Sparse bounds for the discrete spherical maximal functions (Q2289360)

The paper under review considers discrete spherical maximal functions defined by \[ A_{\ast}f(x):= \sup \{\lambda ^{2-d} \sum _{n \in \mathbb{Z}^{d};\ |n|=\lambda} f(x-n);\ \lambda^{2} \in \mathbb{N}\} \quad \forall x \in \mathbb{Z}^{d} \] for all \(f \in \ell_{2}(\mathbb{Z}^{d})\) (where \(d \geq 5\)). These are discrete analogues of the spherical maximal function defined by \( \mathcal{A}_{\ast}f:= \sup_{\lambda>0} d\sigma_{\lambda} \star f \) for the normalised surface measure \(d\sigma_{\lambda}\) of a sphere of radius \(\lambda\). In the continuous case, Lacey has obtained in [\textit{M. T. Lacey}, J. Anal. Math. 139, No. 2, 613--635 (2019; Zbl 1433.42016)] sparse bounds in the optimal range of values \((p,q)\). These are estimates of the form \[ \langle \mathcal{A}_{\ast}f,g \rangle \lesssim \sup_{\mathcal{S}\text{ sparse}} \sum _{Q \in S} |Q| (\frac{1}{|Q|}\int _{Q} |f|^{p})^{\frac{1}{p}} (\frac{1}{|Q|}\int _{Q} |g|^{q})^{\frac{1}{q}}, \] where the supremum runs over sparse families of cubes. Establishing sparse bounds has recently become a preferred approach to various problems in harmonic analysis, thanks, in particular, to the pioneering work of \textit{A. K. Lerner} [Int. Math. Res. Not. 2013, No. 14, 3159--3170 (2013; Zbl 1318.42018)]. This is because, on the one hand, proving sparse bounds is often quite natural (a combination of Littlewood-Paley and stopping times arguments reminiscent of \(T(b)\) theorems), and, on the other hand, because sparse bound directly imply various important \(L^{p}\) estimates, including sharp weighted estimates, and vector-valued estimates. The authors obtain sparse bounds for \(A_{\ast}\) for parameters \(p\), \(q\) such that \((\frac{1}{p},\frac{1}{q})\) lies in the interior of the polygon with vertices \(Z_{0} = (\frac{d-2}{d},\frac{2}{d})\), \(Z_{1} = (\frac{d-2}{d},\frac{d-2}{d})\), \(Z_{2} = (\frac{d^{3}-4d^{2}+4d+1}{d^{3}-2d^{2}+d-2},\frac{d^{3}-4d^{2}+6d-7}{d^{3}-2d^{2}+d-2})\), \(Z_{3}=(0,1)\) (and an endpoint estimate). Interestingly, this region is slightly smaller than its continuous counterpart for \(\mathcal{A}_{\ast}\), and conjectured to be optimal. A counterexample supporting this conjecture is given in Section 5. The proof given by the authors is striking. Like all sparse domination arguments, it uses stopping times \(\tau\) (selecting an appropriate radius \(\lambda\) given \(f\) and \(g\)) and induction. The terms corresponding to small values of \(\tau\) can be estimated using the continuous result (applied to appropriate piecewise constant functions). For large values of \(\tau\), the authors then fundamentally rely on a version of the Hardy-Littlewood circle method decomposition developed in [\textit{Á. Magyar} et al., Ann. Math. (2) 155, No. 1, 189--208 (2002; Zbl 1036.42018)]. It decomposes \(A_{\tau}f\) (using Fourier multipliers) in pieces localised in regions of the sphere around points of the form \(\frac{\ell}{q}\) where \(q \leq \tau\) and \(\ell \in \mathbb{Z}_{q}^{d}\). Terms that correspond to large values of \(q\) can be directly estimated using \(L^2\) bounds from [Magyar et al., loc. cit.]. Some of the terms that correspond to small values of \(q\) can be estimated using a generic result about maximal functions based on Fourier multipliers, proved by \textit{N. Lohoué} [C. R. Acad. Sci., Paris, Sér. I 312, No. 8, 561--566 (1991; Zbl 0746.58073)]. This leaves the most challenging terms, for which one needs to exploit particularly strong cancellation properties. Fortunately, these terms involve Ramanujan sums of the form \[ m\mapsto \sum _{a \in \mathbb{Z}_{q} ^{\times}} \exp(2i\pi \frac{am}{q}), \] for which subtle estimates are known; see Lemma 2.25 and the core of the proof just after it (tertiary decomposition).