The Fourier transform is one of the central tools of modern analysis, providing a systematic bridge between spatial and frequency descriptions of [functions](/page/Function). Its power lies in its ability to convert differential operations into algebraic ones, making it indispensable in the study of partial differential equations, signal processing, and quantum mechanics. This article develops the transform on $L^1(\mathbb{R}^n)$, extends it to $L^2(\mathbb{R}^n)$ via the Plancherel theorem, and establishes its key interactions with [differentiation](/page/Derivative) and convolution. ## Motivation [motivation] ### From [Fourier Series](/page/Fourier%20Series) to the Fourier Transform The idea of decomposing a function into oscillatory components begins with periodic functions. On the circle $\mathbb{R} / 2\pi\mathbb{Z}$, any square-integrable function $f$ can be expanded as a Fourier series \begin{align*} f(x) &= \sum_{k \in \mathbb{Z}} c_k \, e^{ikx}, \end{align*} where the coefficients $c_k$ capture the contribution of each integer frequency $k$. The complex exponentials $\{e^{ikx}\}_{k \in \mathbb{Z}}$ form an orthonormal basis for $L^2$ of the circle, and projection onto this basis diagonalises the derivative operator: \begin{align*} \frac{d}{dx} e^{ikx} &= ik \, e^{ikx}. \end{align*} When one passes from periodic to non-periodic functions — from the circle to the full Euclidean space $\mathbb{R}^n$ — the discrete set of frequencies $k \in \mathbb{Z}$ must be replaced by a [continuous](/page/Continuity) frequency variable $\xi \in \mathbb{R}^n$, and the sum over Fourier coefficients becomes an integral. The resulting operation is the Fourier transform: it takes a function of the spatial variable $x$ and produces a function of the frequency variable $\xi$ that encodes the amplitude and phase of each oscillatory component $e^{ix \cdot \xi}$. ### The PDE Perspective The deepest reason for introducing the Fourier transform is its interaction with differential operators. Consider a constant-coefficient linear PDE such as the [heat equation](/page/Heat%20Equation) $\partial_t u = \Delta u$ on $\mathbb{R}^n$. Applying the Fourier transform in the spatial variables converts the Laplacian $\Delta$ into multiplication by $-|\xi|^2$, reducing the PDE to the family of ordinary differential equations \begin{align*} \partial_t \hat{u}(\xi, t) &= -|\xi|^2 \, \hat{u}(\xi, t), \end{align*} one for each frequency $\xi$. Each of these ODEs can be solved explicitly, and the solution in physical space is recovered by inverting the transform. Without this device, obtaining such explicit solution formulas would require entirely different — and often more laborious — techniques. More broadly, the Fourier transform reveals the spectral content of a differential operator: it replaces a differential operator $P(D)$ with the multiplication operator $P(\xi)$, so that the invertibility, regularity, and decay properties of solutions can be read directly from the algebraic symbol $P(\xi)$. This philosophy underlies much of modern PDE theory, from the construction of fundamental solutions to the definition of [Sobolev spaces](/page/Sobolev%20Space) via frequency-side decay conditions. [/motivation] ## The Fourier Transform on $L^1(\mathbb{R}^n)$ We adopt the convention in which the forward Fourier transform carries no prefactor. This is the standard convention in harmonic analysis and PDE theory (used, for instance, by Stein–Weiss and Grafakos). The normalisation constant $(2\pi)^n$ appears instead in the inversion formula. A principal advantage of this choice is that [convolutions](/page/Convolution) transform cleanly: \begin{align*} \widehat{f * g} &= \hat{f} \, \hat{g} \end{align*} with no stray factors. [definition:Fourier Transform] Let $f \in L^1(\mathbb{R}^n; \mathbb{C})$, where [integrability](/page/Integral) is with respect to Lebesgue measure $\mathcal{L}^n$. The **Fourier transform** of $f$ is the function $\hat{f}$ defined by \begin{align*} \hat{f}: \mathbb{R}^n &\to \mathbb{C} \\ \xi &\mapsto \int_{\mathbb{R}^n} f(x) \, e^{-i x \cdot \xi} \, d\mathcal{L}^n(x). \end{align*} We also write $\mathcal{F}f = \hat{f}$. [/definition] The integral converges absolutely for every $\xi \in \mathbb{R}^n$ because $|f(x) e^{-ix \cdot \xi}| = |f(x)|$ and $f \in L^1$. A direct estimate gives \begin{align*} \|\hat{f}\|_{L^\infty} &\leq \|f\|_{L^1}, \end{align*} so the Fourier transform maps $L^1(\mathbb{R}^n)$ continuously into $L^\infty(\mathbb{R}^n)$. The following classical result sharpens this observation: $\hat{f}$ is not merely bounded but also continuous and vanishing at infinity. [quotetheorem:245] The Riemann–Lebesgue lemma says that high-frequency oscillatory components of an integrable function must have vanishingly small amplitude. Uniform continuity follows from dominated convergence: if $\xi_k \to \xi$, then $f(x) e^{-ix \cdot \xi_k} \to f(x) e^{-ix \cdot \xi}$ pointwise and the integrand is dominated by $|f|$, so $\hat{f}(\xi_k) \to \hat{f}(\xi)$. The decay at infinity is more subtle. One first verifies it for indicator functions of rectangles by explicit computation (the resulting integral is a product of terms of the form $\sin(a\xi_j)/\xi_j$, which decay), and then extends to general $L^1$ functions by density, since simple functions are dense in $L^1$ and the Fourier transform is continuous $L^1 \to L^\infty$. An important limitation: the lemma does *not* assert any quantitative rate of decay — for that, one needs additional regularity. If $f$ has $k$ derivatives in $L^1$, then the differentiation rule (stated below) gives \begin{align*} \hat{f}(\xi) &= O(|\xi|^{-k}) \quad \text{as } |\xi| \to \infty. \end{align*} ### The Gaussian as a Test Case The most important explicit Fourier transform computation is that of the Gaussian. This example serves both as a sanity check for the definition and as a building block for many later arguments, including the heat kernel, approximations to the identity, and the proof of the Plancherel theorem. [example:Fourier Transform Of The Gaussian] Let $\sigma > 0$ and define $g_\sigma \in L^1(\mathbb{R}^n; \mathbb{C})$ by \begin{align*} g_\sigma: \mathbb{R}^n &\to \mathbb{R} \\ x &\mapsto e^{-\sigma |x|^2 / 2}. \end{align*} We claim that \begin{align*} \hat{g}_\sigma(\xi) &= \left(\frac{2\pi}{\sigma}\right)^{n/2} e^{-|\xi|^2 / (2\sigma)}. \end{align*} Since the exponential factors over coordinates — $g_\sigma(x) = \prod_{j=1}^n e^{-\sigma x_j^2/2}$ and $e^{-ix \cdot \xi} = \prod_{j=1}^n e^{-ix_j \xi_j}$ — it suffices by Fubini's theorem to verify the one-dimensional case $n = 1$. We compute \begin{align*} \hat{g}_\sigma(\xi) &= \int_{\mathbb{R}} e^{-\sigma x^2 / 2} \, e^{-ix\xi} \, d\mathcal{L}^1(x). \end{align*} Completing the square in the exponent: \begin{align*} -\frac{\sigma}{2} x^2 - ix\xi &= -\frac{\sigma}{2}\left(x + \frac{i\xi}{\sigma}\right)^2 - \frac{\xi^2}{2\sigma}. \end{align*} Substituting $u = x + i\xi/\sigma$ and shifting the contour of integration: the function $z \mapsto e^{-\sigma z^2/2}$ is entire, and on a rectangular contour with vertices $\pm R$ and $\pm R + i\xi/\sigma$, the contributions from the two vertical sides are bounded by $C e^{-\sigma R^2/2} \to 0$ as $R \to \infty$. By Cauchy's theorem the integral along the shifted contour equals the integral along $\mathbb{R}$, giving \begin{align*} \hat{g}_\sigma(\xi) &= e^{-\xi^2/(2\sigma)} \int_{\mathbb{R}} e^{-\sigma u^2/2} \, d\mathcal{L}^1(u) = \left(\frac{2\pi}{\sigma}\right)^{1/2} e^{-\xi^2/(2\sigma)}. \end{align*} The $n$-dimensional result follows by taking the product over all coordinates $j = 1, \ldots, n$. [/example] The Gaussian example illustrates a fundamental principle: smooth, rapidly decaying functions in the spatial domain correspond to smooth, rapidly decaying functions in the frequency domain. In particular, the Gaussian is (up to dilation) a fixed point of the Fourier transform in the sense that $\hat{g}_\sigma$ is again a Gaussian. The special case $\sigma = 1$ gives \begin{align*} \hat{g}_1(\xi) &= (2\pi)^{n/2} \, e^{-|\xi|^2/2} = (2\pi)^{n/2} \, g_1(\xi), \end{align*} confirming that the standard Gaussian is an eigenfunction of $\mathcal{F}$ with eigenvalue $(2\pi)^{n/2}$. ### An Obstruction: The Indicator Function A second instructive example reveals an important limitation of the $L^1$ theory. [example:Fourier Transform Of An Indicator] Let $a > 0$ and define $f = \mathbb{1}_{[-a, a]} \in L^1(\mathbb{R}; \mathbb{C})$. For $\xi \neq 0$, \begin{align*} \hat{f}(\xi) &= \int_{-a}^{a} e^{-ix\xi} \, d\mathcal{L}^1(x) = \left[\frac{e^{-ix\xi}}{-i\xi}\right]_{x = -a}^{x = a} = \frac{e^{-ia\xi} - e^{ia\xi}}{-i\xi} = \frac{2\sin(a\xi)}{\xi}. \end{align*} At $\xi = 0$, we have $\hat{f}(0) = \int_{-a}^{a} d\mathcal{L}^1(x) = 2a$, which is consistent with $\lim_{\xi \to 0} 2\sin(a\xi)/\xi = 2a$. The function $\xi \mapsto 2\sin(a\xi)/\xi$ is continuous and decays like $|\xi|^{-1}$, but it is *not* integrable on $\mathbb{R}$: the harmonic-type divergence \begin{align*} \int_1^\infty \frac{|\sin(a\xi)|}{\xi} \, d\mathcal{L}^1(\xi) &= +\infty \end{align*} prevents absolute convergence. [/example] This example demonstrates two important phenomena. First, discontinuities in the spatial domain produce slow decay in the frequency domain — here only $|\xi|^{-1}$, consistent with the general principle that smoothness and frequency decay are dual. Second, the Fourier transform does *not* map $L^1(\mathbb{R})$ into itself. This non-integrability of $\hat{f}$ is precisely the obstruction that prevents the naive application of the inversion formula and motivates the introduction of the [Schwartz space](/page/Schwartz%20Space) $\mathcal{S}(\mathbb{R}^n)$, where both $f$ and $\hat{f}$ are guaranteed to be rapidly decreasing. ## Inversion The Fourier transform is injective on $L^1(\mathbb{R}^n)$ — if $\hat{f} = 0$ $\mathcal{L}^n$-almost everywhere, then $f = 0$ $\mathcal{L}^n$-almost everywhere. This follows, for instance, from the density of $L^1 \cap L^2$ and the Plancherel theorem developed below. Recovering $f$ from $\hat{f}$ pointwise, however, requires the Fourier transform itself to be integrable. [definition:Inverse Fourier Transform] Let $g \in L^1(\mathbb{R}^n; \mathbb{C})$. The **inverse Fourier transform** of $g$ is the function $\check{g}$ defined by \begin{align*} \check{g}: \mathbb{R}^n &\to \mathbb{C} \\ x &\mapsto \frac{1}{(2\pi)^n} \int_{\mathbb{R}^n} g(\xi) \, e^{i x \cdot \xi} \, d\mathcal{L}^n(\xi). \end{align*} We also write $\mathcal{F}^{-1}g = \check{g}$. [/definition] The factor $(2\pi)^{-n}$ in the inverse transform is the price of the clean (unscaled) forward definition. In the symmetric convention used by some authors, both forward and inverse transforms carry $(2\pi)^{-n/2}$; in our convention, the full normalisation constant resides in the inverse. This is purely a matter of bookkeeping and does not affect any analytic conclusion. [quotetheorem:246] The hypothesis $\hat{f} \in L^1$ is essential and cannot be dropped. The indicator function example above exhibits an $L^1$ function whose Fourier transform is not integrable; for such functions the inversion integral does not converge absolutely, and one must interpret the reconstruction in a different sense — either