# The [Fourier Transform](/page/Fourier%20Transform) The Fourier transform is the central tool of harmonic analysis, converting problems about derivatives in physical space into problems about polynomial multiplication in frequency space. This chapter develops the Fourier transform on the Schwartz class, extends it to $L^2(\mathbb{R}^n)$ via the Plancherel theorem, and uses it to construct a scale of [Sobolev spaces](/page/Inhomogeneous%20Sobolev%20Space) $H^s(\mathbb{R}^n)$ that measure smoothness by frequency decay. The resulting embedding theorems — translating control of derivatives into improved [integrability](/page/Integral) or [continuity](/page/Continuity) — form the functional-analytic backbone of modern PDE theory. The second part of the chapter turns to the complementary theme of oscillatory integrals, where the interplay between rapid phase oscillation and amplitude decay governs phenomena ranging from mixing in fluid dynamics to dispersive decay in [wave equations](/page/Wave%20Equation). The stationary phase principle, Van der Corput estimates, and Strichartz inequalities are developed as tools for extracting quantitative decay from these oscillations. ## The Schwartz Class test addition A natural first attempt at building a Fourier theory might begin with smooth, compactly supported [functions](/page/Function) $C_c^\infty(\mathbb{R}^n)$. However, the Fourier transform does not preserve compact support: if $f \in L^1(\mathbb{R})$ has compact support and $\hat{f}$ also has compact support, then $f \equiv 0$. This is because the function $z \mapsto \int e^{-ixz} f(x)\, d\mathcal{L}^1(x)$ extends to an entire function on $\mathbb{C}$, and if it vanishes on an open subset of the real line, the identity theorem forces it to vanish identically. The correct resolution is to work with a class of functions that trades compact support for rapid decay at infinity. The Schwartz class achieves exactly this: its elements are smooth and decay faster than any polynomial, together with all their derivatives. [citedefinition:Schwartz Semi-Norm] [citedefinition:Schwartz Space] The function $e^{-|x|^2}$ belongs to $\mathcal{S}(\mathbb{R}^n)$, but $e^{-|x|}$ does not, since it fails to be differentiable at the origin. The inclusion $C_c^\infty(\mathbb{R}^n) \subset \mathcal{S}(\mathbb{R}^n)$ holds, and $C_c^\infty$ is dense in $\mathcal{S}$ — this can be seen by multiplying a Schwartz function by a [sequence](/page/Sequence) of smooth cutoffs $\varphi(\cdot/m)$ with $\varphi|_{B(0,1)} \equiv 1$. An equivalent characterisation is often more convenient in practice: $f \in \mathcal{S}(\mathbb{R}^n)$ if and only if for all $N \in \mathbb{N}$ and all multi-indices $\alpha$ there exists $C_{\alpha, N} > 0$ such that $|\partial^\alpha f(x)| \le C_{\alpha, N}(1 + |x|)^{-N}$ for all $x \in \mathbb{R}^n$. Each $\rho_k$ is a seminorm, and the [Schwartz space](/page/Schwartz%20Space) equipped with the metric $\mathrm{d}(f,g) = \sum_{k \in \mathbb{N}} 2^{-k} \frac{\rho_k(f - g)}{1 + \rho_k(f - g)}$ is a complete, locally convex [metric space](/page/Metric%20Space). [citedefinition:Schwartz Topology] The Schwartz class enjoys closure under multiplication, convolution, [differentiation](/page/Derivative), and multiplication by polynomials. Crucially, every Schwartz function belongs to $L^p(\mathbb{R}^n)$ for every $1 \le p \le \infty$. [citedefinition:Fourier Transform] The key algebraic properties of the Fourier transform are that it interchanges differentiation and multiplication by monomials: \begin{align*} \widehat{\partial_{x_j} f}(\xi) &= i\xi_j \hat{f}(\xi), \\ \widehat{x_j f}(\xi) &= -i\partial_{\xi_j} \hat{f}(\xi). \end{align*} The first identity follows by [integration by parts](/theorems/210) (with [boundary](/page/Boundary) terms vanishing due to the rapid decay of $f$), and the second by differentiation under the integral sign (justified by dominated convergence). Additionally, the Fourier transform converts [convolutions](/page/Convolution) into products: $\widehat{f * g} = \hat{f} \hat{g}$, and satisfies the scaling relation $\widehat{f(\lambda \,\cdot\,)}(\xi) = \lambda^{-n} \hat{f}(\lambda^{-1}\xi)$ for $\lambda > 0$. [citedefinition:Convolution] The fact that the Fourier transform maps $\mathcal{S}(\mathbb{R}^n)$ to itself — and is in fact a [topological](/page/Topology) automorphism — is the foundational result that enables the entire theory. The differentiation-to-multiplication identities above show that if $f \in \mathcal{S}$, then $\xi^\alpha \hat{f}(\xi)$ and $\partial^\beta \hat{f}(\xi)$ are both bounded for all multi-indices, which is exactly the Schwartz condition on $\hat{f}$. [quotetheorem:228] The proof proceeds by showing that the seminorms $\rho_k(\hat{f})$ are controlled by finitely many seminorms of $f$, using the differentiation-multiplication exchange identities. The inverse is then constructed via the [Fourier inversion formula](/theorems/528), which we establish next. [citeproof:228] ## Fourier Inversion and the Plancherel Theorem The automorphism property tells us that $\mathcal{F}$ maps $\mathcal{S}$ bijectively onto itself, but it does not by itself give an explicit formula for the inverse. The inversion formula recovers $f$ from $\hat{f}$ by a second integration against the complex exponential, now with the opposite sign in the exponent. [citedefinition:Inverse Fourier Transform] The proof of the inversion formula relies on Gaussians. With the convention $\hat{f}(\xi) = \int f(x) e^{-ix \cdot \xi}\, d\mathcal{L}^n(x)$ (no prefactor), one computes that for any $a > 0$, \begin{align*} \mathcal{F}(e^{-a|x|^2})(\xi) = \left(\frac{\pi}{a}\right)^{n/2} e^{-|\xi|^2/(4a)}. \end{align*} This is verified by factoring the Gaussian into a product of one-dimensional Gaussians, reducing to the case $n = 1$ and $a = 1/2$. In that case, if $g(x) = e^{-x^2/2}$, the ODE $\hat{g}'(\xi) = -\xi \hat{g}(\xi)$ (obtained by integrating by parts) together with $\hat{g}(0) = \int e^{-x^2/2}\, d\mathcal{L}^1(x) = \sqrt{2\pi}$ gives $\hat{g}(\xi) = \sqrt{2\pi}\, e^{-\xi^2/2}$. [quotetheorem:644] The proof regularises the integral $\int \hat{f}(\xi)\, d\mathcal{L}^n(\xi)$ with a Gaussian factor $e^{-\varepsilon|\xi|^2}$, exploits the known Fourier transform of a Gaussian to evaluate the regularised identity at $x = 0$, and takes $\varepsilon \to 0$ via dominated convergence. The extension to general $x$ follows by applying the result to the translate $h(y) = f(x+y)$, whose Fourier transform is $e^{ix \cdot \xi}\hat{f}(\xi)$. [citeproof:644] The Fourier transform interacts naturally with the $L^2$ inner product. By Fubini's theorem, $\langle \mathcal{F}(f), g \rangle_{L^2} = \langle f, \mathcal{F}^{-1}(g) \rangle_{L^2}$, so the $L^2$ adjoint of $\mathcal{F}$ is $(2\pi)^n$ times its inverse. The rescaled operator $(2\pi)^{-n/2}\mathcal{F}$ is therefore unitary on $L^2(\mathbb{R}^n)$, and the Plancherel theorem makes this precise: $\|\hat{f}\|_{L^2}^2 = (2\pi)^n \|f\|_{L^2}^2$. [quotetheorem:247] The power of this result is that it provides an isometric identification (up to the universal constant $(2\pi)^{n/2}$) between a function and its frequency representation. Unlike the $L^1$ Fourier transform, which maps $L^1$ into $C_0$ but is not surjective, the $L^2$ theory is symmetric: a function and its Fourier transform live in the same space, and the rescaled operator $(2\pi)^{-n/2}\mathcal{F}$ is a unitary isomorphism. This symmetry is the reason the $L^2$ setting is particularly natural for PDE applications. The proof constructs the extension by verifying the norm identity on Schwartz functions and extending by density. [citeproof:247] ## The Fourier Transform on $L^2$ The Plancherel theorem opens the door to extending the Fourier transform beyond the Schwartz class. Since $\mathcal{S}(\mathbb{R}^n)$ is dense in $L^2(\mathbb{R}^n)$, and $(2\pi)^{-n/2}\mathcal{F}$ is an isometry on $\mathcal{S}$, $\mathcal{F}$ extends uniquely to a bounded operator $\mathcal{F}: L^2(\mathbb{R}^n) \to L^2(\mathbb{R}^n)$ satisfying $\|\hat{f}\|_{L^2} = (2\pi)^{n/2}\|f\|_{L^2}$. For $f \in L^2$, we choose an approximating sequence $(f_j) \subset \mathcal{S}$ with $f_j \to f$ in $L^2$; then $(\hat{f}_j)$ is Cauchy in $L^2$ (by Plancherel), and its [limit](/page/Limit) defines $\mathcal{F}_{L^2}(f)$. This construction is well-posed: the limit does not depend on the choice of approximating sequence. Indeed, if $(f_j)$ and $(g_j)$ both converge to $f$ in $L^2$, then $\|\hat{f}_j - \hat{g}_j\|_{L^2} = (2\pi)^{n/2}\|f_j - g_j\|_{L^2} \to 0$. The extended transform inherits the key properties of the Schwartz-level transform: the rescaled operator $(2\pi)^{-n/2}\mathcal{F}$ is a unitary isomorphism of $L^2(\mathbb{R}^n)$. By the Riesz–Thorin interpolation theorem, the Fourier transform also extends to a bounded operator $\mathcal{F}: L^p(\mathbb{R}^n) \to L^q(\mathbb{R}^n)$ for $1 \le p \le 2$ and $1/p + 1/q = 1$. However, these extensions are no longer isometric (even up to a constant) for $p \ne 2$, and the connection with derivatives becomes less direct — this is one of the reasons the $L^2$ setting is privileged. [example:Heat Equation Via Fourier Transform] The Fourier transform reduces the [heat equation](/page/Heat%20Equation) to an ODE in frequency space. Consider the Cauchy problem $\partial_t u - \Delta u = 0$ on $\mathbb{R}^n$ with initial data $u(0,x) = u_0(x) \in \mathcal{S}(\mathbb{R}^n)$. Taking the Fourier transform in $x$ and using $\mathcal{F}(\Delta u) = -|\xi|^2 \hat{u}$, the PDE becomes $\partial_t \hat{u}(t, \xi) + |\xi|^2 \hat{u}(t, \xi) = 0$, which is an ODE with solution $\hat{u}(t, \xi) = e^{-t|\xi|^2} \hat{u}_0(\xi)$. Inverting the Fourier transform and using the Gaussian identity gives \begin{align*} u(t, x) = (G_t * u_0)(x), \qquad G_t(x) = \frac{1}{(4\pi t)^{n/2}} \exp\!\left(-\frac{|x|^2}{4t}\right). \end{align*} The heat kernel $G_t$ satisfies $\int G_t\, d\mathcal{L}^n = 1$ (mass conservation), $G_t \in C^\infty$ for $t > 0$ (smoothing), $\|G_t\|_{L^\infty} = (4\pi t)^{-n/2}$ (decay), and the semigroup property $G_t * G_s = G_{t+s}$ (which follows from $e^{-t|\xi|^2} e^{-s|\xi|^2} = e^{-(t+s)|\xi|^2}$ on the Fourier side). [/example] ## Sobolev Spaces via the Fourier Transform The Fourier transform converts differentiation to multiplication, providing a natural and quantitative way to measure smoothness. If $f \in \mathcal{S}(\mathbb{R}^n)$, then $\widehat{\partial_{x_j} f}(\xi) = i\xi_j \hat{f}(\xi)$, and by Plancherel, $\|\nabla f\|_{L^2}^2 = (2\pi)^{-n}\||\xi| \hat{f}\|_{L^2}^2$. In particular, controlling all derivatives up to order $k$ in $L^2$ is equivalent to controlling $\|(1 + |\xi|^2)^{k/2} \hat{f}\|_{L^2}$. This observation motivates the definition of $H^s(\mathbb{R}^n)$, which extends to arbitrary real orders $s$ — not just integers. The [Inhomogeneous Sobolev Space](/page/Inhomogeneous%20Sobolev%20Space) page develops the full distributional definition: for $s \in \mathbb{R}$, the space $H^s(\mathbb{R}^n)$ consists of those tempered distributions $u \in \mathcal{S}'(\mathbb{R}^n)$ for which $(1 + |\xi|^2)^{s/2}\hat{u} = T_g$ for some $g \in L^2(\mathbb{R}^n)$, with norm $\|u\|_{H^s} := \|g\|_{L^2}$. Here $T_g$ denotes the [regular distribution](/page/Regular%20Distribution) associated to $g$. [citedefinition:Tempered Distribution] [citedefinition:Slowly Increasing Function] [citedefinition:Product Of A Tempered Distribution With A Slowly Increasing Function] [citedefinition:Inhomogeneous Sobolev Space] For the purposes of these notes, where we primarily work with functions in $\mathcal{S}(\mathbb{R}^n)$ or $L^2(\mathbb{R}^n)$, the Fourier transform $\hat{f}$ is itself a function and the norm takes the integral form \begin{align*} \|f\|_{H^s}^2 = \int_{\mathbb{R}^n} (1 + |\xi|^2)^{s} |\hat{f}(\xi)|^2\, d\mathcal{L}^n(\xi). \end{align*} The space $H^s(\mathbb{R}^n)$ is a [Hilbert space](/page/Hilbert%20Space) with inner product $\langle f, g \rangle_{H^s} := \int_{\mathbb{R}^n} (1 + |\xi|^2)^{s} \hat{f}(\xi) \overline{\hat{g}(\xi)}\, d\mathcal{L}^n(\xi)$, and the inclusions $C_c^\infty \subset \mathcal{S} \subset H^s$ are dense. For $s > t \ge 0$, the inclusion $H^s \hookrightarrow H^t$ is continuous and strict, with the interpolation inequality $\|f\|_{H^r} \le \|f\|_{H^t}^\theta \|f\|_{H^s}^{1-\theta}$ for $r = \theta t + (1-\theta)s$. When $s = k$ is a non-negative integer, the Fourier-based norm $\|f\|_{H^k}$ is equivalent to $(\sum_{|\alpha| \le k} \|\partial^\alpha f\|_{L^2}^2)^{1/2}$, recovering the classical definition via weak derivatives on the [Sobolev Space](/page/Sobolev%20Space) page. The [homogeneous Sobolev spaces](/page/Homogeneous%20Sobolev%20Space) capture the idea of having "exactly $s$ derivatives in $L^2$" without lower-order control. As developed on that page, the rigorous definition requires working in the quotient $\mathcal{S}'(\mathbb{R}^n)/\mathcal{P}$ ([tempered distributions](/page/Tempered%20Distributions) modulo polynomials), since the weight $|\xi|^s$ vanishes at the origin and cannot distinguish distributions differing by a polynomial. For $[f] \in \mathcal{S}'(\mathbb{R}^n)/\mathcal{P}$, the norm is \begin{align*} \|[f]\|_{\dot{H}^s} := \left(\int_{\mathbb{R}^n} |\xi|^{2s} |\hat{f}_{T\text{-rep}}(\xi)|^2\, d\mathcal{L}^n(\xi)\right)^{1/2}, \end{align*} where $\hat{f}_{T\text{-rep}}$ denotes the $T$-representative of $\hat{f}$ on $\mathbb{R}^n_0$ — the unique locally integrable function representing the [distributional](/page/Distribution) Fourier transform away from the origin. When $s < n/2$, the [functional realization theorem](/theorems/224) shows that the quotient is superfluous: $\hat{f}_{T\text{-rep}}$ extends to an $L^1_{\mathrm{loc}}$ function on all of $\mathbb{R}^n$, and $f$ is a well-defined tempered distribution rather than merely an equivalence class. A key feature of $\dot{H}^s(\mathbb{R}^n)$ is scaling invariance: defining $f_\lambda(x) = \lambda^{n/2 - s} f(\lambda x)$, one has $\|f_\lambda\|_{\dot{H}^s} = \|f\|_{\dot{H}^s}$ for all $\lambda > 0$. No such property holds for the inhomogeneous spaces $H^s$. ### Sobolev Embedding Into Continuous Functions The first embedding result asks: when does controlling $s$ derivatives in $L^2$ force the function to be bounded and continuous? The answer is $s > n/2$, which ensures that $\hat{f}$ is integrable — and therefore $f$ is bounded by the Fourier inversion formula. [quotetheorem:226] The proof is elementary: the Cauchy–Schwarz inequality gives $\|\hat{f}\|_{L^1} \le \|(1 + |\xi|^2)^{s/2} \hat{f}\|_{L^2} \cdot \|(1 + |\xi|^2)^{-s/2}\|_{L^2}$, and the second factor is finite precisely when $s > n/2$ (so that $(1 + |\xi|^2)^{-s}$ is integrable over $\mathbb{R}^n$). The Fourier inversion formula then gives $\|f\|_{L^\infty} \le (2\pi)^{-n} \|\hat{f}\|_{L^1} \le C_s \|f\|_{H^s}$. This is a clean result but it is not sharp: for $0 < s \le n/2$, one still gains integrability even though continuity may fail. [citeproof:226] ### Sobolev Embedding for Sub-Critical Exponents When $0 \le s < n/2$, controlling $s$ derivatives in $L^2$ does not give continuity, but it does improve the integrability of the function. The critical exponent $p = 2n/(n - 2s)$ is determined by scaling: the inequality $\|f\|_{L^p} \le C \||\nabla|^s f\|_{L^2}$ can only hold if both sides scale identically under $f \mapsto f(\lambda\,\cdot\,)$, which forces $-n/p = s - n/2$. [quotetheorem:225] The proof due to Chemin and Xu avoids hard analysis by decomposing $f$ into low- and high-frequency parts. For a threshold $N > 0$, the low-frequency part $f_{1,N} = \mathcal{F}^{-1}(\mathbb{1}_{B(0,N)} \hat{f})$ is bounded in $L^\infty$ by $C N^{n/2 - s} \|f\|_{\dot{H}^s}$ (using Cauchy–Schwarz), while the high-frequency part $f_{2,N}$ is controlled in $L^2$. Choosing $N$ to balance these contributions and integrating via the layer cake formula yields the sharp $L^p$ bound. This argument is particularly transparent because it reduces the Sobolev embedding to the interplay between Chebyshev's inequality and frequency localisation. [citeproof:225] ### The Gagliardo–Nirenberg–Sobolev Inequality The Sobolev embedding gives a sharp bound at the endpoint $q = 2n/(n-2s)$. For intermediate exponents $2 \le q \le 2n/(n-2s)$, a stronger interpolation inequality holds. [quotetheorem:634] The proof decomposes $f$ into low- and high-frequency parts at a threshold $N$, bounds the low-frequency part in $L^\infty$ using Cauchy–Schwarz and the high-frequency part in $L^2$ using the $\dot{H}^s$ norm. Choosing $N$ as a function of a level parameter $\lambda$ (via the layer cake representation) and integrating