minor corrections to derivatives.rst
Change-Id: I6bb5e3c507113c8873c4c516e9cae3239cca247c
diff --git a/docs/source/derivatives.rst b/docs/source/derivatives.rst
index 3919f86..3838d43 100644
--- a/docs/source/derivatives.rst
+++ b/docs/source/derivatives.rst
@@ -651,8 +651,8 @@
Forward Differences and the remarkable accuracy improvements of
Ridders' method cost an order of magnitude more runtime.
-Recommendation
---------------
+Recommendations
+---------------
Numeric differentiation should be used when you cannot compute the
derivatives either analytically or using automatic differention. This
@@ -929,6 +929,7 @@
Indeed, this is essentially how :class:`AutoDiffCostFunction` works.
+
Pitfalls
--------
@@ -992,14 +993,12 @@
TODO
====
-#. Inverse function theorem
-#. Add references in the various sections about the things to
- do. NIST, RIDDER's METHOD, Numerical Recipes.
-#. Calling iterative routines.
+#. Why does the quality of derivatives matter?
#. Discuss, forward v/s backward automatic differentiation and
relation to backprop, impact of large parameter block sizes on
differentiation performance.
-#. Why does the quality of derivatives matter?
+#. Inverse function theorem
+#. Calling iterative routines.
#. Reference to how numeric derivatives lead to slower convergence.
#. Pitfalls of Numeric differentiation.
#. Ill conditioning of numeric differentiation/dependence on curvature.