via principal-value [limits](/page/Limit) \begin{align*} f(x) &= \lim_{R \to \infty} \frac{1}{(2\pi)^n} \int_{|\xi| \leq R} \hat{f}(\xi) \, e^{ix \cdot \xi} \, d\mathcal{L}^n(\xi), \end{align*} or by working within the Schwartz class $\mathcal{S}(\mathbb{R}^n)$, where $\mathcal{F}$ is an automorphism and the inversion formula holds without any additional hypotheses. The Schwartz space framework is developed on the [Schwartz Space](/pages/1050) page. ## The $L^2$ Theory The definition of the Fourier transform as an integral against $e^{-ix \cdot \xi}$ requires $f \in L^1$, but many functions of interest — particularly in quantum mechanics and signal processing, where $\|f\|_{L^2}^2$ represents probability or energy — belong to $L^2$ without belonging to $L^1$. For instance, the function $x \mapsto (1 + |x|)^{-(n+1)/2}$ lies in $L^2(\mathbb{R}^n)$ but not in $L^1(\mathbb{R}^n)$ when $n \geq 1$. Extending the transform to $L^2$ requires a limiting argument: one approximates $f \in L^2$ by functions in $L^1 \cap L^2$ (for example, by truncation $f \cdot \mathbb{1}_{B(0,R)}$) and shows the resulting [sequence](/page/Sequence) of Fourier transforms converges in $L^2$. [quotetheorem:247] The proof proceeds in two stages. First, one verifies the identity on the Schwartz space $\mathcal{S}(\mathbb{R}^n)$, where both sides are well-defined integrals. The key step is to compute $\int |\hat{f}|^2 \, d\mathcal{L}^n$ by writing $|\hat{f}|^2 = \hat{f} \cdot \overline{\hat{f}}$ and using the fact that complex conjugation interchanges the forward and inverse transforms (up to normalisation), together with the Gaussian computation above. Second, since $\mathcal{S}(\mathbb{R}^n)$ is dense in $L^2(\mathbb{R}^n)$, the isometry extends uniquely by the BLT (bounded linear transformation) theorem. The factor of $(2\pi)^n$ in the Plancherel identity is the trade-off for having a clean convolution theorem: the two conventions distribute the same total normalisation $(2\pi)^n$ differently between the various formulas. Polarising the Plancherel identity via the standard trick \begin{align*} \langle f, g \rangle &= \frac{1}{4}\left(\|f+g\|^2 - \|f-g\|^2 + i\|f+ig\|^2 - i\|f-ig\|^2\right) \end{align*} yields the inner product version. [quotetheorem:248] Parseval's identity says that the Fourier transform preserves the $L^2$ inner product up to the normalisation constant $(2\pi)^n$. In physical terms, this is the statement that total energy (the left-hand side) equals the integral of the spectral energy density (the right-hand side) — the frequency-domain decomposition of energy is lossless. The identity is the polarised form of the [Plancherel Theorem](/theorems/247), and the two results are logically equivalent: either can be derived from the other. ## Interaction with Calculus Operations ### Differentiation The deepest utility of the Fourier transform lies in its conversion of differentiation into algebraic multiplication. This is what makes it the natural tool for constant-coefficient PDEs. [quotetheorem:249] The proof is [integration by parts](/theorems/210): each application of $\partial / \partial x_j$ to $f$ inside the integral transfers onto $e^{-ix \cdot \xi}$, producing a factor of $i\xi_j$. The [boundary](/page/Boundary) terms vanish because $f \in L^1$ forces sufficient decay at infinity. Conversely, multiplication by $x^\alpha$ on the spatial side corresponds to differentiation on the frequency side: \begin{align*} \widehat{x^\alpha f}(\xi) &= i^{|\alpha|} \, D_\xi^\alpha \hat{f}(\xi). \end{align*} This duality between smoothness and decay is one of the central organising principles of Fourier analysis: the smoother a function is, the faster its Fourier transform decays, and conversely, the more rapidly a function