Schrödinger evolution returns to rough data at the exact threshold

← All digests

Family 080 · Real and complex analysis
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-14

1. The question

A solution of an evolution equation should approach its initial data as time goes to zero. For smooth data this expectation is usually straightforward. For rough data, convergence in an average norm can hold even when the values at individual points behave badly.

Carleson’s Schrödinger convergence problem asks how much Sobolev regularity guarantees almost-everywhere return to the initial state. Earlier theorems identified the critical regularity in dimensions at least two up to a strict inequality. The question left at the boundary was whether the exact critical exponent also works.

The two manuscripts in family 080 claim that it does: one handles the plane, and the other all higher dimensions. They also specify a common pointwise representative of the evolution, constructed by removing Gaussian regularization uniformly in time on one full-measure spatial set.

2. Background and essential definitions

The free evolution

The free Schrödinger equation is

$$i\partial_tu+\Delta u=0,\qquad u(x,0)=f(x),$$

where $\Delta=\sum_{j=1}^n\partial_{x_j}^2$. With the Fourier convention $\widehat f(\xi)=\int e^{-ix\cdot\xi}f(x)\,dx$, its formal solution is

$$u(x,t)=\frac1{(2\pi)^n}\int_{\mathbb R^n} e^{ix\cdot\xi-it|\xi|^2}\widehat f(\xi)\,d\xi.$$

For general square-integrable $f$, the Fourier transform is defined by completion in $L^2$. The integral need not be absolutely convergent at each point. This is why the choice of a pointwise representative matters.

The evolution always converges to $f$ in $L^2$ as $t\to0$. That fact alone does not force convergence at almost every point along all times. Norm convergence can conceal large values moving through smaller spatial sets.

Sobolev regularity as a frequency penalty

The inhomogeneous Sobolev space $H^s(\mathbb R^n)$ consists of functions satisfying

$$\|f\|_{H^s}^2=\frac1{(2\pi)^n}\int_{\mathbb R^n} (1+|\xi|^2)^s|\widehat f(\xi)|^2\,d\xi<\infty.$$

The exponent $s$ measures a fractional number of square-integrable derivatives. Larger $s$ penalizes high frequencies more strongly. For data near frequency $N$, the Sobolev norm contributes roughly the factor $N^s$.

The critical exponent in dimensions $n\ge2$ is

$$s_n=\frac{n}{2(n+1)}.$$

Thus $s_2=1/3$, $s_3=3/8$, and $s_4=2/5$. An estimate that works only for $s>s_n$ does not establish membership at equality, however close $s$ is allowed to be.

Gaussian regularization and Lebesgue values

For $a>0$, define

$$U_af(x,t)=\frac1{(2\pi)^n}\int_{ \mathbb R^n}e^{ix\cdot\xi-it|\xi|^2-a|\xi|^2} \widehat f(\xi)\,d\xi.$$

The Gaussian factor damps high frequencies, making the integral absolutely convergent for $L^2$ data. The manuscripts define the desired evolution by taking $a\downarrow0$ first.

At almost every point, a locally integrable function has a Lebesgue value

$$f^*(x)=\lim_{r\downarrow0}\frac1{|B(x,r)|} \int_{B(x,r)}f(y)\,dy.$$

Using this representative avoids dependence on arbitrary changes to $f$ on a null set.

3. Development history and previous partial results

Carleson proved the one-dimensional positive result at $s=1/4$, and Dahlberg and Kenig showed that lower regularity cannot work in general. The one-dimensional problem therefore already included its endpoint.

In higher dimensions, Sjölin and Vega obtained convergence for $s>1/2$. Restriction and bilinear methods gradually lowered the planar threshold. Among the milestones recalled by the manuscripts are $s>2/5$ from Tao’s bilinear restriction theorem and $s>3/8$ from Lee’s work.

Bourgain’s counterexample forces the necessary condition $s\ge n/(2(n+1))$ in dimensions at least two. Crucially, it does not show failure at equality.

Du, Guth and Li proved planar convergence for $s>1/3$ in 2017, combining polynomial partitioning and decoupling. Du and Zhang reached $s>n/(2(n+1))$ in higher dimensions. Their fractal restriction estimates also influenced related geometric problems.

These theorems and Bourgain’s obstruction determined the threshold except for the equality case. The new papers address that remaining case. They also build on the older question of selecting a representative continuous in time outside one spatial exceptional set, studied by Sjögren and Sjölin at higher regularity.

4. What the new manuscripts claim

The endpoint in every dimension at least two

The planar paper treats $f\in H^{1/3}(\mathbb R^2)$. The higher-dimensional paper treats every fixed $n\ge3$ and $f\in H^{s_n}(\mathbb R^n)$.

In either case, the manuscript states that there is a full-measure set $E_f$ of spatial points and a representative $v(x,t)$ such that, for each $x\in E_f$, the function of time is continuous on $[0,1)$, satisfies $v(x,0)=f^*(x)$, and obeys

$$\lim_{a\downarrow0}\sup_{0<t<1} |U_af(x,t)-v(x,t)|=0.$$

Consequently the Gaussian limit exists for every real $0<t<1$ on the same set $E_f$, and

$$\lim_{t\downarrow0}v(x,t)=f^*(x).$$

No radial symmetry, compact Fourier support or additional logarithmic strengthening of $H^{s_n}$ is assumed.

The maximal estimate behind the representative

For bounded-frequency data, the papers establish a local maximal estimate of the form

$$\int_{B(x_0,1)}\sup_{0\le t<1}|e^{it\Delta}f(x)|\,dx \le C_n\|f\|_{H^{s_n}},$$

uniformly in the center and the frequency cutoff. The output is local $L^1$. In particular, the planar paper does not assert the stronger endpoint local $L^3$ maximal estimate.

The higher-dimensional manuscript uses specific information-calculus and time-tree ingredients from the planar companion. It supplies further packet estimates in the required dimensions rather than treating the planar conclusion itself as a higher-dimensional theorem.

5. Why the advance matters

Importance: 5/5. If correct, this family completes the Sobolev threshold for almost-everywhere Schrödinger convergence in all dimensions at least two. It reaches equality at the exponent already singled out by the preceding positive results and counterexamples.

Endpoint questions are structurally different from merely improving a decimal exponent. Away from the endpoint, a small positive regularity margin can absorb losses when frequency pieces are summed. At equality that reserve disappears, forcing a method to preserve the precise square-summation structure of the Sobolev norm.

The common representative is also substantive. A separate almost-everywhere statement for each fixed time does not automatically give one spatial set valid for uncountably many times. The claimed uniform Gaussian limit supplies that missing coherence and makes the limiting evolution precise.

6. The main idea

Geometry of high-frequency packets

Schrödinger wave packets travel along tubes in spacetime. Their directions come from the paraboloid $(\xi,|\xi|^2)$, which records the relation between spatial frequency and temporal oscillation.

To study a time supremum, select a time at each spatial point where the evolution is large. The resulting graph in spacetime carries a measure that reflects ordinary spatial volume. Estimating waves against this measure connects the maximal problem to fractal restriction estimates.

The manuscripts separate transverse interactions, where several packet directions are genuinely independent, from narrow configurations, where directions or positions concentrate near lower-dimensional structures. In transverse regions they claim a definite power saving. Narrow regions require more careful geometry to avoid losing that gain during rescaling.

The planar paper develops a bilinear estimate for smooth packet arrays that remains uniform over the number of scales. The higher-dimensional paper analyzes concentration near polynomial walls, then uses additional curvature tests to organize the relevant portions into approximately affine plates. Estimates in one fewer dimension become applicable after that flattening.

Sum frequencies on one time tree

A high-frequency estimate with an extra factor $N^\varepsilon$ would prove only $H^{s_n+\varepsilon}$ convergence. Reaching $H^{s_n}$ requires removal of that growing loss before the final frequency sum.

The papers use one common selected-time graph for all frequency pieces. A binary tree organizes the time choices, distinguishing a heavier child from a lighter sibling at each split. Spatial localization supplies decay as one moves down the tree, while a weak estimate with output exponent above two supplies a summable advantage for the lighter branches.

The claimed outcome retains a sum of squared frequency contributions, matching the definition of the $H^{s_n}$ norm. A crude sum of their absolute norms would generally impose an extra requirement on the initial data.

Remove regularization simultaneously in time

Once the uniform maximal control is available, smooth approximations can be made close outside suitably small spatial exceptional sets. The paper organizes countably many error envelopes so that the Gaussian approximate identity converges uniformly over the time interval on the remaining full-measure set.

This avoids taking an uncountable intersection of time-dependent full-measure sets. Such an intersection need not have full measure. The order of limits and the common exceptional set are therefore part of the theorem’s construction, not just conventions added afterward.

7. Scope, limitations, and open questions

The result concerns the free Euclidean Schrödinger equation. Potentials, nonlinear terms, boundaries and curved spatial manifolds are different settings, with potentially different thresholds and representatives.

The Gaussian parameter tends to zero before the time limit is taken. The statement does not authorize interchanging these limits arbitrarily or identifying the result with every possible pointwise summation prescription without further argument.

The claimed maximal control is local in space and has $L^1$ output. Stronger endpoint maximal estimates remain separate questions. Almost-everywhere convergence also permits an exceptional spatial null set and provides no pointwise rate of convergence.

Bourgain’s obstruction below $s_n$ explains the sharp regularity threshold for a theorem covering every datum in the Sobolev class. It does not say that every individual function below that threshold must diverge.

8. Sources and further reading

The essential primary comparisons are Bourgain’s necessary condition, Du–Guth–Li’s planar theorem, and Du–Zhang’s higher-dimensional theorem. The two source manuscripts below give the endpoint claims and the exact Gaussian-first representative.

Optional background on Sobolev spaces and Lebesgue points explains the regularity and representative conventions used here.

Source manuscripts

• M1. OpenAI, Endpoint convergence for the planar Schrodinger equation (September 24, 2026).
• M2. OpenAI, Endpoint pointwise convergence for the Schrodinger equation in higher dimensions (September 24, 2026).

Sources consulted on 14 October 2026 at repository revision 1581d4c. Download the manuscript citations.