1: \begin{proof}

2: Formally substituting $\alpha=\lam_j$ and $N(\alpha)=j$ takes us from the first asymptotic to the second. The difficulty with making this substitution rigorous is that if $\lam_j$ is a multiple eigenvalue, then $N(\lam_j)$ can exceed $j$.

3: 

4: To circumvent the problem, we argue as follows. Given $\e>0$ we know from $N(\alpha) \sim \alpha/c$ that

5: \[

6: (1-\e) \frac{\alpha}{c} < N(\alpha) < (1+\e) \frac{\alpha}{c}

7: \]

8: for all large $\alpha$. Substituting $\alpha=\lam_j$ into the right hand inequality implies that

9: \[

10: j < (1+\e) \frac{\lam_j}{c}

11: \]

12: for all large $j$. Substituting $\alpha=\lam_j-\delta$ into the left hand inequality implies that

13: \[

14: (1-\e) \frac{\lam_j-\delta}{c} < j

15: \]

16: for each large $j$ and $0<\delta<1$, and hence (by letting $\delta \to 0$) that

17: \[

18: (1-\e) \frac{\lam_j}{c} \leq j .

19: \]

20: We conclude that

21: \[

22: \frac{1}{1+\e} < \frac{\lam_j}{cj} \leq \frac{1}{1-\e}

23: \]

24: for all large $j$, so that

25: \[

26: \lim_{j \to \infty} \frac{\lam_j}{cj}  = 1

27: \]

28: as desired.

29: \end{proof}

30: