Technical Summary: Gauge Transforms, Random Averaging Operator Ansatz, and Improved Probabilistic Well-Posedness for the Radial NLS on the 3d Ball

1. Problem Statement and Context

The paper addresses the defocusing cubic nonlinear Schrödinger equation (NLS) on the three-dimensional unit ball $B^3$ with Dirichlet boundary conditions and radial initial data:
$$(i\partial_t + \Delta)u = |u|^2u, \quad u|_{\partial B^3} = 0.$$
The study focuses on the probabilistic well-posedness of this equation when the initial data is randomized according to a Gaussian measure $\mu_\alpha$ supported on Sobolev spaces $H^s_{rad}(B^3)$ with $s < \alpha - 1/2$. The initial data is defined via a random Fourier series:
$$\phi_\alpha(\omega, x) = \sum_{n \ge 1} \frac{g_n(\omega)}{n^\alpha} e_n(x),$$
where $g_n$ are independent standard complex Gaussian variables and $e_n$ are radial eigenfunctions of the Dirichlet Laplacian.

The Challenge:
In three dimensions, the critical Sobolev regularity for the NLS is $s=1/2$. The support of the Gibbs measure corresponds to data with regularity slightly below $1/2$ (specifically $s < 1/2$ when $\alpha=1$). Previous work by Bourgain and Bulut [BB14] established almost sure global existence and uniqueness for $\alpha=1$ (data in $H^{1/2-}$). However, their approach relied on a finite-dimensional approximation where the nonlinearity was truncated by a spectral projector $\Pi_N$ on the right-hand side. This method yielded a convergence rate with a logarithmic divergence in the frequency parameter $N$, limiting the ability to handle data with lower regularity (i.e., $\alpha < 1$) or to treat the full equation without the projector.

The paper aims to improve these results by establishing almost sure local well-posedness for $\alpha > 15/16$, a regime that is supercritical with respect to the probabilistic scaling introduced by Deng, Nahmod, and Yue [DNY24a, DNY22].

2. Methodology and Key Innovations

The authors employ a combination of gauge transformations, random averaging operators, and refined modulation analysis to overcome the singularities inherent in the radial 3D setting.

2.1. Decomposition of the Nonlinearity and Gauge Transform

The core difficulty identified in [BB14] is a specific resonant interaction (high $\times$ high $\times$ high) that produces a logarithmic divergence when $\alpha=1$. In the radial setting, the Fourier coefficients $u_n$ satisfy an equation where a divergent real-valued coefficient $C_n(u_N)$ depends on the frequency $n$:
$$(i\partial_t - n^2) \hat{u}_N(n) = C_n(u_N) \hat{u}_N(n) + \text{other terms}.$$
Unlike the torus case, where global gauge transformations can eliminate such terms, the frequency-dependent nature of $C_n$ in the ball prevents a single global transformation.

The Solution: The authors introduce a refined gauge transformation adapted to the specific algebraic structure of the divergent coefficient. They define a modified dispersion relation $\lambda_n^2 = n^2 + \mu_n^2$, where $\mu_n^2$ captures the expected value of the resonant interaction. By subtracting the linear part associated with $\mu_n^2$ and utilizing cancellations in the probabilistic framework (specifically $|u_n(t)|^2 - |u_n(0)|^2$), they convert the problematic resonant contribution into non-resonant quintic terms. This transformation is linear in the frequencies, contrasting with the nonlinear gauge transforms used in other contexts (e.g., [OTW20]).

2.2. Random Averaging Operator Ansatz

To handle the remaining high $\times$ low frequency interactions, the authors utilize the random averaging operator framework introduced by Deng, Nahmod, and Yue [DNY24a].

• Decomposition: The solution $u_N$ is decomposed into a "colored" Gaussian term $\psi_N$ (capturing the most singular high $\times$ low interactions) and a smoother remainder $w_N$.
• Radial Reduction: In the radial setting, the random averaging operators reduce to multiplication by complex phases. Specifically, for a mode $n$, the operator acts as $H_n(t) = \exp(-i \int_0^t \Theta_n(\tau) d\tau)$, where $\Theta_n$ is a real-valued function determined by lower-frequency components.
• Induction Scheme: The proof proceeds via an induction scheme over dyadic frequency scales. At each scale $N$, the authors construct the random averaging phase $H_N$ and the remainder $w_N$ using fixed-point arguments in Fourier-Lebesgue spaces ($X^{s,b}$ spaces adapted to the twisted Laplacian).

2.3. Modulation Analysis and New Estimates

The analysis is conducted entirely in frequency space, reflecting the one-dimensional nature of the radial problem. Key technical contributions include:

• New Strichartz-type Estimates: The authors derive refined counting estimates for the resonance function $\Omega(\vec{n}) = \lambda_{n_1}^2 - \lambda_{n_2}^2 + \lambda_{n_3}^2 - \lambda_{n_4}^2$ on the ball. These estimates account for the perturbed dispersion relation $\lambda_n^2 = n^2 + \mu_n^2$ and the boundary effects.
• Tensor and Operator Bounds: They establish deterministic and probabilistic bounds for multilinear interactions (trilinear and quintic). Crucially, they prove that the random averaging operators allow for the absorption of the most singular high $\times$ low interactions, leaving only terms that can be controlled perturbatively.
• Handling the Logarithmic Divergence: By combining the gauge transform with the random averaging ansatz, the authors derive an estimate for the difference of solutions $x_{N,M}(t)$ of the form:
  $$x_{N,M}(t) \le C N^{-\delta} + C \int_0^t x_{N,M}(\tau) d\tau,$$
  which lacks the logarithmic divergence ($\log N$) present in the [BB14] estimate. This allows the application of Gronwall's inequality to prove convergence.

3. Main Results

Theorem 1.2 (Almost Sure Local Well-Posedness):
Let $\alpha > 15/16$. There exists a measurable set $\Sigma_\alpha$ with $\mu_\alpha(\Sigma_\alpha) = 1$ such that for every initial data $\phi \in \Sigma_\alpha$, there exists a time $T > 0$ for which the sequence of solutions $(u_N)_{N \ge 1}$ to the frequency-truncated equation (evolving under the full cubic nonlinearity) converges in $L^\infty([-T, T]; H^{\alpha - 1/2 -}_{rad}(B^3))$ to a unique limit $u$. This limit solves the cubic NLS in the distributional sense.

Theorem 1.3 (Quantitative Statement):
The convergence holds on a time interval $[-\tau_R, \tau_R]$ with $\tau_R = R^{-C_1}$, where the probability of the initial data failing to satisfy the necessary bounds is exponentially small ($< C_1 e^{-c_1 R^{\delta_0}}$). The solution admits a decomposition $u = \psi_\infty + w_\infty$, where $\psi_\infty$ is the limit of the colored Gaussian terms and $w_\infty$ is a smoother remainder in $C([-T, T]; H^{s_0}_{rad})$ for some $s_0 > 1/2$.

4. Significance and Claims

The paper claims to be a significant improvement over the results of Bourgain and Bulut [BB14] in three dimensions.

• Supercritical Regime: The result establishes well-posedness for initial data with regularity strictly below the typical regularity of the Gibbs measure support (for $\alpha < 1$), placing the problem in a supercritical regime relative to the probabilistic scaling of [DNY24a].
• Removal of the Projector: Unlike [BB14], which required the nonlinearity to be truncated by $\Pi_N$ (a finite-dimensional approximation), this work proves convergence for the full cubic equation. The limit is shown to be the same for both the truncated and full approximations.
• Elimination of Logarithmic Divergence: The primary technical achievement is the elimination of the $\log N$ divergence in the convergence estimates. This is achieved through the novel combination of gauge transformations (to handle resonant high-high-high interactions) and random averaging operators (to handle high-low interactions).
• Radial Geometry: The work extends the probabilistic well-posedness theory to bounded domains with boundary, specifically the 3D ball, where boundary effects and the lack of translation invariance complicate the analysis compared to the torus.

The authors note that the constraint $\alpha > 15/16$ is not optimal (they conjecture $\alpha > 3/4$ is the true threshold) but is sufficient to surpass the critical threshold $\alpha=1$. They also remark that the case of dimensions $d \ge 4$ remains out of reach due to energy-criticality and derivative losses in Strichartz estimates on spheres.

5. Conclusion

This work provides a robust framework for handling low-regularity probabilistic solutions to the NLS on bounded domains. By refining the analysis of nonlinear interactions through gauge transforms and random averaging, the authors successfully navigate the supercritical regime, offering a stronger convergence result than previously available for the radial 3D NLS. The methodology suggests a path forward for similar problems where standard dispersive smoothing is insufficient.