produces the interpolated $L^q$ bound. The exponent condition $1/q = 1/2 - \theta s/n$ ensures dimensional consistency. [citeproof:634] ### The Hardy–Littlewood–Sobolev Inequality The Sobolev embedding can also be deduced from a classical result on fractional integration, which controls the convolution with a power-law kernel. The key difficulty is that the kernel $|x|^{-\alpha}$ does not belong to any Lebesgue space $L^{n/\alpha}(\mathbb{R}^n)$ — the integral diverges both near the origin and at infinity — so the classical Young convolution inequality does not apply. Instead, one must work with the weak $L^p$ space $L^{n/\alpha,\infty}(\mathbb{R}^n)$ and the corresponding [weak Young inequality](/theorems/649). [citedefinition:Weak Lebesgue Space] [quotetheorem:469] The connection to the Sobolev embedding is through the Fourier-side identity $\mathcal{F}(|x|^{-\alpha})(\xi) = c_\alpha |\xi|^{\alpha - n}$: convolution with $|x|^{-\alpha}$ corresponds to multiplication by $|\xi|^{\alpha - n}$ in frequency, which is precisely the operation of "removing $\alpha$ derivatives." The special case $q = 2$ of the Hardy–Littlewood–Sobolev inequality can conversely be deduced from the [Sobolev embedding for homogeneous spaces](/theorems/225) via this Fourier correspondence. The proof itself proceeds by decomposing the kernel and the functions into dyadic level [sets](/page/Set) and applying a combinatorial trilinear estimate. [citeproof:469] # Oscillatory Integrals Oscillatory integrals of the form $I(\lambda) = \int_{\mathbb{R}^n} e^{i\lambda \varphi(y)} a(y)\, d\mathcal{L}^n(y)$ appear throughout analysis, from the asymptotic evaluation of special functions to the study of dispersive PDE. The fundamental principle is that when $\lambda$ is large, rapid oscillations of the integrand cause extensive cancellation, and the dominant contribution comes from neighbourhoods of **stationary points** — points where $\nabla \varphi = 0$. This chapter develops the tools needed to make this principle quantitative. ## The Stationary Phase Principle The simplest manifestation of the cancellation principle is the non-stationary phase lemma: if the phase gradient never vanishes on the support of the amplitude, the integral decays faster than any power of $\lambda$. [quotetheorem:635] The proof uses the differential operator $L = (i\lambda\varphi')^{-1} d/dy$ (in one dimension) satisfying $L(e^{i\lambda\varphi}) = e^{i\lambda\varphi}$. Each integration by parts transfers one derivative from the exponential onto the amplitude and produces a factor of $\lambda^{-1}$. The hypothesis $\nabla\varphi \ne 0$ ensures the denominator $|\varphi'|$ is bounded below on $\operatorname{supp}(a)$. Iterating $m$ times gives the $O(\lambda^{-m})$ bound. [citeproof:635] When the phase does have a stationary point, the leading-order contribution is determined by the Hessian at that point. The prototypical case is the quadratic phase $\varphi(y) = |y|^2$, for which \begin{align*} \left|\int_{\mathbb{R}^n} e^{i\lambda|y|^2} a(y)\, d\mathcal{L}^n(y)\right| \le C_a \lambda^{-n/2}. \end{align*} This is proved by decomposing the integral into a small ball of radius $R$ around the origin (estimated by $C\|a\|_{L^\infty} R^n$) and the complement (estimated by $C_a \lambda^{-N} R^{-2N+n}$ using integration by parts), then choosing $R = \lambda^{-1/2}$ to balance the two contributions. For a general non-degenerate stationary point, one obtains a full asymptotic expansion. [quotetheorem:645] The proof for the model case $\varphi(y) = |y|^2$ regularises the integral with a Gaussian factor $e^{-\varepsilon|y|^2}$, evaluates the resulting Gaussian integral via Parseval, and expands the residual phase $e^{|\xi|^2/(4i\lambda)}$ in a Taylor [series](/page/Series) to extract the leading term $(2i\lambda)^{-n/2}(2\pi)^{n}a(0)$. The general case follows from the Morse lemma, which provides a local diffeomorphism reducing the phase to its quadratic Taylor expansion near the critical point. [citeproof:645] ### The Van der Corput Lemma In one dimension, a sharper and more flexible tool is available: the Van der Corput lemma gives decay bounds for oscillatory integrals in terms of lower bounds on a single derivative of the phase, without requiring non-degeneracy of the Hessian. [quotetheorem:637] The proof proceeds by induction on $k$. The first step reduces the problem from an integral with amplitude $f$ to a pure exponential integral (without amplitude) via a single integration by parts. The base case $k = 1$ uses monotonicity of $\varphi'$ to control the total variation of $1/\varphi'$ after integration by parts. The inductive step splits the domain at the minimiser of $|\varphi^{(k)}|$ and balances the trivial estimate on a $\delta$-neighbourhood against the inductive estimate on the complement, with $\delta = \lambda^{-1/(k+1)}$ yielding the recursion $c_{k+1} = 2c_k + 2$. [citeproof:637] ## Mixing by Shear Flows Oscillatory integral techniques arise naturally in the study of transport equations, where phase oscillations produce effective mixing even in the absence of diffusion. ### The Transport Equation and Mixing Norms Consider a passive scalar $\theta(t,x)$ transported by an incompressible velocity field $u(x)$: \begin{align*} \partial_t \theta + u \cdot \nabla \theta = 0, \qquad \nabla \cdot u = 0. \end{align*} Although the $L^2$ norm of $\theta$ is conserved (multiply by $\theta$ and integrate, using $\nabla \cdot u = 0$), the scalar field can undergo significant homogenisation as the flow creates progressively finer spatial scales. The $\dot{H}^{-1}$ norm captures this transfer of energy to high frequencies: for a mean-zero function on $\mathbb{T}^n$, \begin{align*} \|\theta(t)\|_{\dot{H}^{-1}}^2 = \sum_{k \in \mathbb{Z}^n \setminus\{0\}} \frac{|\hat{\theta}(t,k)|^2}{|k|^2}. \end{align*} This norm weights high-frequency modes less heavily, so when the flow transfers energy to fine scales, the $\dot{H}^{-1}$ norm decays even though $\|\theta\|_{L^2}$ does not. ### The Couette Flow The simplest nontrivial shear flow is the Couette profile $u(x,y) = (y, 0)$ on $\mathbb{T} \times \mathbb{R}$. The transport equation becomes $\partial_t \theta + y \partial_x \theta = 0$, with explicit solution $\theta(t, x, y) = \theta_{\mathrm{in}}(x - yt, y)$. Taking the Fourier transform in both variables: \begin{align*} \hat{\theta}(t, k, \eta) = \hat{\theta}_{\mathrm{in}}(k, \eta + kt). \end{align*} The linear shear in physical space manifests as a translation $\eta \mapsto \eta + kt$ in frequency space: energy migrates to higher $\eta$-frequencies at a rate proportional to the wavenumber $k$. Computing the $\dot{H}^{-1}$ norm directly from this representation, one finds that for each nonzero mode $k$, the denominator $k^2 + \eta^2$ in the original norm becomes effectively $k^2 + (t + \eta/k)^2$ after the frequency shift. A case analysis shows \begin{align*} \frac{1}{(1 + (\eta/k)^2)(1 + (t + \eta/k)^2)} \le \frac{4}{1 + t^2} \end{align*} uniformly in $\eta$ and $k \ne 0$, yielding the mixing estimate $\|\theta(t)\|_{\dot{H}^{-1}} \le \frac{2}{\sqrt{1 + t^2}} \|\theta_{\mathrm{in}}\|_{\dot{H}^1}$. [remark:Couette Mixing Rate] The $O(1/t)$ mixing rate for Couette flow in $\dot{H}^{-1}$ is algebraic, not exponential — this is characteristic of shear-driven mixing. For the more general $\dot{H}^{-s}$ norm, the same argument yields $\|\theta(t)\|_{\dot{H}^{-s}} \lesssim (1 + t^2)^{-s/2} \|\theta_{\mathrm{in}}\|_{\dot{H}^s}$ for any $s > 0$. [/remark] ### General Shear Flows For a general shear flow $u(x,y) = (v(y), 0)$ on $\mathbb{T} \times (0,1)$, we can still take the partial Fourier transform in $x$ to obtain $\hat{\theta}(t, k, y) = e^{-ikv(y)t} \hat{\theta}_{\mathrm{in}}(k, y)$. For fixed $k \ne 0$, the $H^{-1}_y$ norm of $\hat{\theta}(t, k, \cdot)$ involves the oscillatory integral $\int_0^1 e^{-ikv(y)t} \hat{\theta}_{\mathrm{in}}(k, y) \phi(y)\, d\mathcal{L}^1(y)$, with phase $\varphi(y) = v(y)$ and parameter $\tau = kt$. The Van der Corput lemma controls this integral. [quotetheorem:638] The proof reduces the $\dot{H}^{-1}$ estimate to one-dimensional oscillatory integrals via duality: the partial Fourier transform in $x$ decouples the modes, and for each nonzero mode $k$, the $H^{-1}_y$ norm is expressed as a supremum of oscillatory integrals with phase $v(y)$ and parameter $kt$. A [partition of unity](/page/Partition%20of%20Unity) subordinate to the non-degeneracy condition localises each integral to a region where a specific derivative $v^{(j)}$ is bounded below, and the [Van der Corput lemma](/theorems/637) then controls each piece. [citeproof:638] ## Linear Dispersive Equations Dispersion is the phenomenon whereby different frequency components of a wave travel at different speeds, causing an initially localised wave packet to spread out over time. Unlike mixing, which arises from advection, dispersion is an intrinsic property of the linear evolution operator. Mathematically, a linear dispersive equation takes the form $\partial_t u = ih(D)u$, where $h: \mathbb{R}^n \to \mathbb{R}$ is the dispersion relation and $h(D)$ denotes the Fourier multiplier with symbol $h(\xi)$. Taking the Fourier transform gives $\hat{u}(t, \xi) = e^{ith(\xi)} \hat{u}_{\mathrm{in}}(\xi)$. By Fourier inversion, the solution in physical space is the oscillatory integral \begin{align*} u(t, x) = \frac{1}{(2\pi)^n} \int_{\mathbb{R}^n} e^{i(x \cdot \xi + th(\xi))} \hat{u}_{\mathrm{in}}(\xi)\, d\mathcal{L}^n(\xi), \end{align*} and the stationary phase principle governs its large-time behaviour. The [group](/page/Group) velocity $\nabla h(\xi)$ determines which frequency reaches a given spatial location at time $t$: the main contribution at $(t, x)$ comes from frequencies $\xi$ satisfying $x/t \approx -\nabla h(\xi)$. ### The Linear Schrödinger Equation The free Schrödinger equation $i\partial_t u + \Delta u = 0$ has dispersion relation $h(\xi) = -|\xi|^2$, giving \begin{align*} u(t, x) = \frac{1}{(4\pi it)^{n/2}} \int_{\mathbb{R}^n} e^{i|x-y|^2/(4t)} u_{\mathrm{in}}(y)\, d\mathcal{L}^n(y). \end{align*} Since $\nabla^2 h = -2\operatorname{Id}$ is non-degenerate everywhere, every frequency contributes a stationary point, and the [stationary phase lemma](/theorems/636) gives the asymptotic expansion $u(t, x) \approx (C_n / t^{n/2}) \hat{u}_{\mathrm{in}}(x/(2t)) e^{i|x|^2/(4t)}$ as $t \to \infty$. The fundamental dispersive estimate is stated and proved in the following result. [quotetheorem:930] The technique is elementary: the explicit kernel has a complex Gaussian integrand whose modulus is identically $1$, so the $L^\infty$ bound reduces to the prefactor $(4\pi|t|)^{-d/2}$ times $\|u_{\mathrm{in}}\|_{L^1}$. The full proof is on the [theorem page](/theorems/930). Interpolating with the $L^2$ conservation law $\|e^{it\Delta} u_{\mathrm{in}}\|_{L^2} = \|u_{\mathrm{in}}\|_{L^2}$ via Riesz–Thorin gives the family of dispersive decay estimates \begin{align*} \|e^{it\Delta} u_{\mathrm{in}}\|_{L^q} \le C_p |t|^{-\frac{n}{2}\left(\frac{1}{p} - \frac{1}{q}\right)} \|u_{\mathrm{in}}\|_{L^p}, \qquad \frac{1}{p} + \frac{1}{q} = 1, \quad 1 \le p \le 2. \end{align*} [remark:Dispersive Decay Versus Energy Conservation] The $L^\infty$ decay is not a contradiction with $L^2$ conservation: dispersion does not destroy energy but spreads it over an ever-increasing spatial region. The total $L^2$ mass is constant, but the amplitude at any fixed point decays. [/remark] ### The Linear KdV Equation The Korteweg–de Vries equation $\partial_t v + \partial_x^3 v = 0$ has dispersion relation $h(\xi) = \xi^3$. The propagator kernel involves the Airy function $\operatorname{Ai}(x) = \frac{1}{2\pi} \int_\mathbb{R} e^{i(x\xi + \xi^3)}\, d\mathcal{L}^1(\xi)$, giving \begin{align*} e^{-t\partial_x^3} f = \frac{1}{t^{1/3}} \operatorname{Ai}\!\left(\frac{\cdot}{t^{1/3}}\right) * f. \end{align*} Since $h''(\xi) = 6\xi$ vanishes at $\xi = 0$, the phase has a degenerate critical point. The Van der Corput lemma with $k = 2$ gives only $\lambda^{-1/2}$ decay — but this is at a single frequency. The net decay rate is $t^{-1/3}$, reflecting the fact that only one dimension's worth of dispersion is available. The precise statement and proof are given by the following result. [quotetheorem:931] The proof combines three techniques: frequency decomposition (low frequencies are trivially bounded), integration by parts for non-stationary regions (where $\varphi_x' \ne 0$), and the [Van der Corput lemma](/theorems/637) with $k = 2$ for the near-stationary region. The condition $\beta \le 1/2$ arises from balancing the amplitude growth $|\xi|^\beta \sim |x|^{\beta/2}$ against the Van der Corput decay $|\varphi_x''|^{-1/2} \sim |x|^{-1/4}$. The full proof is on the [theorem page](/theorems/931). ### The Linear Wave Equation The wave equation $\partial_t^2 u - \Delta u = 0$ factors as $(\partial_t + i|\nabla|)(\partial_t - i|\nabla|)u = 0$, giving two half-wave propagators $e^{\pm it|\nabla|}$ with dispersion relation $h(\xi) = \pm|\xi|$. A key difference from Schrödinger is that $h$ is homogeneous of degree $1$, so the Hessian $\nabla^2 h$ has rank $n - 1$ (it vanishes in the radial direction). The $L^\infty$ decay rate is therefore $t^{-(n-1)/2}$ rather than $t^{-n/2}$: \begin{align*} \|e^{\pm it|\nabla|} f\|_{L^\infty} \le C \langle t \rangle^{-(n-1)/2} \|f\|_{L^1}, \end{align*} for data $f$ with Fourier support in $\{1/2 \le |\xi| \le 2\}$. In $n = 1$, d'Alembert's formula $u(t,x) = \frac{1}{2}(f(x+t) + f(x-t)) + \frac{1}{2}\int_{x-t}^{x+t} g(y)\, d\mathcal{L}^1(y)$ shows no amplitude decay at all — the initial data are simply transported. ## Strichartz Estimates The dispersive decay estimates established above give pointwise-in-time control of spatial $L^q$ norms. A natural question is whether these decay rates yield time-integrated bounds: does the $L^q_x$ norm decay in a time-integrable fashion? The answer, provided by Strichartz estimates, is affirmative, and these estimates form the foundation of the modern theory of nonlinear dispersive PDE. ### The Abstract Framework Let $(U(t))_{t \in \mathbb{R}}$ be a family of operators satisfying the unitarity and dispersive assumptions \begin{align*} \|U(t)f\|_{L^2} &= \|f\|_{L^2}, \\ \|U(t)U^*(s)f\|_{L^\infty} &\le C_0 |t - s|^{-\sigma} \|f\|_{L^1}, \end{align*} for some $\sigma > 0$. The Schrödinger propagator $U(t) = e^{it\Delta}$ satisfies these with $\sigma = n/2$. A pair $(q, r)$ with $2 < q, r \le \infty$ is called **admissible** if it satisfies the scaling condition $1/q + \sigma/r = \sigma/2$. This condition is dictated by dimensional analysis: rescaling $t \mapsto \lambda t$ corresponds to $d\mathcal{L}^n(x) \mapsto \lambda^{-\sigma} d\mathcal{L}^n(x)$, and the mixed norm $L^q_t L^r_x$ is scale-invariant only when the admissibility condition holds. [quotetheorem:639] The key mechanism is the $TT^*$ method: the homogeneous estimate is equivalent to its dual by unitarity, and the dual estimate is proved by expanding the squared $L^2$ norm as a double time integral, bounding the integrand using the dispersive decay, and applying the [Hardy–Littlewood–Sobolev inequality](/theorems/469) in the time variable. The admissibility condition ensures the HLS exponents match. The inhomogeneous estimate then follows by composing the homogeneous and dual estimates with different admissible pairs. [citeproof:639] For the Schrödinger equation specifically, an admissible pair $(q, r)$ satisfies $2/q + n/r = n/2$. The Strichartz estimates and Duhamel's formula combine to give control of solutions to the inhomogeneous problem. [quotetheorem:640] The proof combines Duhamel's formula with the [abstract Strichartz estimates](/theorems/639): the homogeneous part $e^{it\Delta}u_{\mathrm{in}}$ is controlled directly by part (1), while the inhomogeneous Duhamel integral $\int_0^t e^{i(t-s)\Delta}F(s)\, d\mathcal{L}^1(s)$ is controlled by part (3) applied to the truncated forcing $\chi_{[0,t]}F$. The fact that the source pair $(a,b)$ and the solution pair $(q,r)$ can be chosen independently is what gives the estimate its flexibility in nonlinear applications. [citeproof:640] [remark:Significance Of Strichartz Estimates] Strichartz estimates demonstrate a fundamental principle: space-time integrability is obtained by combining the oscillatory decay of the propagator with fractional integration in time. For nonlinear dispersive PDE, they provide the basic framework for fixed-point arguments — one places the nonlinearity in a dual Strichartz space and recovers the solution in a primal Strichartz space, closing the contraction via the inhomogeneous estimate. [/remark]