Multiple testing of local maxima for detection of peaks on the (celestial) sphere (Q2278653)

The authors extend the STEM procedure of \textit{D. Cheng} and \textit{A. Schwartzman} [Ann. Stat. 45, No. 2, 529--556 (2017; Zbl 1369.62144)] to the case of a topological multiple testing scheme for detecting peaks on the sphere under isotropic Gaussian noise where tests are performed at local maxima of the observed field filtered by the spherical needlet transform. They also develop a statistical procedure to control the probability of observing false detections, the fraction of false detections (FDP), and the expected FDP in a population. The performance of the proposed procedures is explored numerically for simulated cosmic microwave background radiation fields. The authors study the sequence of signal-plus-noise models \[ y_{N}(x) = \mu_{N}(x)+ z(x),~x \in \mathbb{S}^{2}, ~N=1,2,\dots \] where \(z\) is a Gaussian, zero-mean, isotropic field on the unit sphere \(\mathbb{S}^{2}\) expressed as \[ z(x)=\sum_{l=1}^{\infty}z_{l}(x),~z_{l}(x)=\sum_{m=-l}^{l}a_{l,m}Y_{l,m}(x) \] with \(Y_{l,m}\) standing for the spherical harmonics, and \(\mu_{N}\) is a sequence of deterministic functions defined by \[ \mu_{N}(x) =\sum_{k=1}^{N}a_{k}h(x, t_{k,N}, \xi_{k}) \] with \(a_{k}>0\) \((k=1, 2, \dots),\) \[ h(x, t, \xi)=\sum_{l=1}^{\infty}\frac{2l+1}{4\pi}e^{-l(l+1)t}P_{l}(\langle x,\xi\rangle), \] \[ P_{l}(u) = \frac{(-1)^{l}}{2^{l}}\frac{d^{l}}{du^{l}}(1 -u^{2})^{l},~ l= 1, 2, \dots, \] and \(\langle \cdot,\cdot\rangle\) standing for the scalar product. It is assumed that \(t_{k,N} \rightarrow 0\) as \(N\rightarrow \infty\). The observable data is \[ \hat y_{N_j}(x) = \hat \mu_{N_j}(x)+ \hat z_{j}(x),~x \in \mathbb{S}^{2}, ~N_j=1,2,\dots \] where \[ \hat z_{j}(x)=\sum_{l=1}^{\infty}b\left(\frac{l}{B^{j}},s\right)z_{l}(x) \] and \[ \hat \mu_{N_{j}}(x) =\sum_{k=1}^{N} \sum_{l=1}^{\infty}a_{k}b\left(\frac{l}{B^{j}},s\right)\frac{2l+1}{4\pi}e^{-l(l+1)t_{k,N}}P_{l}(\langle x,\xi_{k}\rangle) \] with \(b(u, s) = u^{2s}e^{-u^{2}}\) and bandwidth parameter \(B>1\). The STEM algorithm is as follows: \begin{itemize} \item[1.] Normalize \(\hat y_{N_j}\) to obtain \(\tilde y_{N_j} = \frac{\hat y_{N_j}}{\sqrt{\mathbb{E}[\hat z_{j}^{2}]}}\), where the denominator is assumed to be known. \item[2.] Find the set of local maxima: \[ T_{j} = \{x \in \mathbb{S}^{2} : \nabla\tilde y_{N_j}(x)=0, \nabla^{2}\tilde y_{N_j}(x) \text{ is negative definite}\}. \] \item[3.] For each \(x \in T_{j}\), compute the \(p\)-value of the hypothesis ``\(\mu_{j}(x')=0\) for all \(x'\in B(x,\rho_{j})\)'' where \(B(x,\rho_{j})\) is a geodesic ball centered at \(x\) on the sphere and of radius \(\rho_{j} \sim j^{\nu}B^{-j}\) for some \(\nu>0\). \item[4.] Apply the multiple testing procedure of \textit{Y. Benjamini} and \textit{Y. Hochberg} [J. R. Stat. Soc., Ser. B 57, No. 1, 289--300 (1995; Zbl 0809.62014)] on the set of \(p\)-values obtained, and declare significant all local maxima whose \(p\)-values are smaller than the significance threshold. \end{itemize} The authors obtain explicit upper bounds for the FDP and the expected FDP, and show that the proportion of detected peaks tends to 1 asymptotically. Numerical analysis using simulated cosmic microwave background radiation fields is then performed to assess the accuracy of the analytic approximation for the \(p\)-values in Step 3 of STEM, and the tightness of the upper bounds for FDP and expected FDP.