Strichartz Estimates via the TT* Method

Intuition Beginner

A wave packet under the Schrödinger equation spreads out and fades. The previous unit measured that fading one instant at a time: at each fixed moment the tallest value of the solution has dropped by a definite amount. Strichartz estimates ask a different and more useful question. Instead of looking at one snapshot, they add up the size of the solution across a whole stretch of time and space at once, and find that this total is controlled purely by the size of the starting bump — no matter how the bump was shaped.

Why is the all-at-once measurement worth the trouble? Because when you want to solve a hard nonlinear equation, you build the answer step by step, feeding the solution back into itself. A snapshot bound tells you the solution is small at each instant, but small-at-each-instant pieces can still add up to something large when you stack many instants together. What you really need is a guarantee about the accumulated size over time. The Strichartz estimate is exactly that guarantee, and it is the single tool that makes the step-by-step construction close.

The trade at the heart of it is this: pointwise the solution might be tall right at the start, but it cannot stay tall for long, because it is busy spreading. So if you measure size in a way that is forgiving of a brief tall spike but punishing of a sustained large value, the spreading works in your favour. The accumulated space-time size stays bounded by the starting size measured in the plain averaged-square sense. You give up asking "how tall right now" and gain a clean bound on "how much, all told."

There is a bookkeeping rule that says which combinations of time-averaging and space-averaging are allowed. You get to choose how harshly to weigh time versus space, but the two choices are locked together by one equation tied to the dimension. Most choices are permitted; one extreme choice, where you weigh time as gently as possible, sits right on the boundary and is genuinely harder to prove — it took twenty years longer than the rest.

The one-sentence takeaway: a Strichartz estimate converts the instant-by-instant fading of a dispersing wave into a single clean bound on its total size across space and time, and that accumulated bound is the fuel the nonlinear theory runs on.

Visual Beginner

Picture the life of a spreading packet drawn as a stack of horizontal strips, one strip per moment of time, time running upward. At the bottom strip the packet is a tall narrow spike. Each strip higher up shows the packet a little lower and a little wider, until near the top it is a broad shallow swell. A snapshot bound looks at one strip and reports its tallest point. A Strichartz estimate instead pours a thin layer of ink over the whole stack, darker where the packet is larger, and weighs the total ink. Because the tall part exists only in the bottom few strips and quickly thins out, the total ink stays modest.

The companion picture is the admissibility dial. Draw two dials side by side: one sets how harshly time is weighed, the other how harshly space is weighed. A chain links them so that turning one forces the other. The rule engraved on the chain subtracts a fixed multiple of the space setting from the time setting and demands the result equal a number fixed by the dimension. Almost every setting of the dials is allowed; the one setting where the time dial is turned all the way to its gentlest mark is the delicate borderline, allowed in high enough dimension but forbidden in the lowest case.

Worked example Beginner

We check the admissibility bookkeeping in three space dimensions and read off one allowed pair of weightings by hand, then test the forbidden extreme.

Step 1. Fix the dimension at $n=3$. The admissibility rule for the Schrödinger equation says: two divided by the time weighting, plus three divided by the space weighting, must equal three divided by two. Write the time weighting as $q$ and the space weighting as $r$. The rule is two over $q$ plus three over $r$ equals one and one half.

Step 2. Try the balanced choice where space is weighed by the demanding sixth power, so $r=6$. Then three over $r$ is three over six, which is one half. The rule needs the two terms to total one and one half, so two over $q$ must be one and one half minus one half, equal to one.

Step 3. Solve for the time weighting. Two over $q$ equals one means $q=2$. So in three dimensions the pair "time weighting two, space weighting six" satisfies the rule exactly. This is the famous endpoint pair in dimension three, and it is allowed here because the dimension is large enough.

Step 4. Try the other easy choice where time and space are weighed equally, $q=r$. The rule becomes two over $q$ plus three over $q$ equals one and one half, so five over $q$ equals one and one half, giving $q=5/1.5=10/3$, about $3.33$. So the pair "time weighting ten-thirds, space weighting ten-thirds" is also allowed.

Step 5. Test the forbidden extreme in a lower dimension. Drop to $n=2$, where the rule is two over $q$ plus two over $r$ equals one. The gentlest possible time weighting is $q=2$, which makes two over $q$ equal one, forcing two over $r$ to be zero, that is $r$ infinite. This corner — time weighting two, space weighting infinite, in dimension two — is exactly the one case the clean estimate fails, the lone exception carved out of the rule.

What this tells us: the admissibility rule is a single straight-line relation between the two weightings; most points on the line give valid estimates, the special endpoint where time is weighed as gently as possible is valid only when the dimension is at least three, and in dimension two that corner is the one genuine gap.

Check your understanding Beginner

Formal definition Intermediate+

Fix $n\ge 1$. Work on ${\mathbb{R}}^n$ with the Fourier transform of 02.10.04 and the free Schrödinger propagator $e^{it\Delta }$ of 02.21.01, a unitary group on $L^2({\mathbb{R}}^n)$ satisfying the pointwise dispersive estimate $\Vert e^{it\Delta }f{\Vert }_{L^{\infty }}\le (4\pi |t|{)}^{-n/2}\Vert f{\Vert }_{L^1}$ and its interpolated form $\Vert e^{it\Delta }f{\Vert }_{L^r}\lesssim |t{|}^{-n(1/2-1/r)}\Vert f{\Vert }_{L^{r^{\prime }}}$ for $2\le r\le \infty$, with $1/r+1/r^{\prime }=1$.

Definition (mixed space-time Lebesgue norm). For $1\le q,r\le \infty$ and $u:{\mathbb{R}}_t\times {\mathbb{R}}_x^n\to \mathbb{C}$ the mixed norm is $$\Vert u{\Vert }_{L_t^q{L}_x^r}={\left({\int }_{\mathbb{R}}\Vert u(t,\cdot ){\Vert }_{L_x^r}^{q}dt\right)}^{1/q}={\left({\int }_{\mathbb{R}}({\int }_{{\mathbb{R}}^n}|u(t,x){|}^{r}dx{)}^{q/r}dt\right)}^{1/q},$$ with the usual essential-supremum modification when $q$ or $r$ is $\infty$. The space $L_t^q{L}_x^r$ is the Banach space of $u$ with $\Vert u{\Vert }_{L_t^q{L}_x^r}<\infty$; its dual is $L_t^{q^{\prime }}{L}_x^{r^{\prime }}$ under the space-time pairing $⟨u,v⟩=\iint u\bar{v}dxdt$.

Definition (Schrödinger-admissible pair). A pair $(q,r)$ is admissible (sharp Schrödinger-admissible) if $2\le q,r\le \infty$, the scaling relation $$\frac{2}{q}+\frac{n}{r}=\frac{n}{2}$$ holds, and $(q,r,n)\ne (2,\infty ,2)$. The relation is forced by the scaling symmetry $u(x,t)\mapsto u(\lambda x,{\lambda }^{2}t)$ of the free equation: it is the unique linear constraint under which $\Vert e^{it\Delta }f{\Vert }_{L_t^q{L}_x^r}$ and $\Vert f{\Vert }_{L^2}$ have the same dilation weight. The excluded corner $(2,\infty ,2)$ is the sole admissible-by-scaling pair at which the estimate fails.

Definition (homogeneous Strichartz estimate). The homogeneous Strichartz estimate for an admissible pair $(q,r)$ is the space-time bound $$\Vert e^{it\Delta }f{\Vert }_{L_t^q{L}_x^r(\mathbb{R}\times {\mathbb{R}}^n)}\le C\Vert f{\Vert }_{L^2({\mathbb{R}}^n)},f\in L^2({\mathbb{R}}^n).$$ Equivalently, the map $f\mapsto e^{it\Delta }f$ is bounded $L_x^2\to L_t^q{L}_x^r$.

Definition (dual and retarded inhomogeneous estimates). The dual estimate is the adjoint bound: for admissible $(q,r)$, $${\Vert {\int }_{\mathbb{R}}{e}^{-is\Delta }F(s,\cdot )ds\Vert }_{L_x^2}\le C\Vert F{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}}.$$ The retarded (inhomogeneous) estimate concerns the Duhamel operator $(\mathcal{D}F)(t)={\int }_{s<t}{e}^{i(t-s)\Delta }F(s,\cdot )ds$ solving $i{\partial }_{t}u+\Delta u=F$, $u(0)=0$; for admissible pairs $(q,r)$ and $(\tilde{q},\tilde{r})$ it asserts $$\Vert \mathcal{D}F{\Vert }_{L_t^q{L}_x^r}\le C\Vert F{\Vert }_{L_t^{{\tilde{q}}^{\prime }}{L}_x^{{\tilde{r}}^{\prime }}}.$$ The time-cutoff $s<t$ — the retardation — is the only difference between $\mathcal{D}$ and the untruncated ${\int }_{\mathbb{R}}$; passing the bound across that cutoff is the role of the Christ-Kiselev lemma.

