ix\cdot\xi}.$$

Its “frequency” is $\xi$. Why? Differentiation in an arbitrary direction $v$ yields

$$(v\cdot\nabla)\,e^{ix\cdot\xi}=i(v\cdot\xi)\,e^{ix\cdot\xi}.$$

Thus $\xi$ tells you directly in which directions the function varies rapidly, and how small the corresponding scale of variation is. More precisely, if $v$ is a unit vector and $v\cdot\xi\ne0$, the oscillation along $v$ has spatial period $2\pi/|v\cdot\xi|$.

In particular, if

$$v\cdot\xi=0,$$

then

$$(v\cdot\nabla)\,e^{ix\cdot\xi}=0.$$

In other words, the wave is completely invariant in the direction $v$.

Consequently, the geometric meaning of $\xi$ is not simply that “the wave travels in the direction of $\xi$” (there is no time variable here), but rather the following: $\xi$ is the gradient of the phase $x\cdot\xi$ and, when $\xi\ne0$, a normal vector to the surfaces of constant phase. The equiphase surfaces

$$x\cdot\xi=\mathrm{constant}$$

are hyperplanes orthogonal to $\xi$.

Directional smoothness and frequency localization

This immediately explains an important phenomenon. Suppose a function is nearly invariant along a unit direction $\nu$:

$$(\nu\cdot\nabla)f\approx 0.$$

Here $\hat f$ denotes the Fourier transform of $f$, defined for sufficiently nice functions by

$$\hat f(\xi)=\int_{\mathbb R^n}f(x)e^{-ix\cdot\xi}\,dx.$$

After Fourier transformation,

$$\widehat{(\nu\cdot\nabla)f}(\xi)=i(\nu\cdot\xi)\,\hat f(\xi).$$

If the left-hand side is small in $L^2$, the right-hand side is also small in $L^2$. More precisely, for any $\delta>0$, the $L^2$ mass of $\hat f$ in the region $|\nu\cdot\xi|\ge\delta$ is controlled by $\delta^{-2}\|(\nu\cdot\xi)\hat f\|_{L^2}^2$. Hence the Fourier mass tends to concentrate where

$$|\nu\cdot\xi|\approx 0,$$

that is, near the hyperplane

$$\nu^\perp=\{\xi\in\mathbb R^n:\nu\cdot\xi=0\}.$$

This is a statement about concentration of Fourier mass, not necessarily about the exact support of $\hat f$.

Spatial scale versus frequency scale

Another intuition worth fixing at this stage is that “large scale” in space corresponds to “small scale” in frequency. For example, for a one-dimensional function

$$f(x)=\phi(x/L),$$

where $L>0$ and $\phi$ is a fixed, sufficiently regular function describing the original shape of $f$. If $\phi$ varies on a spatial scale of order one, then the spatial scale of variation of $f$ is of order $L$. Its Fourier transform satisfies the exact identity

$$\hat f(\xi)=L\,\hat\phi(L\xi).$$

Thus, if $\hat\phi(\eta)$ is mainly concentrated where $|\eta|\lesssim1$, then the Fourier transform of $f$ is mainly concentrated where

$$|\xi|\lesssim L^{-1}.$$

Therefore: slow variation corresponds to low frequency, and rapid variation corresponds to high frequency.

Discussion

Questions, corrections, and further thoughts are welcome. All language editions share this discussion.

Comments load only when you choose to open them. This connects to Giscus and GitHub. A GitHub account is needed to post.

Open discussions on GitHub ↗