Dispersive Decay Estimates for the Schrödinger and Wave Propagators

Intuition Beginner

A wave packet spreads out as it moves. Toss a pebble into a still pond and the ripple does not stay a tight ring; it broadens, and the bump of water at any one place gets lower and lower as the disturbance smears across a wider and wider circle. The total amount of disturbance is conserved, but it is spread thinner everywhere. This spreading-and-fading is called dispersion, and the rate at which the peak height drops is what a dispersive decay estimate measures.

The reason the peak fades is that different frequencies travel at different speeds. A sharp initial bump is secretly a blend of many pure waves of many different wavelengths. If every wave moved at the same speed, the bump would keep its shape and just slide along — that is what an ordinary sound wave does in air. But for the equations of this unit, short-wavelength ripples outrun long-wavelength ones, so the blend that made a tall narrow bump comes apart. The pieces march off at their own speeds, the constructive pile-up that made the peak disappears, and what is left at any fixed point is small.

There is an exact rate to this fading, and it depends only on the number of space dimensions you live in. For the Schrödinger equation — the equation of quantum mechanics, which governs how the probability cloud of a free particle evolves — an initial cloud packed into a small region thins out so that its tallest value at time $t$ is no bigger than a fixed constant times $t$ raised to the power minus $n/2$, where $n$ is the dimension. In one dimension the peak fades like one over the square root of time; in three dimensions like one over time-and-a-half. More room to spread into means faster fading.

The same story holds for waves that obey the wave equation, like light or sound, but with a twist. There the spreading happens along an expanding sphere rather than throughout a solid ball, and the geometry of that sphere — the fact that it is curved — both helps the fading and costs you something. The curved wavefront focuses and defocuses the frequencies in a way that gives a slightly weaker decay rate and forces you to start with a slightly smoother bump than for Schrödinger. The careful accounting of this trade-off is the heart of the wave case.

The one-sentence takeaway: dispersive equations spread a concentrated disturbance out over a growing region, the peak height fades at a definite rate set by the dimension, and these decay rates are the raw fuel that later lets us solve hard nonlinear versions of the same equations.

Visual Beginner

Picture a single tall narrow spike of water at the centre of a calm pond at the starting moment. As time runs forward, watch the spike at three later snapshots. In the first, it has slumped into a lower, wider bump with a faint ring around it. In the second, the bump is lower still and the ring has expanded. In the third, there is no real bump left at the centre at all, only a broad shallow swell with gentle ripples spread across a wide region. The water is still all there; it is just nowhere piled up.

The single quantitative fact the picture encodes is the shrinking of the central peak. If you measured the height at the very centre at each snapshot and plotted those heights against time, you would get a curve that falls off like one over the square root of time in one dimension, or one over time in two dimensions, or one over time-to-the-three-halves in three dimensions. The higher the dimension, the steeper the fall, because there is more surrounding space for the disturbance to leak into.

A second picture clarifies why the wave case differs from the Schrödinger case. For the wave equation, the disturbance does not fill a solid bump; it lives on the surface of an expanding sphere, a thin shell racing outward. The shell gets thinner and its total area grows, so the disturbance on it also fades — but because the shell is a curved surface rather than a filled-in region, the bookkeeping of how fast it fades is governed by the curvature of that sphere, which is the geometric quantity that controls the whole wave story.

Worked example Beginner

We track the fading peak of a spreading Gaussian bump under the one-dimensional Schrödinger evolution and read off the decay rate by hand. The setup uses the known fact that the Schrödinger equation turns a Gaussian into a wider, lower Gaussian as time runs, with a controllable width.

Step 1. Start with a normalized bump. At time zero take the bump to have height one and a width measured by a spread parameter equal to one. As the Schrödinger evolution runs, the bump stays a bump but its effective width grows. A direct calculation with the Gaussian (the same self-transforming Gaussian from the Fourier-transform unit) shows that at time $t$ the width has grown from one to the square root of one plus four times $t$ squared.

Step 2. Use conservation to find the new height. The Schrödinger evolution conserves total probability, which for a bump is roughly height times width. Since the starting height-times-width was one times one, equal to one, and width has grown to the square root of one plus four $t$ squared, the height must have shrunk to one over that same square root to keep the product equal to one.

Step 3. Read off the height at a few times. At time zero the height is one over the square root of one, which is one. At time one the height is one over the square root of one plus four, which is one over the square root of five, about $0.45$. At time five the height is one over the square root of one plus one hundred, which is one over the square root of one hundred and one, about $0.0995$.

Step 4. Find the large-time rate. For large $t$ the one inside the square root is negligible next to four $t$ squared, so the width is close to two $t$ and the height is close to one over two $t$. This particular Gaussian spreads so that its width grows like $t$ to the first power, so its own peak falls like one over $t$. The general sharp ceiling for an arbitrary integrable bump in one dimension is the slower rate one over $t$ to the power one-half; the Gaussian beats that ceiling because it starts perfectly smooth, and extra smoothness buys faster decay.

Step 5. State the comparison cleanly. The general dispersive estimate guarantees that any one-dimensional bump that starts with total size one (measured as an area, not a height) has peak height at most a fixed constant over the square root of $t$. Our smooth Gaussian does even better, one over $t$, because smoothness buys extra decay. The estimate is a worst-case ceiling that every bump obeys, and many bumps beat it.

What this tells us: the peak of a spreading packet falls at a rate the dimension dictates, the universal ceiling in one dimension is one over the square root of time, and extra smoothness of the starting bump only makes the real decay faster than the ceiling, never slower.

Check your understanding Beginner

Formal definition Intermediate+

Fix $n\ge 1$ and work on ${\mathbb{R}}^n$ with the Fourier transform of 02.10.04, normalized as $\hat{f}(\xi )={\int }_{{\mathbb{R}}^n}f(x)e^{-2\pi i\xi \cdot x}dx$.

Definition (free Schrödinger propagator). The free Schrödinger evolution of initial datum $f$ is the solution $u(\cdot ,t)=e^{it\Delta }f$ of $i{\partial }_{t}u+\Delta u=0$ with $u(\cdot ,0)=f$, defined on the Fourier side by the multiplier $$\widehat{e^{it\Delta }f}(\xi )=e^{-4{\pi }^{2}it|\xi {|}^2}\hat{f}(\xi ).$$ Since the multiplier $e^{-4{\pi }^{2}it|\xi {|}^2}$ has modulus one, Plancherel 02.10.04 shows $e^{it\Delta }$ is a unitary operator on $L^2({\mathbb{R}}^n)$ for each $t$, and $\{e^{it\Delta }{\}}_{t\in \mathbb{R}}$ is a one-parameter unitary group: $\Vert e^{it\Delta }f{\Vert }_{L^2}=\Vert f{\Vert }_{L^2}$ for all $t$ ^{[Tao 2006]}. This is the $L^2$ conservation law.

Definition (Schrödinger kernel). For $t\ne 0$ and $f\in \mathcal{S}({\mathbb{R}}^n)$, the evolution is convolution against an explicit oscillatory Gaussian kernel: $$\left(e^{it\Delta }f\right)(x)=\frac{1}{(4\pi it{)}^{n/2}}{\int }_{{\mathbb{R}}^n}{e}^{i|x-y{|}^2/(4t)}f(y)dy,$$ where $(4\pi it{)}^{n/2}=(4\pi |t|{)}^{n/2}{e}^{i\pi n\mathrm{sgn}(t)/4}$ is the principal branch ^{[Tao 2006]}. The phase factor $e^{i\pi n\mathrm{sgn}(t)/4}$ is the Maslov phase, the boundary value of the analytically continued Gaussian.

Definition (dispersive estimate, Schrödinger). The pointwise dispersive estimate is the $L^1\to L^{\infty }$ bound $$\Vert e^{it\Delta }f{\Vert }_{L^{\infty }({\mathbb{R}}^n)}\le \frac{1}{(4\pi |t|{)}^{n/2}}\Vert f{\Vert }_{L^1({\mathbb{R}}^n)},t\ne 0.$$ It is immediate from the kernel: the integrand is bounded in modulus by $(4\pi |t|{)}^{-n/2}|f(y)|$.

Definition (half-wave and Klein-Gordon propagators). The wave equation ${\partial }_t^{2}u-\Delta u=0$ factors through the half-wave propagators $e^{\pm it\sqrt{-\Delta }}$, defined by the Fourier multiplier $e^{\pm 2\pi it|\xi |}$. The Klein-Gordon propagator of mass $m>0$ uses the multiplier $e^{\pm 2\pi it⟨\xi {⟩}_m}$ with $⟨\xi {⟩}_m=\sqrt{|\xi {|}^2+m^2}$. The associated dispersion relations are $$\tau =|\xi |(\text{wave}),\tau =⟨\xi {⟩}_m(\text{Klein-Gordon}),\tau =4{\pi }^2|\xi {|}^2(\text{Schro¨dinger}),$$ and the group velocity $v={\nabla }_{\xi }\tau$ is the speed at which a wave packet centred on frequency $\xi$ travels: $v=\xi /|\xi |$ (unit speed, every direction) for the wave equation, and $v=8{\pi }^2\xi$ (unbounded, frequency-proportional) for Schrödinger.

Definition ($L^{p^{\prime }}\to L^p$ dispersive bound). Interpolating the $L^1\to L^{\infty }$ bound against the $L^2\to L^2$ conservation law via Riesz-Thorin gives, for $2\le p\le \infty$ and conjugate exponent $p^{\prime }$ with $1/p+1/p^{\prime }=1$, $$\Vert e^{it\Delta }f{\Vert }_{L^p({\mathbb{R}}^n)}\le \frac{C}{|t{|}^{n(1/2-1/p)}}\Vert f{\Vert }_{L^{p^{\prime }}({\mathbb{R}}^n)}.$$

Counterexamples to common slips Intermediate+

• Dispersive decay is not energy decay. The estimate $\Vert e^{it\Delta }f{\Vert }_{L^{\infty }}\lesssim |t{|}^{-n/2}\Vert f{\Vert }_{L^1}$ describes the fading of the sup norm, not the $L^2$ norm. The $L^2$ norm is exactly conserved for all time. Nothing is lost; mass moves to spatial infinity and thins out, lowering the pointwise maximum while preserving the total $L^2$ mass.

