8:42 a.m.
Dear All,
To reproduce
x = {1,1}'
catch eval numhess(x,x'*x)
catch eval fdjac(x,x'*x)
catch eval fdjac(x,(x'*x))
function matrix i(matrix x)
return x
end function
eval numhess(x,i(x'*x))
eval fdjac(x,i(x'*x))
Why it may look trivial, I spent
some time to guess to use
identity function
I can imagine 2 possibilities:
1) to include native identity
function in hansl
2)fdjac and numhess should do
something similar inside
???
Oleh
9:11 a.m.
On Wed, 4 Oct 2017, oleg_komashko(a)ukr.net wrote:
Ok, first things first: DON'T EVER use the Matlab-style syntax "x'*x";
it
works, but it's inefficient, both in terms of memory usage and CPU time.
Use "x'x" instead. Believe me, it does make a difference. And if you
don't
believe me, try this:
<hansl>
set verbose off
X = mnormal(2000,100)
XX = zeros(100,100)
set stopwatch
loop 100 -q
XX = X'*X
endloop
t0 = $stopwatch
loop 100 -q
XX = X'X
endloop
t1 = $stopwatch
print t0 t1
</hansl>
Having said this, fdjac and numhess are supposed to work on _functions_,
not _arbitrary expressions_. Of course Allin has the final word here, but
I imagine that allowing for arbitrary genr expressions as the second
argument would complicate the parsing quite a bit, and I don't really see
the scope for it. But I may be wrong, of course.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
9:32 a.m.
Use "x'x" instea several packd - Thank you. Sometimes
ago
I rewrote several packages computing quadratic forms
and quadratic forms with inverse. x'x was my gap
Having said this, fdjac and numhess are supposed to work on
_functions_,
not _arbitrary expressions
arbitrary expressions are turned into functions via identity function,
so, may be make native identity function in hansl?
Oleh
4 жовтня 2017, 12:11:45, від "Riccardo (Jack) Lucchetti"
<r.lucchetti(a)univpm.it>:
7:52 p.m.
On Wed, 4 Oct 2017, oleg_komashko(a)ukr.net wrote:
Turns out that it's easy enough to support the pattern you gave on
your third "eval" attempt above -- that is, you put the raw
expression into parentheses and we can then treat it as an
"anonymous inline function":
<hansl>
x = {1,1}'
eval numhess(x,(x'x))
eval fdjac(x,(x'x))
</hansl>
That's now in git. In principle we could parse the unparenthesized
case, but it would be a lot more fiddly and error-prone.
Allin
8:23 p.m.
Dear Allin,
1)
<hansl>
x = {1,1}'
eval numhess(x,(x'x))
eval fdjac(x,(x'x))
</hansl>
Thank you!!!
As for me, my problem is 100% solved!
2) I think its a very good idea
to consider grad and Hessian via
complex numbers
The formula is trivial, one step
dy/dx ~ Im((y(x0)-y(x0+i*eps))/eps)
The trick is in that the preciseness
does not depends on real x!
So eps ~ 3*10^-16!
clib contains all elementary
functions of complex argument.
To make it fully functional
it is sufficient to find a c code
for normal cdf for complex
argument and basic matrix
operations for complex arguments.
If it could be realized in c
sufficiently quickly it would be
a real bomb: no need for
automatic/symbolic differentiation!
Oleh
5:57 a.m.
On Wed, 4 Oct 2017, oleg_komashko(a)ukr.net wrote:
I'm working on a function package which, among other things, will include
support for complex matrices. If you want, I can send you a pre-release
version.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
6:38 a.m.
It is very interesting,
But this is a somewhat different
question. For gradient and
hessian hansl support is not
needed: all functions are
inside c code: gretl input
are real function
c output is real number
matrix. The question is
the speed.
The script attached
illustrates the preciseness
It does not use {numDeriv}
only R shared library
But it is too slow to
be practical
The Hessian uses the
simplest finite difference
formula
Oleh
5 жовтня 2017, 08:59:28, від "Riccardo (Jack) Lucchetti"
<r.lucchetti(a)univpm.it>:
6:58 a.m.
On Thu, 5 Oct 2017, oleg_komashko(a)ukr.net wrote:
It is very interesting,
But this is a somewhat different
question.
I'm sorry, I don't think I understand what you're talking about. But never
mind. :)
However, if anybody on the list needs complex matrices, the corresponding
funciton package should be out Real Soon Now.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
7:04 a.m.
What I mean by functionality:
To compute grads for vast majority
of functions of the real argument
the c code inside
should support these functions
for complex arguments:
to differentiate , say exp(x^2 + x) + sin(log(x^2+1)
the c code should support exp(), sin(), log()
and polynomials for the complex argument
Oleh
5 жовтня 2017, 09:58:18, від "Riccardo (Jack) Lucchetti"
<r.lucchetti(a)univpm.it>:
2739
days inactive
2740
days old
9 comments
3 participants
participants (3)
-
Allin Cottrell -
oleg_komashko＠ukr.net -
Riccardo (Jack) Lucchetti