1: \begin{proof}[\bf Proof of \eqref{eq:lambdabar}]

2: To derive the above asymptotic result rigorously, we use the notations in \cite{Kato66} and invoke (3.2), (3.3), (3.4) and (3.6) in Chap.\ Two of \cite{Kato66} to bound the residual term in the asymptotic expansion. Under Assumption~\ref{ass:span}, there is a constant gap between $\lambda_r(\bV^* \bW^* (\bV^*))$ and $\lambda_{r+1}(\bV^* \bW^* (\bV^*))$. Since $\bW \xrightarrow{p} \bW^*$, we deduce that with probability $1-o(1)$, the gap between $\lambda_r(\bV^* \bW (\bV^*))$ and $\lambda_{r+1}(\bV^* \bW (\bV^*))$ is bounded away from zero by a constant. Thus, choosing any circle with constant radius enclosing $\lambda=0$, we have with probability $1-o(1)$,  $\rho$ in (3.4) and $\max_{\zeta \in \Gamma}\| R(\zeta) \|$ and both bounded by a constant. Since $\| \bT^{(1)} \| = O_P(1)$, $\| \bT^{(2)} \| = O_P(2)$ and $\bT^{(n)}$ ($n\ge3$) are simply zero, (3.2) is satisfied with probability $1-o(1)$ for $|x| < r_n$ where $r_n>0$ is any vanishing sequence. This implies with probability $1-o(1)$, $r_0$ in (3.3) satisfies $r_0\ge r_n$. We fix the `$n$' in (3.6) by $2$ (not to confuse with our sample size $n$). Then, the upper bound in (3.6) is $O(n^{-3/2} / r_0^3) = O(n^{-1})$ with the choice, say, $r_n = n^{-1/10}$. Therefore, we conclude that the residual term in the second-order expansion of $\lambda(\omega_n)$ is $o_P(\omega_n^2)$.

3: \end{proof}

4: