1: \begin{abstract}

2: The data augmentation (DA) algorithm is a widely used Markov chain

3: Monte Carlo algorithm that is easy to implement but often suffers

4: from slow convergence. The sandwich algorithm is an alternative

5: that can converge much faster while requiring roughly the same

6: computational effort per iteration. Theoretically, the sandwich

7: algorithm always converges at least as fast as the corresponding DA

8: algorithm in the sense that $\norm{K^*} \le\norm{K}$, where $K$ and

9: $K^*$ are the Markov operators associated with the DA and sandwich

10: algorithms, respectively, and \mbox{$\Vert\cdot\Vert$} denotes operator norm.

11: In this paper, a substantial refinement of this operator norm

12: inequality is developed. In particular, under regularity conditions

13: implying that $K$ is a trace-class operator, it is shown that $K^*$

14: is also a positive, trace-class operator, and that the spectrum of

15: $K^*$ dominates that of $K$ in the sense that the ordered elements

16: of the former are all less than or equal to the corresponding

17: elements of the latter. Furthermore, if the sandwich algorithm is

18: constructed using a group action, as described by

19: Liu and Wu [\textit{J. Amer. Statist. Assoc.} \textbf{94} (1999)

20: 1264--1274] and Hobert and Marchev

21: [\textit{Ann. Statist.} \textbf{36} (2008)

22: 532--554], then there is strict

23: inequality between at least one pair of eigenvalues. These results

24: are applied to a new DA algorithm for Bayesian quantile regression

25: introduced by Kozumi and Kobayashi

26: [\textit{J. Stat. Comput. Simul.} \textbf{81} (2011) 1565--1578].

27: \end{abstract}