The largest sample eigenvalue distribution in the rank 1 quaternionic spiked model of Wishart ensemble (Q838000)

Consider an \(N\)-variate normal distribution with mean \(0\) and covariance matrix \(\Sigma\) over the quaternions. Then \(\Sigma\) is a quaternionic Hermitian matrix and without loss in generality may be assumed to be diagonal. Let \(X\) be an \(N\times M\) quaternionic matrix whose columns are \(M\) independent observations from this distribution. The eigenvalues \(\lambda=(\lambda_{1},\dots,\lambda_{N})\) of the sample matrix \(S:=\frac{1} {M}XX^{\ast}\) (where \(\ast\) denotes the quaternionic Hermitian transpose) are called the sample eigenvalues and the probability space of the \(\lambda_{i}\) is called the Wishart ensemble. The case of interest is where \(M\) and \(N\) are of comparable size and \(N\rightarrow\infty\). When \(\Sigma=I\) we have the Laguerre ensemble (LSE) and this has been well studied. The ``spiked model'' is the case where the first \(k\) diagonal entries of \(\Sigma\) are of the form \(1+a_{i}\) and the remaining diagonal entries are \(1\). For the analogous situation over the complex field, it was shown by \textit{J. Baik, G. Ben Arous} and \textit{S. Péché} [Ann. Probab. 33, No.~5, 1643--1697 (2005; Zbl 1086.15022)] that if \(\max~(a_{1},\dots,a_{k})\) is smaller than a certain threshold \(\gamma^{-1}\), the distribution of the largest sample eigenvalue is the same as for that for \(\Sigma=I\), but for larger values the distribution is Gaussian with mean and variance depending on \(\max(a_1,\dots,a_k)\). The present paper considers the rank \(1\) (that is, \(k=1\), \(a_{1}=a\)) spiked model for the quaternionic case. The main theorem states that with the same threshold \(\gamma^{-1}\) as the complex case, in the rank \(1\) quaternionic spiked model we have: (1) if \(-1<a<\gamma^{-1}\), then the distribution of the largest sample eigenvalue is the same as for the case \(\Sigma=I\); (2) if \(a=\gamma^{-1}\), then the limit and fluctuation scale are the same as for \(\Sigma=I\), but the distribution function is different; and (3) if \(a>\gamma^{-1}\) then limit, fluctuation scale and distribution are all changed. In (2) the distribution is the GOE Tracy-Widom distribution and in (3) the distribution is a specified Gaussian distribution. In view of the results for the rank \(1\) spiked model for the complex and quaternionic cases, it is conjectured that the same theorem will hold in the real case (which so far has remained open).