decays, the smoother its transform. ### Why This Matters for PDEs To see the [Fourier Differentiation Rule](/theorems/249) in action, consider the Poisson equation $-\Delta u = f$ on $\mathbb{R}^n$ with $n \geq 3$. Applying $\mathcal{F}$ to both sides and using the differentiation rule twice gives \begin{align*} \widehat{\Delta u}(\xi) &= -|\xi|^2 \, \hat{u}(\xi), \end{align*} so the equation becomes \begin{align*} |\xi|^2 \, \hat{u}(\xi) &= \hat{f}(\xi), \end{align*} and formally \begin{align*} \hat{u}(\xi) &= \frac{\hat{f}(\xi)}{|\xi|^2}. \end{align*} The solution in physical space is then $u = \mathcal{F}^{-1}(\hat{f} / |\xi|^2)$. Recognising $|\xi|^{-2}$ as the Fourier transform of a constant multiple of $|x|^{2-n}$ (for $n \geq 3$) recovers the classical Newton potential \begin{align*} u(x) &= c_n \int_{\mathbb{R}^n} |x-y|^{2-n} \, f(y) \, d\mathcal{L}^n(y). \end{align*} The entire derivation reduces to algebra on the frequency side. ### Convolution The convolution of two functions represents a "smearing" or "averaging" operation and arises naturally in probability ([distributions](/page/Distribution) of sums of independent random variables), PDE theory (solutions via fundamental solutions), and signal processing (linear time-invariant filtering). The Fourier transform converts this integral operation into pointwise multiplication. [definition:Convolution] Let $f, g \in L^1(\mathbb{R}^n; \mathbb{C})$. The **convolution** of $f$ and $g$ is the function $f * g$ defined by \begin{align*} f * g: \mathbb{R}^n &\to \mathbb{C} \\ x &\mapsto \int_{\mathbb{R}^n} f(x - y) \, g(y) \, d\mathcal{L}^n(y). \end{align*} By [Young's convolution inequality](/theorems/463), \begin{align*} \|f * g\|_{L^1} &\leq \|f\|_{L^1} \, \|g\|_{L^1}, \end{align*} so $f * g \in L^1(\mathbb{R}^n)$. [/definition] [quotetheorem:250] The proof is a direct application of Fubini's theorem: writing out the definitions and exchanging the order of integration separates the double integral into a product. The exchange is justified because $\|f\|_{L^1} \, \|g\|_{L^1} < \infty$ ensures absolute convergence of the double integral via Tonelli's theorem, and the substitution $z = x - y$ together with the multiplicativity of the exponential $e^{-i(z+y)\cdot\xi} = e^{-iz\cdot\xi} \, e^{-iy\cdot\xi}$ factors the result into $\hat{f}(\xi) \, \hat{g}(\xi)$. The cleanness of this formula — no stray factors of $(2\pi)$ — is the principal reason many analysts prefer the convention adopted on this page. In the symmetric convention, the [Convolution Theorem](/theorems/250) reads $\widehat{f * g} = (2\pi)^{n/2} \hat{f} \hat{g}$, which is less elegant and more error-prone when iterated. ### Translation and Modulation Two further identities complete the basic Fourier calculus by describing the interaction with spatial and frequency shifts. [quotetheorem:251] These identities express a fundamental duality: translation in the spatial domain corresponds to multiplication by a complex exponential (phase shift) in the frequency domain, while multiplication by a complex exponential in the spatial domain corresponds to translation in the frequency domain. Both are verified by direct substitution in the definition — the translation identity uses the change of variables $z = x - a$, and the modulation identity uses the algebraic identity $e^{ia \cdot x} e^{-ix \cdot \xi} = e^{-ix \cdot (\xi - a)}$. Together with the differentiation and convolution rules, they form the core operational calculus of the Fourier transform. ## References 1. E. M. Stein and G. Weiss, *Introduction to Fourier Analysis on Euclidean Spaces* (1971). 2. L. Grafakos, *Classical Fourier Analysis*, 3rd ed. (2014). 3. E. M. Stein and R. Shakarchi, *Fourier Analysis: An Introduction* (2003).