summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorJaron Kent-Dobias <jaron@kent-dobias.com>2024-10-25 19:00:16 +0200
committerJaron Kent-Dobias <jaron@kent-dobias.com>2024-10-25 19:00:16 +0200
commitd73d3cdc03337fde998db900ed3232151e75f729 (patch)
tree5a2d67591eef5b0cb7bfd84ae2f3dfbb99d29c17
parent157b8b12bdad4646773d1c596f99af6a1b4d9c9d (diff)
downloadmarginal-d73d3cdc03337fde998db900ed3232151e75f729.tar.gz
marginal-d73d3cdc03337fde998db900ed3232151e75f729.tar.bz2
marginal-d73d3cdc03337fde998db900ed3232151e75f729.zip
Clarified the definitions of marginal minima and pseudogap in the
opening paragraphs
-rw-r--r--marginal.tex5
1 files changed, 4 insertions, 1 deletions
diff --git a/marginal.tex b/marginal.tex
index acd3d4a..d9e1d47 100644
--- a/marginal.tex
+++ b/marginal.tex
@@ -61,7 +61,10 @@ dynamics would get stuck at a specific energy level, called the threshold
energy. The threshold energy is the energy level at which level sets of the
landscape transition from containing mostly saddle points to containing mostly
minima. The level set associated with this threshold energy contains mostly \emph{marginal
-minima}, or minima that have a pseudogap in the spectrum of their Hessian.
+minima}, or minima whose Hessian matrix has a continuous spectral density over
+all sufficiently small positive eigenvalues. In most circumstances the spectrum
+is \emph{pseudogapped}, which means that the spectral density smoothly
+approaches zero as zero eigenvalue is approached from above.
However, recent work found that the threshold energy is not important even for
simple gradient descent dynamics \cite{Folena_2020_Rethinking, Folena_2023_On, ElAlaoui_2020_Algorithmic}.