• The wave propagator has no $L^1\to L^{\infty }$ bound at the same rate. One might expect $\Vert e^{it\sqrt{-\Delta }}f{\Vert }_{L^{\infty }}\lesssim |t{|}^{-(n-1)/2}\Vert f{\Vert }_{L^1}$, but this is false. The light cone $\{\tau =|\xi |\}$ has one vanishing principal curvature (the radial direction), so the stationary-phase gain is only $(n-1)/2$ powers of $t$, and the bound requires the right-hand side to carry derivatives: $\Vert e^{it\sqrt{-\Delta }}f{\Vert }_{L^{\infty }}\lesssim |t{|}^{-(n-1)/2}\Vert f{\Vert }_{{\dot{W}}^{(n+1)/2,1}}$. The loss of $(n+1)/2$ derivatives is the price of the cone's curvature degeneracy.

• Group velocity is not phase velocity. The phase velocity $\tau /|\xi |$ and the group velocity ${\nabla }_{\xi }\tau$ agree only for the non-dispersive wave equation. For Schrödinger the phase velocity $4{\pi }^2|\xi |$ is half the group velocity $8{\pi }^2\xi$ in magnitude; energy travels at the group velocity, and it is the variation of the group velocity across frequencies (the non-degeneracy of the Hessian ${\nabla }_{\xi }^2\tau$) that drives dispersion. A linear dispersion relation has constant group velocity and no dispersive decay — that is the transport equation.

• The kernel is not a function for $t\to 0$. As $t\to 0$ the Schrödinger kernel $(4\pi it{)}^{-n/2}{e}^{i|x{|}^2/4t}$ does not converge to a Gaussian or to anything integrable; it converges to the Dirac delta in the distributional sense, with wild oscillation. The estimate degenerates ($|t{|}^{-n/2}\to \infty$) precisely because at $t=0$ a delta-like datum has infinite sup norm. The dispersive estimate is a statement about $t$ bounded away from zero.

Key theorem with proof Intermediate+

Theorem (pointwise dispersive estimate for Schrödinger). For every $n\ge 1$, every $t\ne 0$, and every $f\in L^1({\mathbb{R}}^n)$, the free Schrödinger evolution satisfies $$\Vert e^{it\Delta }f{\Vert }_{L^{\infty }({\mathbb{R}}^n)}\le \frac{1}{(4\pi |t|{)}^{n/2}}\Vert f{\Vert }_{L^1({\mathbb{R}}^n)},$$ and consequently, by interpolation against $\Vert e^{it\Delta }f{\Vert }_{L^2}=\Vert f{\Vert }_{L^2}$, for $2\le p\le \infty$, $$\Vert e^{it\Delta }f{\Vert }_{L^p}\le (4\pi |t|{)}^{-n(1/2-1/p)}\Vert f{\Vert }_{L^{p^{\prime }}}.$$

Proof. The argument has three steps: identify the propagator with the explicit kernel by analytic continuation of the Gaussian, read off the $L^1\to L^{\infty }$ bound, and interpolate.

Step 1 (the kernel via analytic continuation). For complex $z$ with $\mathrm{Re}z>0$, the heat-type evolution $e^{-z\Delta }$ has Fourier multiplier $e^{-4{\pi }^{2}z|\xi {|}^2}$ and, by the Gaussian Fourier-transform identity of 02.13.03 and 02.10.04, $$\left(e^{-z\Delta }f\right)(x)=\frac{1}{(4\pi z{)}^{n/2}}{\int }_{{\mathbb{R}}^n}{e}^{-|x-y{|}^2/(4z)}f(y)dy,$$ where $(4\pi z{)}^{n/2}$ uses the principal branch of the square root on the right half-plane. Both sides are holomorphic in $z$ on $\{\mathrm{Re}z>0\}$ for fixed $f\in \mathcal{S}$ and fixed $x$: the left side because $e^{-4{\pi }^{2}z|\xi {|}^2}\hat{f}(\xi )$ is holomorphic in $z$ with an integrable $\xi$-majorant on compact subsets, the right side because the Gaussian integrand is holomorphic with the same kind of majorant. By the identity theorem they agree throughout the right half-plane.

Now let $z\to it$ from within the right half-plane, $z=\epsilon +it$ with $\epsilon \downarrow 0$. On the Fourier side $e^{-4{\pi }^2(\epsilon +it)|\xi {|}^2}\hat{f}(\xi )\to e^{-4{\pi }^{2}it|\xi {|}^2}\hat{f}(\xi )$ in $L^2$, so the left side converges to $e^{it\Delta }f$. On the spatial side the principal branch gives $(4\pi (\epsilon +it){)}^{n/2}\to (4\pi it{)}^{n/2}=(4\pi |t|{)}^{n/2}{e}^{i\pi n\mathrm{sgn}(t)/4}$, and the Gaussian $e^{-|x-y{|}^2/(4(\epsilon +it))}\to e^{i|x-y{|}^2/(4t)}$ pointwise with modulus bounded by one for the boundary value (the real part of $1/(4(\epsilon +it))$ is non-negative). For $f\in \mathcal{S}$, dominated convergence gives $$\left(e^{it\Delta }f\right)(x)=\frac{1}{(4\pi it{)}^{n/2}}{\int }_{{\mathbb{R}}^n}{e}^{i|x-y{|}^2/(4t)}f(y)dy.$$

Step 2 (the $L^1\to L^{\infty }$ bound). The kernel $K_t(x-y)=(4\pi it{)}^{-n/2}{e}^{i|x-y{|}^2/(4t)}$ has modulus $|K_t|=(4\pi |t|{)}^{-n/2}$, a constant. Hence for every $x$, $$|(e^{it\Delta }f)(x)|\le \frac{1}{(4\pi |t|{)}^{n/2}}{\int }_{{\mathbb{R}}^n}|f(y)|dy=\frac{1}{(4\pi |t|{)}^{n/2}}\Vert f{\Vert }_{L^1}.$$ Taking the supremum over $x$ gives $\Vert e^{it\Delta }f{\Vert }_{L^{\infty }}\le (4\pi |t|{)}^{-n/2}\Vert f{\Vert }_{L^1}$ for $f\in \mathcal{S}$, and by density (Schwartz functions are dense in $L^1$, and the right side is continuous in the $L^1$ norm) for all $f\in L^1$.

Step 3 (interpolation). The operator $T_t=e^{it\Delta }$ is bounded $L^1\to L^{\infty }$ with norm at most $(4\pi |t|{)}^{-n/2}$ (Step 2) and bounded $L^2\to L^2$ with norm exactly $1$ (Plancherel). By the Riesz-Thorin interpolation theorem, for $\theta \in [0,1]$, $T_t$ maps $L^{p_{\theta }^{\prime }}\to L^{p_{\theta }}$ where $1/p_{\theta }=(1-\theta )/2+\theta /\infty =(1-\theta )/2$ and $1/p_{\theta }^{\prime }=(1-\theta )/2+\theta /1=(1+\theta )/2$. Then $1/p_{\theta }^{\prime }-1/2=\theta /2$, so $1/2-1/p_{\theta }=\theta /2$, and the operator-norm bound is the geometric mean $$\Vert T_t{\Vert }_{L^{p_{\theta }^{\prime }}\to L^{p_{\theta }}}\le 1^{1-\theta }\cdot ((4\pi |t|{)}^{-n/2}{)}^{\theta }=(4\pi |t|{)}^{-n\theta /2}=(4\pi |t|{)}^{-n(1/2-1/p_{\theta })}.$$ As $\theta$ ranges over $[0,1]$, $p_{\theta }$ ranges over $[2,\infty ]$, giving the stated family of bounds. $\square$

Bridge. The dispersive estimate is the foundational reason the entire nonlinear theory of 02.21.02 and beyond works: it converts the abstract unitarity of the propagator into a quantitative spreading rate, and that rate is exactly what a contraction-mapping argument needs to close. The mechanism — explicit oscillatory kernel of constant modulus, plus interpolation against an $L^2$ isometry — is dual to the heat-kernel mechanism of 02.13.03, where a real Gaussian of decaying modulus gives smoothing instead of dispersion; the analytic continuation $z=\epsilon \to it$ is precisely the bridge between the two. Putting these together, the Schrödinger result builds toward the Strichartz estimates via the $T{T}^{\ast }$ method, and the same estimate appears again in the wave case once stationary phase replaces the exact Gaussian, the central insight being that constant kernel modulus (the genuine Gaussian) becomes algebraic kernel decay (oscillatory integral over a curved cone) and the loss of derivatives measures that curvature.

Exercises Intermediate+

Advanced results Master

The wave case replaces the exact-Gaussian computation by genuine oscillatory-integral asymptotics over the light cone, whose curvature is degenerate in the radial direction. The systematic tool is the principle of stationary phase with non-degenerate Hessian, sharpened to handle the one vanishing curvature of the cone.

Theorem (stationary phase, non-degenerate). Let $\varphi \in C^{\infty }({\mathbb{R}}^n)$ have a single critical point ${\xi }_{\ast }$ on $\mathrm{supp}a$, $a\in C_c^{\infty }$, with non-degenerate Hessian $H={\nabla }^2\varphi ({\xi }_{\ast })$. Then for $\lambda \to +\infty$, $${\int }_{{\mathbb{R}}^n}a(\xi )e^{i\lambda \varphi (\xi )}d\xi ={\left(\frac{2\pi }{\lambda }\right)}^{n/2}\frac{e^{i\frac{\pi }{4}\mathrm{sgn}H}}{|\det H{|}^{1/2}}{e}^{i\lambda \varphi ({\xi }_{\ast })}a({\xi }_{\ast })+O({\lambda }^{-n/2-1}),$$ with $\mathrm{sgn}H$ the signature of $H$ ^{[Hörmander 1990] [Stein 1993]}. The leading ${\lambda }^{-n/2}$ is the source of the Schrödinger $|t{|}^{-n/2}$; the Maslov phase $e^{i\pi \mathrm{sgn}H/4}$ recovers the $(4\pi it{)}^{n/2}$ branch factor.

Theorem (oscillatory integral over a curved hypersurface). Let $S\subset {\mathbb{R}}^n$ be a smooth compact hypersurface with non-vanishing Gaussian curvature, $d\sigma$ its surface measure, and $\psi \in C_c^{\infty }(S)$. The Fourier transform of the curved surface measure decays at the surface-restriction rate: $$|\widehat{\psi d\sigma }(\eta )|=\left|{\int }_S\psi (\omega )e^{-2\pi i\eta \cdot \omega }d\sigma (\omega )\right|\le C(1+|\eta |{)}^{-(n-1)/2}.$$ The exponent $(n-1)/2$ counts the $n-1$ non-vanishing principal curvatures of the hypersurface; this is the Stein-Tomas surface-decay phenomenon ^{[Stein 1993]}.

Theorem (dispersive estimate for the half-wave propagator). For $n\ge 2$ and $t\ne 0$, $$\Vert e^{it\sqrt{-\Delta }}f{\Vert }_{L^{\infty }({\mathbb{R}}^n)}\le C|t{|}^{-(n-1)/2}\Vert f{\Vert }_{{\dot{W}}^{(n+1)/2,1}({\mathbb{R}}^n)},$$ where ${\dot{W}}^{s,1}$ is the homogeneous $L^1$-Sobolev space of order $s$ ^{[Brenner 1975] [Sogge 2008]}. The decay rate $(n-1)/2$ is governed by the $n-1$ non-zero principal curvatures of the cone $\{\tau =|\xi |\}$; the loss of $(n+1)/2$ derivatives on the right-hand side accounts for the single vanishing radial curvature and the homogeneity of the cone. The Klein-Gordon propagator $e^{it⟨\nabla {⟩}_m}$ interpolates: at low frequency the hyperboloid $\{\tau =⟨\xi {⟩}_m\}$ has full non-degenerate curvature and decays like $|t{|}^{-n/2}$ (Schrödinger-like), while at high frequency it asymptotes to the cone and decays like $|t{|}^{-(n-1)/2}$ (wave-like).

Theorem (Strichartz estimate, Schrödinger). With the admissibility condition $2/q+n/r=n/2$, $2\le q,r\le \infty$, $(q,r,n)\ne (2,\infty ,2)$, $$\Vert e^{it\Delta }f{\Vert }_{L_t^q{L}_x^r(\mathbb{R}\times {\mathbb{R}}^n)}\le C\Vert f{\Vert }_{L^2({\mathbb{R}}^n)}.$$ This is obtained from the pointwise dispersive estimate by the $T{T}^{\ast }$ method together with the Hardy-Littlewood-Sobolev inequality, the endpoint $q=2$ requiring the bilinear refinement of Keel-Tao ^{[Keel-Tao 1998] [Ginibre-Velo 1985]}. The dispersive estimate is the sole analytic input; everything else is abstract functional analysis.

Synthesis. The two propagators are two faces of one principle of stationary phase, and the bridge is the dispersion relation's Hessian. For Schrödinger the phase $x\cdot \xi -2\pi t|\xi {|}^2$ has a single non-degenerate critical point with full-rank Hessian, so the full $n/2$ decay appears and the kernel modulus is exactly constant; the foundational reason the estimate is so clean is that the phase is exactly quadratic and stationary phase truncates with no remainder. For the wave equation the phase lives on the cone, one principal curvature vanishes, and this is exactly the curvature degeneracy that drops the decay to $(n-1)/2$ and forces the loss of $(n+1)/2$ derivatives — the curvature bookkeeping of the surface-decay theorem is dual to the Hessian bookkeeping of stationary phase. Putting these together, both estimates feed the same $T{T}^{\ast }$ machine that generates the Strichartz estimates, which generalises the pointwise bound into the space-time bound that the nonlinear theory consumes; the central insight is that dispersion, surface restriction, and Strichartz are one phenomenon — the decay of an oscillatory integral over a curved characteristic variety — read at three levels of abstraction, and this same curved-variety decay appears again in the restriction conjecture and in the local smoothing estimates of the wave equation.

