diff --git a/lib/matplotlib/backends/backend_pgf.py b/lib/matplotlib/backends/backend_pgf.py
index f44975690c84..7250f39b89dd 100644
--- a/lib/matplotlib/backends/backend_pgf.py
+++ b/lib/matplotlib/backends/backend_pgf.py
@@ -212,9 +212,10 @@ def _build_latex_header():
         # fonts later when we don't expect the additional output on stdout.
         # TODO: is this sufficient?
         latex_header = [
+            r"\documentclass{minimal}",
             # Include TeX program name as a comment for cache invalidation.
+            # TeX does not allow this to be the first line.
             r"% !TeX program = {}".format(rcParams["pgf.texsystem"]),
-            r"\documentclass{minimal}",
             latex_preamble,
             latex_fontspec,
             r"\begin{document}",