Counterexamples to common slips Intermediate+

• Admissibility is two conditions, not one. The scaling relation $2/q+n/r=n/2$ alone is necessary but not sufficient: one must also have $q\ge 2$ (and $r\ge 2$). A pair with $q<2$ violates the Galilean-/translation-invariance obstruction (a Knapp-type example in time) and no homogeneous Strichartz estimate holds, even though scaling is satisfied.

• The endpoint is excluded only in dimension two. The pair $(q,r)=(2,2n/(n-2))$ is admissible and valid for every $n\ge 3$; it fails only at $n=2$, where it degenerates to $(2,\infty )$. Writing the endpoint as $(2,\infty )$ and declaring it always forbidden confuses the dimension-two degeneracy with the general endpoint, which is the hard but true Keel-Tao case.

• The TT^ exponent must be admissible for HLS to close.* The non-endpoint argument feeds the dispersive decay $|t-s{|}^{-n(1/2-1/r)}$ into the one-dimensional Hardy-Littlewood-Sobolev inequality of 02.19.05. That fractional-integration step requires the time-decay exponent to equal $2/q^{\prime }$ with the HLS scaling line satisfied, i.e. exactly $2/q+n/r=n/2$ with $q>2$. If one tries an inadmissible exponent the HLS hypothesis $1<q^{\prime }<\infty$ on the scaling line is violated and the integral diverges.

• Christ-Kiselev needs strictly ordered exponents. The retarded estimate is not a free corollary of the untruncated bound: passing from ${\int }_{\mathbb{R}}$ to ${\int }_{s<t}$ uses the Christ-Kiselev lemma, which requires $q^{\prime }<q$ strictly. At the endpoint $q=q^{\prime }=2$ the lemma fails, which is precisely why the endpoint inhomogeneous estimate needs the separate Keel-Tao bilinear treatment rather than Christ-Kiselev.

Key theorem with proof Intermediate+

Theorem (homogeneous Strichartz estimate via TT^*; Strichartz 1977, Ginibre-Velo 1992). Let $n\ge 1$ and let $(q,r)$ be admissible with $q>2$. Then there is $C=C(n,q,r)$ such that $$\Vert e^{it\Delta }f{\Vert }_{L_t^q{L}_x^r(\mathbb{R}\times {\mathbb{R}}^n)}\le C\Vert f{\Vert }_{L^2({\mathbb{R}}^n)}(f\in L^2).$$

Proof. Write $T:L_x^2\to L_t^q{L}_x^r$, $Tf=e^{it\Delta }f$. The claim is $\Vert T\Vert <\infty$. The TT^* method reduces this to a bound on a single space-time operator built from the dispersive decay.

Step 1 (TT^ duality).* For a bounded linear $T:H\to B$ between a Hilbert space $H$ and a Banach space $B$, the three statements $$\Vert Tf{\Vert }_B\le C\Vert f{\Vert }_H,\Vert T^{\ast }g{\Vert }_H\le C\Vert g{\Vert }_{B^{\ast }},\Vert T{T}^{\ast }g{\Vert }_B\le C^2\Vert g{\Vert }_{B^{\ast }}$$ are equivalent, with the same constant $C$. The first two are adjoint to each other; the third follows from the first two by $\Vert T{T}^{\ast }g{\Vert }_B\le C\Vert T^{\ast }g{\Vert }_H\le C^2\Vert g{\Vert }_{B^{\ast }}$, and conversely $\Vert T^{\ast }g{\Vert }_H^2=⟨T{T}^{\ast }g,g⟩\le \Vert T{T}^{\ast }g{\Vert }_B\Vert g{\Vert }_{B^{\ast }}\le C^2\Vert g{\Vert }_{B^{\ast }}^2$ recovers the second. Here $H=L_x^2$, $B=L_t^q{L}_x^r$, $B^{\ast }=L_t^{q^{\prime }}{L}_x^{r^{\prime }}$. The adjoint is $T^{\ast }g={\int }_{\mathbb{R}}{e}^{-is\Delta }g(s,\cdot )ds$, so $$T{T}^{\ast }g(t)={\int }_{\mathbb{R}}{e}^{i(t-s)\Delta }g(s,\cdot )ds.$$ It suffices to prove $\Vert T{T}^{\ast }g{\Vert }_{L_t^q{L}_x^r}\le C^2\Vert g{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}}$.

Step 2 (insert the dispersive decay). Fix $t$. By Minkowski's integral inequality and the interpolated dispersive estimate from 02.21.01 applied to $e^{i(t-s)\Delta }$, $$\Vert T{T}^{\ast }g(t){\Vert }_{L_x^r}\le {\int }_{\mathbb{R}}\Vert e^{i(t-s)\Delta }g(s,\cdot ){\Vert }_{L_x^r}ds\lesssim {\int }_{\mathbb{R}}|t-s{|}^{-n(1/2-1/r)}\Vert g(s,\cdot ){\Vert }_{L_x^{r^{\prime }}}ds.$$ Set $\sigma =n(1/2-1/r)$ and $h(s)=\Vert g(s,\cdot ){\Vert }_{L_x^{r^{\prime }}}$, a non-negative function of $t$ alone after taking norms in $x$. The right side is the one-dimensional convolution $(|\cdot {|}^{-\sigma }\ast h)(t)$, a fractional integral of order $1-\sigma$.

Step 3 (Hardy-Littlewood-Sobolev in time). Taking the $L_t^q$ norm and applying the one-dimensional Hardy-Littlewood-Sobolev inequality of 02.19.05 (here $\alpha =1-\sigma$, dimension one), $$\Vert T{T}^{\ast }g{\Vert }_{L_t^q{L}_x^r}\lesssim \Vert |\cdot {|}^{-\sigma }\ast h{\Vert }_{L_t^q}\lesssim \Vert h{\Vert }_{L_t^{q^{\prime }}}=\Vert g{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}},$$ provided the HLS scaling line $1/q=1/q^{\prime }-(1-\sigma )$ holds with $1<q^{\prime }<q<\infty$. Compute: $1/q=1/q^{\prime }-1+\sigma$ rearranges to $1/q+1/q^{\prime }=1+\sigma -1=\sigma$ is not the relation; instead, using $1/q^{\prime }=1-1/q$, the HLS line reads $1/q-(1-1/q)=-(1-\sigma )$, i.e. $2/q-1=\sigma -1$, so $2/q=\sigma =n(1/2-1/r)$. That is exactly the admissibility relation $2/q+n/r=n/2$. The HLS hypothesis $1<q^{\prime }<q$ becomes $q>2$, the standing assumption. This proves $\Vert T{T}^{\ast }g{\Vert }_{L_t^q{L}_x^r}\lesssim \Vert g{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}}$, hence by Step 1 the homogeneous estimate. $\square$

Bridge. This argument is the foundational reason the entire dispersive-PDE chapter exists: it builds toward the local well-posedness theory of nonlinear Schrödinger equations, where the homogeneous and retarded estimates of this unit are the contraction-mapping norms, and it appears again in the wave and Klein-Gordon settings once the cone's curvature loss of 02.21.01 is folded into the admissibility arithmetic. The central insight is that the only analytic input is the pointwise dispersive decay of 02.21.01; everything downstream — duality, Minkowski, the one-dimensional fractional integral — is abstract, so this is exactly the mechanism that converts a single decay rate into an entire family of space-time bounds. The TT^* reduction is dual to the restriction-extension duality of the Stein-Tomas theorem, and putting these together the Strichartz estimate generalises the dispersive estimate of 02.21.01 from a fixed-time inequality to a space-time inequality, with the Hardy-Littlewood-Sobolev inequality of 02.19.05 as the bridge that the admissibility relation $2/q+n/r=n/2$ encodes.

Exercises Intermediate+

Advanced results Master

Theorem (abstract Keel-Tao Strichartz; Keel-Tao 1998). Let $(X,dx)$ be a measure space, $H$ a Hilbert space, and $\{U(t){\}}_{t\in \mathbb{R}}$ a family of operators $U(t):H\to L^2(X)$ obeying the energy estimate $\Vert U(t)f{\Vert }_{L_x^2}\lesssim \Vert f{\Vert }_H$ and the decay estimate $\Vert U(t)U(s{)}^{\ast }g{\Vert }_{L_x^{\infty }}\lesssim |t-s{|}^{-\beta }\Vert g{\Vert }_{L_x^1}$ for some $\beta >0$ and all $t\ne s$. Call $(q,r)$ $\beta$-admissible if $q,r\ge 2$, $(q,r,\beta )\ne (2,\infty ,1)$, and $1/q+\beta /r=\beta /2$. Then for all $\beta$-admissible $(q,r)$ and $(\tilde{q},\tilde{r})$, $$\Vert U(t)f{\Vert }_{L_t^q{L}_x^r}\lesssim \Vert f{\Vert }_H,\Vert {\int }_{s<t}U(t)U(s{)}^{\ast }F(s)ds{\Vert }_{L_t^q{L}_x^r}\lesssim \Vert F{\Vert }_{L_t^{{\tilde{q}}^{\prime }}{L}_x^{{\tilde{r}}^{\prime }}},$$ *including the endpoint* $q=2$ whenever $\beta >1$ ^{[Keel-Tao 1998]}. For Schrödinger $\beta =n/2$, recovering $2/q+n/r=n/2$; the endpoint $q=2$ is then available for $n>2$.

Theorem (why $q=2$ is delicate). At the endpoint $q=2$ the time-decay exponent $\sigma =n(1/2-1/r)=1$, so the convolution kernel $|t-s{|}^{-1}$ in the TT^* reduction is not locally integrable and lies exactly at the failure threshold of the one-dimensional Hardy-Littlewood-Sobolev inequality (the weak-type, non-strong, $L^1\to L^{1,\infty }$ borderline). The naive HLS step diverges logarithmically. Keel-Tao replace it by a dyadic decomposition in time-separation: split $T{T}^{\ast }={\sum }_{j\in \mathbb{Z}}{T}_j$ with $T_j$ carrying $|t-s|\sim 2^j$, prove for each $j$ a bilinear estimate $$|⟨T_{j}F,G⟩|\lesssim 2^{-j\beta (1/2-1/r)}{2}^{j(1-2/q)}\Vert F{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}}\Vert G{\Vert }_{L_t^{q^{\prime }}{L}_x^{r^{\prime }}}$$ interpolated between the energy and decay endpoints at two different spatial exponents $r$, and sum the geometric series in $j$. The summation converges for $q>2$ termwise and, at $q=2$, only after the off-diagonal $r\ne \tilde{r}$ refinement and an atomic decomposition of the Lorentz space $L^{r,2}$; the diagonal $L^{r,r}=L^r$ argument fails, which is the precise sense in which $q=2$ is harder.

Theorem (failure at $(2,\infty ,2)$). In dimension $n=2$ the would-be endpoint is $(q,r)=(2,\infty )$, and $\Vert e^{it\Delta }f{\Vert }_{L_t^2{L}_x^{\infty }}\lesssim \Vert f{\Vert }_{L^2}$ is false. A counterexample is built from a Knapp wave packet: taking $\hat{f}$ the indicator of a unit cube and tracking the focusing of $e^{it\Delta }f$ near the spatial origin over a unit time interval produces a logarithmic divergence in the $L_t^2{L}_x^{\infty }$ norm while $\Vert f{\Vert }_{L^2}$ stays bounded ^{[Keel-Tao 1998]}. The obstruction is the failure of $L_x^{\infty }$ to be a Banach space with the requisite duality/interpolation properties at the borderline decay $\sigma =1$; replacing $L^{\infty }$ by BMO or by the Lorentz refinement at higher dimension is what rescues the endpoint for $n\ge 3$.

Theorem (Strichartz from restriction-extension; Strichartz 1977, Tomas 1975). The homogeneous estimate is equivalent, by the $T{T}^{\ast }$/Plancherel dictionary, to the adjoint Fourier restriction estimate for the paraboloid $\Sigma =\{(\xi ,|\xi {|}^2)\}\subset {\mathbb{R}}^{n+1}$: $\Vert (gd{\sigma }_{\Sigma }{)}^{\vee }{\Vert }_{L_{t,x}^q}\lesssim \Vert g{\Vert }_{L^2(d{\sigma }_{\Sigma })}$ on the diagonal $q=r$. The Stein-Tomas exponent $q=2(n+2)/n$ is the diagonal admissible Strichartz exponent, and the curvature of the paraboloid — the same non-degenerate Hessian that drives the dispersive decay of 02.21.01 — is what powers both ^{[Strichartz 1977] [Tomas 1975]}.

Synthesis. The TT^* method is the central insight that organises this entire unit: it shows the homogeneous estimate, the dual estimate, and the retarded estimate are one statement read through duality, and the foundational reason all three hold is the single pointwise dispersive decay of 02.21.01 fed into a one-dimensional fractional integral. Putting these together, the admissibility relation $2/q+n/r=n/2$ is exactly the Hardy-Littlewood-Sobolev scaling line of 02.19.05 transcribed into the time variable, which is the bridge between fixed-time decay and space-time integrability; this is exactly the same arithmetic that, in the abstract Keel-Tao formulation, becomes $1/q+\beta /r=\beta /2$ with the dispersive exponent $\beta$ as the only datum, so the wave equation's curvature loss of 02.21.01 enters merely by changing $\beta$ from $n/2$ to $(n-1)/2$. The endpoint $q=2$ is delicate because the fractional-integral kernel $|t-s{|}^{-1}$ sits on the HLS failure threshold, and the Keel-Tao dyadic-bilinear argument that rescues it is dual to the Stein-Tomas restriction theorem; what looks like three separate estimates — dispersive decay, Strichartz, Fourier restriction — generalises to one curved-variety phenomenon, and the same TT^* duality appears again in the local-smoothing and bilinear-restriction theory that the nonlinear well-posedness program of the dispersive chapter consumes.

Full proof set Master

Proposition 1 (Minkowski integral inequality controls the Duhamel term). For measurable $g:{\mathbb{R}}_s\times {\mathbb{R}}^n\to \mathbb{C}$ and $1\le r\le \infty$, $\Vert {\int }_{\mathbb{R}}{e}^{i(t-s)\Delta }g(s,\cdot )ds{\Vert }_{L_x^r}\le {\int }_{\mathbb{R}}\Vert e^{i(t-s)\Delta }g(s,\cdot ){\Vert }_{L_x^r}ds$.

Proof. This is Minkowski's integral inequality: for any Banach-space-valued integrable family $\{F_s\}$, $\Vert \int F_{s}ds{\Vert }_{L_x^r}\le \int \Vert F_s{\Vert }_{L_x^r}ds$, valid for $1\le r\le \infty$ by the triangle inequality and Fubini applied to the dual pairing. Apply it with $F_s=e^{i(t-s)\Delta }g(s,\cdot )\in L_x^r$. $\square$

Proposition 2 (the admissibility relation is the HLS scaling line in time). With $\sigma =n(1/2-1/r)$, the one-dimensional fractional integration $h\mapsto |\cdot {|}^{-\sigma }\ast h$ maps $L_t^{q^{\prime }}\to L_t^q$ with $q>2$ if and only if $2/q+n/r=n/2$.

Proof. The one-dimensional Hardy-Littlewood-Sobolev inequality of 02.19.05 with kernel $|t{|}^{-\sigma }$, $0<\sigma <1$, maps $L_t^p\to L_t^a$ on the scaling line $1/a=1/p-(1-\sigma )$ for $1<p<a<\infty$. Set $p=q^{\prime }$, $a=q$: the line reads $1/q=1/q^{\prime }-(1-\sigma )=(1-1/q)-1+\sigma =\sigma -1/q$, hence $2/q=\sigma =n(1/2-1/r)$, i.e. $2/q+n/r=n/2$. The strict ordering $q^{\prime }<q$ forces $q>2$, and $0<\sigma <1$ corresponds to $r$ strictly between $2$ and $\infty$ (the open non-endpoint range). $\square$

Proposition 3 (necessity of $q\ge 2$). No homogeneous Strichartz estimate $\Vert e^{it\Delta }f{\Vert }_{L_t^q{L}_x^r}\lesssim \Vert f{\Vert }_{L^2}$ holds when $q<2$, even if $2/q+n/r=n/2$.

Proof. Translation invariance in time gives a Galilean-boost family: for the modulated datum $f_v(x)=e^{iv\cdot x}f(x)$, $e^{it\Delta }{f}_v(x)=e^{iv\cdot x-it|v{|}^2}(e^{it\Delta }f)(x-2tv)$, so $\Vert e^{it\Delta }{f}_v(\cdot ){\Vert }_{L_x^r}=\Vert e^{it\Delta }f(\cdot ){\Vert }_{L_x^r}$ is unchanged. Superpose $N$ such packets at well-separated velocities $v_1,\dots ,v_N$ whose space-time tubes are disjoint; then $\Vert f{\Vert }_{L^2}^2=\sum \Vert f{\Vert }_{L^2}^2=N\Vert f{\Vert }^2$ grows like $N^{1/2}$ in $L^2$, while the disjoint tubes make $\Vert e^{it\Delta }\sum f_{v_j}{\Vert }_{L_t^q{L}_x^r}^q={\sum }_j\Vert e^{it\Delta }{f}_{v_j}{\Vert }^q$ grow like $N^{1/q}$. The estimate would force $N^{1/q}\lesssim N^{1/2}$ for all $N$, i.e. $1/q\le 1/2$, that is $q\ge 2$. $\square$

Proposition 4 (Christ-Kiselev truncation lemma). Let $S$ have kernel $K(t,s)$ and be bounded $L_t^p(B_1)\to L_t^{\tilde{q}}(B_2)$ with norm $A$, $1\le p<\tilde{q}\le \infty$. Then $S_{<}F(t)={\int }_{s<t}K(t,s)F(s)ds$ is bounded with norm $\lesssim A(1-2^{-\delta }{)}^{-1}$, $\delta =1/p-1/\tilde{q}>0$.

Proof (sketch in the diadic-stopping-time form). Normalise $\Vert F{\Vert }_{L_t^p(B_1)}=1$ and define the distribution function $\lambda (t)={\int }_{-\infty }^t\Vert F(s){\Vert }_{B_1}^{p}ds\in [0,1]$. For each dyadic level partition $[0,1]$ into $2^k$ intervals and let ${\Omega }_{k,i}={\lambda }^{-1}(I_{k,i})$ be the corresponding time intervals; the retarded region $\{s<t\}$ decomposes as a disjoint union over the binary tree of pairs $({\Omega }_{k,i}^{\text{left}},{\Omega }_{k,i}^{\text{right}})$. On each pair the truncated kernel coincides with the full kernel of $S$ restricted to a product set, so the untruncated bound applies and contributes $A\Vert F{1}_{\text{left}}{\Vert }_{L^p}\Vert \cdot \Vert$ with $\Vert F{1}_{\Omega }{\Vert }_{L^p}^p=$ (length of $I_{k,i}$) $=2^{-k}$. Summing over the $2^k$ pieces at level $k$ and over levels with the ${\ell }^{\tilde{q}}$/${\ell }^p$ embedding ($\tilde{q}>p$) gives a geometric series ${\sum }_k{2}^{k/\tilde{q}}{2}^{-k/p}={\sum }_k{2}^{-k\delta }=(1-2^{-\delta }{)}^{-1}$, finite precisely because $\delta >0$. Hence $\Vert S_{<}F{\Vert }_{L_t^{\tilde{q}}(B_2)}\lesssim A(1-2^{-\delta }{)}^{-1}$ ^{[Christ-Kiselev 2001]}. $\square$

Proposition 5 (non-endpoint retarded estimate). For admissible $(q,r)$ and $(\tilde{q},\tilde{r})$ with $q,\tilde{q}>2$, $\Vert \mathcal{D}F{\Vert }_{L_t^q{L}_x^r}\lesssim \Vert F{\Vert }_{L_t^{{\tilde{q}}^{\prime }}{L}_x^{{\tilde{r}}^{\prime }}}$.

Proof. The untruncated operator $g\mapsto {\int }_{\mathbb{R}}{e}^{i(t-s)\Delta }g(s)ds=T{\tilde{T}}^{\ast }g$ factors as $T$ (homogeneous, bounded $L_x^2\to L_t^q{L}_x^r$ by the Key Theorem) composed with ${\tilde{T}}^{\ast }$ (dual of the homogeneous estimate for $(\tilde{q},\tilde{r})$, bounded $L_t^{{\tilde{q}}^{\prime }}{L}_x^{{\tilde{r}}^{\prime }}\to L_x^2$ by Exercise 6). Hence the untruncated bound holds with the product constant. Apply Proposition 4 with $B_1=L_x^{{\tilde{r}}^{\prime }}$, $B_2=L_x^r$, $p={\tilde{q}}^{\prime }$, ${\tilde{q}}_{\text{lemma}}=q$: the hypothesis $p<{\tilde{q}}_{\text{lemma}}$ is ${\tilde{q}}^{\prime }<q$, true since ${\tilde{q}}^{\prime }<2<q$. The truncation $\mathcal{D}=S_{<}$ is therefore bounded with the stated norm. $\square$

Proposition 6 (endpoint admissible exponent in dimension $n$). For $n\ge 3$ the endpoint pair is $(q,r)=(2,2n/(n-2))$, and it satisfies $2/q+n/r=n/2$ with $2<r<\infty$.

Proof. At $q=2$, the relation gives $n/r=n/2-1=(n-2)/2$, so $r=2n/(n-2)$. For $n\ge 3$ this is finite and exceeds $2$ (since $2n/(n-2)>2⟺2n>2n-4$, always), so $2<r<\infty$ and the pair is admissible. At $n=2$ the formula gives $r=\infty$, the excluded corner $(2,\infty ,2)$; at $n=1$, $q=2$ forces $n/r=-1/2<0$, impossible, so $q=2$ is unavailable in one dimension and the admissible range is $q>4$ with $r=\infty$ excluded throughout the $n=1$ analysis. $\square$

Connections Master

• The sole analytic input to every estimate here is the pointwise dispersive decay $\Vert e^{it\Delta }f{\Vert }_{L^r}\lesssim |t{|}^{-n(1/2-1/r)}\Vert f{\Vert }_{L^{r^{\prime }}}$ of 02.21.01; the TT^* method is the abstract machine that converts that fixed-time bound into the space-time bound, and the wave/Klein-Gordon Strichartz estimates follow by replacing the Schrödinger decay exponent $n/2$ with the curvature-degenerate cone exponent $(n-1)/2$ from that same unit, changing only $\beta$ in the Keel-Tao admissibility line.

• The time-convolution step is exactly the one-dimensional Hardy-Littlewood-Sobolev inequality of 02.19.05: the admissibility relation $2/q+n/r=n/2$ is the HLS scaling line $1/q=1/q^{\prime }-(1-\sigma )$ transcribed to the time axis, so the fractional-integration theory of that unit is the engine of the non-endpoint Strichartz proof, and its $p=1$ weak-type failure is the precise reason the $q=2$ endpoint sits at the HLS borderline.

• The mixed-norm duality $(L_t^q{L}_x^r{)}^{\ast }=L_t^{q^{\prime }}{L}_x^{r^{\prime }}$ and the Minkowski and Hölder inequalities underlying the TT^* bookkeeping are the $L^p$-space theory of 02.10.04 and 02.07.06; the Plancherel theorem of 02.10.04 additionally supplies the restriction-extension equivalence that recasts the diagonal Strichartz estimate as the Stein-Tomas Fourier restriction theorem for the paraboloid.

Historical & philosophical context Master

The space-time integrability estimates now bearing the name originate with Robert Strichartz, whose 1977 Duke Mathematical Journal paper Restrictions of Fourier transforms to quadratic surfaces and decay of solutions of wave equations ^{[Strichartz 1977]} proved the diagonal $L_{t,x}^q$ estimates for the wave and Schrödinger equations by recognising them as instances of the Stein-Tomas Fourier-restriction theorem for the paraboloid and cone; Peter Tomas had established the relevant restriction estimate in his 1975 Bulletin of the AMS note ^{[Tomas 1975]}. The decoupling of the time and space exponents into the full admissible family, and the systematic $T{T}^{\ast }$ formulation that derives the estimates from the dispersive decay plus an abstract duality, is due to Jean Ginibre and Giorgio Velo, whose 1985 and 1992 papers — the latter, Smoothing properties and retarded estimates for some dispersive evolution equations in Communications in Mathematical Physics 144 ^{[Ginibre-Velo 1992]} — gave the retarded inhomogeneous estimates and the off-diagonal range, with related contributions by Kenji Yajima ^{[Yajima 1987]}.

The endpoint $q=2$ resisted the elementary Hardy-Littlewood-Sobolev approach because the relevant fractional-integration kernel sits exactly at the inequality's failure threshold, and it was settled only in 1998 by Markus Keel and Terence Tao in Endpoint Strichartz estimates in the American Journal of Mathematics ^{[Keel-Tao 1998]}, whose abstract theorem replaces the divergent HLS step with a dyadic-in-time bilinear interpolation against the Lorentz scale and simultaneously establishes the failure of the $(2,\infty )$ corner in dimension two. The truncation lemma that transfers the untruncated estimate to the retarded Duhamel operator was isolated by Michael Christ and Alexander Kiselev in their 2001 Journal of Functional Analysis paper on maximal functions associated to filtrations ^{[Christ-Kiselev 2001]}, originally in the spectral-theory of one-dimensional Schrödinger operators, and recognised to give the clean ordered-exponent passage used throughout dispersive PDE.

Bibliography Master

@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}
}

@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{GinibreVelo1992,
  author  = {Ginibre, Jean and Velo, Giorgio},
  title   = {Smoothing properties and retarded estimates for some dispersive evolution equations},
  journal = {Communications in Mathematical Physics},
  volume  = {144},
  year    = {1992},
  pages   = {163--188}
}

@article{ChristKiselev2001,
  author  = {Christ, Michael and Kiselev, Alexander},
  title   = {Maximal functions associated to filtrations},
  journal = {Journal of Functional Analysis},
  volume  = {179},
  year    = {2001},
  pages   = {409--425}
}

@article{Tomas1975,
  author  = {Tomas, Peter A.},
  title   = {A restriction theorem for the {F}ourier transform},
  journal = {Bulletin of the American Mathematical Society},
  volume  = {81},
  year    = {1975},
  pages   = {477--478}
}

@article{Yajima1987,
  author  = {Yajima, Kenji},
  title   = {Existence of solutions for {S}chr\"odinger evolution equations},
  journal = {Communications in Mathematical Physics},
  volume  = {110},
  year    = {1987},
  pages   = {415--426}
}

@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}
}