Full proof set Master

Proposition 1 (group velocity is the propagation speed of a wave packet). Let $\tau \in C^2$ and let $f$ have Fourier transform $\hat{f}$ concentrated in a small ball around a frequency ${\xi }_0$. Then the bulk of the packet $u(\cdot ,t)=e^{it\tau (\nabla /2\pi i)}f$ travels with velocity $v={\nabla }_{\xi }\tau ({\xi }_0)$, up to spreading of order $t\Vert {\nabla }^2\tau \Vert$ times the frequency width.

Proof. Write $u(x,t)=\int \hat{f}(\xi )e^{2\pi i(x\cdot \xi )+it\tau (\xi )}d\xi$. With $\hat{f}$ supported in $|\xi -{\xi }_0|\le \delta$, Taylor-expand $\tau (\xi )=\tau ({\xi }_0)+\nabla \tau ({\xi }_0)\cdot (\xi -{\xi }_0)+\frac{1}{2}(\xi -{\xi }_0{)}^{\top }{\nabla }^2\tau ({\xi }_{\ast })(\xi -{\xi }_0)$. The constant term contributes a global phase $e^{it\tau ({\xi }_0)}$. The linear term combines with $x\cdot \xi$ into $e^{2\pi i\xi \cdot (x+(t/2\pi )\nabla \tau ({\xi }_0))}$ evaluated against $\hat{f}$, which by the translation rule 02.10.04 is the original profile $f$ shifted by $-(t/2\pi )\nabla \tau ({\xi }_0)$; identifying the propagation speed, the packet centre moves at velocity proportional to $\nabla \tau ({\xi }_0)$, the group velocity. The quadratic remainder has size at most $t\Vert {\nabla }^2\tau \Vert {\delta }^2$ in the phase, producing spreading of the stated order. $\square$

Proposition 2 (sharpness of the Schrödinger exponent). The exponent $n/2$ in $\Vert e^{it\Delta }f{\Vert }_{L^{\infty }}\lesssim |t{|}^{-n/2}\Vert f{\Vert }_{L^1}$ cannot be improved: there is $f\in L^1\cap L^2$ and a sequence $t_k\to \infty$ with $\Vert e^{i{t}_k\Delta }f{\Vert }_{L^{\infty }}\ge c{t}_k^{-n/2}\Vert f{\Vert }_{L^1}$.

Proof. Take $f(x)=e^{-\pi |x{|}^2}$, a Schwartz Gaussian with $\Vert f{\Vert }_{L^1}=1$. By the exact kernel and the Gaussian integral (analytic continuation as in the Key theorem), $$(e^{it\Delta }f)(0)=\frac{1}{(4\pi it{)}^{n/2}}{\int }_{{\mathbb{R}}^n}{e}^{i|y{|}^2/4t}{e}^{-\pi |y{|}^2}dy=\frac{1}{(4\pi it{)}^{n/2}}(\pi -i/(4t){)}^{-n/2}\cdot {\pi }^{n/2}\cdots ,$$ and evaluating the Gaussian integral $\int e^{-(\pi -i/4t)|y{|}^2}dy=(\pi -i/4t{)}^{-n/2}$ gives $$|(e^{it\Delta }f)(0)|=\frac{1}{(4\pi |t|{)}^{n/2}}|\pi -i/4t{|}^{-n/2}=\frac{1}{(4\pi |t|{)}^{n/2}}({\pi }^2+1/16t^2{)}^{-n/4}.$$ As $t\to \infty$, $({\pi }^2+1/16t^2{)}^{-n/4}\to {\pi }^{-n/2}$, a positive constant. Hence $\Vert e^{it\Delta }f{\Vert }_{L^{\infty }}\ge |(e^{it\Delta }f)(0)|\ge c|t{|}^{-n/2}$ for all large $t$, with $c=(4\pi {)}^{-n/2}{\pi }^{-n/2}/2$ say. The upper bound is attained up to a constant, so the exponent is sharp. $\square$

Proposition 3 (Knapp obstruction: the wave exponent cannot reach $n/2$). For the half-wave propagator in dimension $n\ge 2$, no estimate $\Vert e^{it\sqrt{-\Delta }}f{\Vert }_{L^{\infty }}\lesssim |t{|}^{-n/2}\Vert f{\Vert }_{L^1}$ holds; the cone's degeneracy forces the slower rate $(n-1)/2$.

Proof. Test against a Knapp example. Fix large $\mu$ and let $\hat{f}$ be the indicator of a cap on the unit cone of angular width ${\mu }^{-1/2}$ and radial thickness $1$ near frequency $\mu e_1$: a slab of dimensions roughly $\mu \times {\mu }^{1/2}\times \cdots \times {\mu }^{1/2}$ (one long direction, $n-1$ short directions of length ${\mu }^{1/2}$), so $\Vert f{\Vert }_{L^1}\gtrsim \Vert \hat{f}{\Vert }_{L^{\infty }}\cdot 1$ while $\Vert \hat{f}{\Vert }_{L^1}\sim {\mu }^{(n-1)/2}$. On the dual tube of dimensions ${\mu }^{-1}\times {\mu }^{-1/2}\times \cdots$ and length $\sim t$ in time, the phases of all frequencies in the cap stay coherent (the cone is flat to within the cap's tolerance), so $|e^{it\sqrt{-\Delta }}f|$ is of size $\sim \Vert \hat{f}{\Vert }_{L^1}\sim {\mu }^{(n-1)/2}$ on a set of positive measure for $|t|\lesssim {\mu }^{1/2}$. Meanwhile $\Vert f{\Vert }_{L^1}\sim$ (Fourier-support volume)$\cdot \Vert \hat{f}{\Vert }_{L^{\infty }}\sim {\mu }^{(n+1)/2}\cdot {\mu }^{-(n+1)/2}\cdot \Vert \hat{f}{\Vert }_{L^1}$; tracking the normalization, the ratio $\Vert e^{it\sqrt{-\Delta }}f{\Vert }_{L^{\infty }}/\Vert f{\Vert }_{L^1}$ is forced to be $\gtrsim |t{|}^{-(n-1)/2}$ and cannot be $\lesssim |t{|}^{-n/2}$, since matching ${\mu }^{(n-1)/2}$ against $|t{|}^{-n/2}\Vert f{\Vert }_{L^1}$ with $|t|\sim {\mu }^{1/2}$ produces a contradiction by a half-power of $\mu$. The coherent cap — the Knapp example — is the obstruction, and it exists precisely because the cone has a vanishing principal curvature in the radial direction. $\square$

Proposition 4 ($T{T}^{\ast }$ reduction of Strichartz to dispersive). If a propagator $U(t)$ satisfies $\Vert U(t){\Vert }_{L^2\to L^2}\lesssim 1$ and $\Vert U(t)U(s{)}^{\ast }{\Vert }_{L^{r^{\prime }}\to L^r}\lesssim |t-s{|}^{-\sigma (r)}$ with $\sigma (r)=n(1/2-1/r)$ admissible, then $\Vert U(t)f{\Vert }_{L_t^q{L}_x^r}\lesssim \Vert f{\Vert }_{L^2}$.

Proof. By duality $\Vert U(t)f{\Vert }_{L_t^q{L}_x^r}\lesssim \Vert f{\Vert }_{L^2}$ is equivalent to the bound $\Vert \int U(s{)}^{\ast }g(s)ds{\Vert }_{L^2}\lesssim \Vert g{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}}$, and squaring reduces this to the bilinear form $\iint ⟨U(t)U(s{)}^{\ast }g(s),g(t)⟩dsdt\lesssim \Vert g{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}}^2$. Insert the dispersive bound $\Vert U(t)U(s{)}^{\ast }g(s){\Vert }_{L^r}\lesssim |t-s{|}^{-\sigma (r)}\Vert g(s){\Vert }_{L^{r^{\prime }}}$ and apply Hölder in $x$: $$\left|\iint ⟨\cdots ⟩dsdt\right|\lesssim \iint |t-s{|}^{-\sigma (r)}\Vert g(t){\Vert }_{L^{r^{\prime }}}\Vert g(s){\Vert }_{L^{r^{\prime }}}dsdt.$$ The right side is a one-dimensional fractional integral of order $1-\sigma (r)$ acting on $\Vert g(\cdot ){\Vert }_{L^{r^{\prime }}}$, and the Hardy-Littlewood-Sobolev inequality bounds it by $\Vert g{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}}^2$ exactly when $\sigma (r)=2/q^{\prime }$ with $q^{\prime }\le 2$, i.e. the admissibility relation $2/q+n/r=n/2$. For the non-endpoint range $q>2$ this is the full argument; the endpoint $q=2$ requires the Keel-Tao bilinear interpolation against dyadic time-frequency pieces ^{[Keel-Tao 1998]}. $\square$

Connections Master

• The propagator's unitarity and its $L^2$ conservation law rest entirely on the Fourier transform and Plancherel theorem of 02.10.04; the explicit Schrödinger kernel is the analytic continuation $z\to it$ of the heat kernel constructed in 02.13.03, and the dispersive-versus-smoothing dichotomy is precisely the contrast between an oscillatory Gaussian of constant modulus and a real Gaussian of decaying modulus.

• The stationary-phase and oscillatory-integral asymptotics that drive the wave estimate are the scalar-symbol prototype of the microlocal calculus of 02.14.02; the half-wave propagator $e^{it\sqrt{-\Delta }}$ is a Fourier integral operator whose canonical relation is the geodesic flow on the cone, and the Hadamard parametrix / geometric-optics construction is the variable-coefficient generalization of the constant-coefficient phase used here.

• The wave-propagator decay quantifies the dispersive content of the Strichartz one-line inequality stated abstractly in 02.13.04; this unit supplies the actual oscillatory-integral theory that the wave unit only names, and the curvature loss of $(n+1)/2$ derivatives is the analytic origin of the energy-method derivative budgets that appear in the nonlinear wave theory descending from 02.13.04.

Historical & philosophical context Master

The pointwise dispersive estimate for the Schrödinger equation follows from the explicit oscillatory Gaussian kernel, known since the early days of quantum mechanics through the path-integral and propagator formalism, but its modern role as a quantitative input to nonlinear PDE was crystallized by Robert Strichartz in his 1977 paper Restrictions of Fourier transforms to quadratic surfaces and decay of solutions of wave equations in the Duke Mathematical Journal ^{[Strichartz 1977]}. Strichartz recognized that the decay of solutions of dispersive equations is the same phenomenon as the restriction of the Fourier transform to curved surfaces studied by Stein and Tomas, unifying harmonic analysis and PDE; the space-time integrability estimates that now bear his name were the direct consequence.

The stationary-phase principle that governs the wave case has a longer pedigree, originating in the asymptotic methods of Stokes and Kelvin in the nineteenth century for water-wave and optical problems and systematized in the twentieth century within the theory of oscillatory integrals. Lars Hörmander's The Analysis of Linear Partial Differential Operators ^{[Hörmander 1990]} gave the definitive modern treatment of stationary phase with the signature-of-the-Hessian phase factor, and Elias Stein's Harmonic Analysis ^{[Stein 1993]} developed the curved-surface restriction estimates that quantify the cone's curvature degeneracy. The fixed-time $L^{p^{\prime }}\to L^p$ decay for the wave equation in the form requiring a derivative loss was established by Philip Brenner in 1975 ^{[Brenner 1975]}.

The estimates became the engine of the modern theory through the abstract $T{T}^{\ast }$ formulation of Ginibre and Velo in 1985 ^{[Ginibre-Velo 1985]} and the resolution of the endpoint case by Markus Keel and Terence Tao in 1998 ^{[Keel-Tao 1998]}, whose bilinear argument closed the delicate $q=2$ Strichartz estimate that the elementary Hardy-Littlewood-Sobolev approach misses. Sergiu Klainerman's vector-field method ^{[Klainerman 1985]} provided a parallel geometric route to decay via the Lorentz symmetries of the wave equation, and Tao's lecture notes ^{[Tao 2006]} gave the synthesis that organizes the dispersive, restriction, and Strichartz estimates as facets of one oscillatory-integral phenomenon.

Bibliography Master

@book{Tao2006,
  author    = {Tao, Terence},
  title     = {Nonlinear Dispersive Equations: Local and Global Analysis},
  series    = {CBMS Regional Conference Series in Mathematics},
  volume    = {106},
  publisher = {American Mathematical Society},
  year      = {2006}
}

@article{Strichartz1977,
  author  = {Strichartz, Robert S.},
  title   = {Restrictions of {F}ourier transforms to quadratic surfaces and decay of solutions of wave equations},
  journal = {Duke Mathematical Journal},
  volume  = {44},
  year    = {1977},
  pages   = {705--714}
}

@book{Stein1993,
  author    = {Stein, Elias M.},
  title     = {Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory Integrals},
  series    = {Princeton Mathematical Series},
  volume    = {43},
  publisher = {Princeton University Press},
  year      = {1993}
}

@article{KeelTao1998,
  author  = {Keel, Markus and Tao, Terence},
  title   = {Endpoint {S}trichartz estimates},
  journal = {American Journal of Mathematics},
  volume  = {120},
  year    = {1998},
  pages   = {955--980}
}

@article{GinibreVelo1985,
  author  = {Ginibre, Jean and Velo, Giorgio},
  title   = {The global {C}auchy problem for the nonlinear {S}chr\"odinger equation revisited},
  journal = {Annales de l'Institut Henri Poincar\'e, Analyse Non Lin\'eaire},
  volume  = {2},
  year    = {1985},
  pages   = {309--327}
}

@book{Hormander1990,
  author    = {H\"ormander, Lars},
  title     = {The Analysis of Linear Partial Differential Operators I},
  edition   = {2},
  series    = {Grundlehren der mathematischen Wissenschaften},
  volume    = {256},
  publisher = {Springer},
  year      = {1990}
}

@book{Sogge2008,
  author    = {Sogge, Christopher D.},
  title     = {Lectures on Non-Linear Wave Equations},
  edition   = {2},
  publisher = {International Press},
  year      = {2008}
}

@article{Brenner1975,
  author  = {Brenner, Philip},
  title   = {On $L^p$--$L^{p'}$ estimates for the wave equation},
  journal = {Mathematische Zeitschrift},
  volume  = {145},
  year    = {1975},
  pages   = {251--254}
}

@article{Klainerman1985,
  author  = {Klainerman, Sergiu},
  title   = {Uniform decay estimates and the {L}orentz invariance of the classical wave equation},
  journal = {Communications on Pure and Applied Mathematics},
  volume  = {38},
  year    = {1985},
  pages   = {321--332}
}