Add email comment explicitly authorizing distributions/special_math.py be released under Apache 2.0.

PiperOrigin-RevId: 215296386

1 files changed, 59 insertions, 2 deletions

diff --git a/tensorflow/python/ops/distributions/special_math.py b/tensorflow/python/ops/distributions/special_math.py
index 31b7a36fd3..ccc667cae3 100644
--- a/tensorflow/python/ops/distributions/special_math.py
+++ b/tensorflow/python/ops/distributions/special_math.py
@@ -12,6 +12,62 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+
+# Functions "ndtr" and "ndtri" are derived from calculations made in:
+# https://root.cern.ch/doc/v608/SpecFuncCephesInv_8cxx_source.html
+# In the following email exchange, the author gives his consent to redistribute
+# derived works under an Apache 2.0 license.
+#
+# From: Stephen Moshier <steve@moshier.net>
+# Date: Sat, Jun 9, 2018 at 2:36 PM
+# Subject: Re: Licensing cephes under Apache (BSD-like) license.
+# To: rif <rif@google.com>
+#
+#
+#
+# Hello Rif,
+#
+# Yes, Google may distribute Cephes files under the Apache 2 license.
+#
+# If clarification is needed, I do not favor BSD over other free licenses.
+# I would agree that Apache 2 seems to cover the concern you mentioned
+# about sublicensees.
+#
+# Best wishes for good luck with your projects!
+# Steve Moshier
+#
+#
+#
+# On Thu, 31 May 2018, rif wrote:
+#
+# > Hello Steve.
+# > My name is Rif. I work on machine learning software at Google.
+# >
+# > Your cephes software continues to be incredibly useful and widely used. I
+# > was wondering whether it would be permissible for us to use the Cephes code
+# > under the Apache 2.0 license, which is extremely similar in permissions to
+# > the BSD license (Wikipedia comparisons). This would be quite helpful to us
+# > in terms of avoiding multiple licenses on software.
+# >
+# > I'm sorry to bother you with this (I can imagine you're sick of hearing
+# > about this by now), but I want to be absolutely clear we're on the level and
+# > not misusing your important software. In former conversation with Eugene
+# > Brevdo (ebrevdo@google.com), you wrote "If your licensing is similar to BSD,
+# > the formal way that has been handled is simply to add a statement to the
+# > effect that you are incorporating the Cephes software by permission of the
+# > author." I wanted to confirm that (a) we could use the Apache license, (b)
+# > that we don't need to (and probably you don't want to) keep getting
+# > contacted about individual uses, because your intent is generally to allow
+# > this software to be reused under "BSD-like" license, and (c) you're OK
+# > letting incorporators decide whether a license is sufficiently BSD-like?
+# >
+# > Best,
+# >
+# > rif
+# >
+# >
+# >
+
 """Special Math Ops."""
 
 from __future__ import absolute_import
@@ -135,7 +191,7 @@ def _ndtri(p):
 
   # Constants used in piece-wise rational approximations. Taken from the cephes
   # library:
-  # https://github.com/scipy/scipy/blob/master/scipy/special/cephes/ndtri.c
+  # https://root.cern.ch/doc/v608/SpecFuncCephesInv_8cxx_source.html
   p0 = list(reversed([-5.99633501014107895267E1,
                       9.80010754185999661536E1,
                       -5.66762857469070293439E1,
@@ -305,7 +361,8 @@ def log_ndtr(x, series_order=3, name="log_ndtr"):
     else:
       raise TypeError("x.dtype=%s is not supported." % x.dtype)
 
-    # The basic idea here was ported from py/scipy/special/cephes/ndtr.c.
+    # The basic idea here was ported from:
+    #   https://root.cern.ch/doc/v608/SpecFuncCephesInv_8cxx_source.html
     # We copy the main idea, with a few changes
     # * For x >> 1, and X ~ Normal(0, 1),
     #     Log[P[X < x]] = Log[1 - P[X < -x]] approx -P[X < -x],