4
$\begingroup$

I've defined a set of rules, which I want to use with FullSimplify, as follows:

transf1[expr_] := expr /. (a_⊕b_)⊕c_ -> a⊕(b⊕c)
transf2[expr_] := expr /. a_⊕b_ -> b⊕a
transf3[expr_] := expr /. a_⊕-a_ -> 0
transf4[expr_] := expr /. a_⊕0 -> a
transf5[expr_] := expr /. a_⊖b_ -> a⊕-b

But when I try to actually use them, I get no result:

FullSimplify[x⊕y⊖x,TransformationFunctions->{transf1,transf2,transf3,transf4,transf5}]

x⊕(y⊖x)

At the same time, there does exist a sequence of transformations, which does yield a simpler result:

transf4@transf3[transf2 /@ transf1@transf2@transf5[x⊕y⊖x]]

y

Am I using FullSimplify wrongly? Isn't the above the correct way to provide custom transformation functions?

$\endgroup$
  • $\begingroup$ FullSimplify won't evaluate all possible compositions.... perhaps a modified ComplexityFunction is necessary to guide it through a good path $\endgroup$ – Dr. belisarius Feb 29 '16 at 21:00
  • $\begingroup$ See 69262 and 96633. $\endgroup$ – Sjoerd C. de Vries Feb 29 '16 at 21:25
3
$\begingroup$

Here is a analysis of the problem which doesn't give a solution :

You can follow what FullSimplifyis doing with this modified code :

transf1[expr_] := (Print[1, " ", expr, " "];expr /. (a_⊕b_)⊕c_ -> a⊕(b⊕c))
transf2[expr_] := (Print[2, " ", expr, " "];expr /. a_⊕b_ -> b⊕a)
transf3[expr_] := (Print[3, " ", expr, " "];expr /. a_⊕-a_ -> 0)
transf4[expr_] := (Print[4, " ", expr, " "];expr /. a_⊕0 -> a)
transf5[expr_] := (Print[5, " ", expr, " "];expr /. a_⊖b_ -> a⊕-b)

FullSimplify[x\[CirclePlus]y\[CircleMinus]x, 
 TransformationFunctions -> {transf1, transf2, transf3, transf4, 
   transf5}]

enter image description here

You see that the transformation N°5 is tried first. It should give x+(y+(-x)), but you see nowhere this expression in the following lines. I presume therefore that the transformation N°5 is not used. There is a explanation for that : The ComplexityFunction, which is similar to LeafCount is worse with x+(y+(-x)) than with x+(y-x).

The problem is that one can hardly find a ComplexityFunction that get better trough the sequence that gives the solution :

enter image description here

| improve this answer | |
$\endgroup$
  • $\begingroup$ Does FullSimplify require that ComplexityFunction monotonically decreases on each transformation? $\endgroup$ – Ruslan Mar 1 '16 at 5:29
  • $\begingroup$ @Ruslan Excellent question, but I don't know the answer. Maybe the answer can be deducted from here ? All the possibilities form a tree of possibilities. I don't know how FullSimplifygo through the tree. I don't even know what is the "current state" that memorize what has already been tried (it could be something like a list of candidates expressions to further optimize) $\endgroup$ – andre314 Mar 1 '16 at 7:40
  • $\begingroup$ @Rusian I fact, I was very surprised to see how basic is the bevavior of FullSimpllify in your very interesting example (where the expressions are inert) $\endgroup$ – andre314 Mar 1 '16 at 7:47
2
$\begingroup$

Instead of defining transformation rules you can try to achieve the same by defining your operations:

CircleMinus[a_, b_] := CirclePlus[a, -b]
CirclePlus[a_, 0] := a
CirclePlus[a_, b_] := If[a == -b, 0]
SetAttributes[CirclePlus, {Orderless, Flat}]

and using a simple Simplify

Simplify[CirclePlus[x, CircleMinus[y, x]], Assuming[x != 0, y != 0]]
(* 0 *)

Correction

Indeed, there were problems with my solution. Now they are corrected:

ClearAll[CirclePlus, CircleMinus]
CircleMinus[a_, b_] := CirclePlus[a, -b]
CirclePlus[a_, 0] := a
CirclePlus[a_] := a
CirclePlus[AA___, a_, BB___, b_, CC___] := CirclePlus[AA, BB, CC] /; a === -b
SetAttributes[CirclePlus, {Orderless, Flat}]

CirclePlus[x, CircleMinus[y, x]]
(* y *)
| improve this answer | |
$\endgroup$
  • $\begingroup$ The problem here is that this way some transformations will be applied unconditionally, even without Simplify. And in your use of Simplify the assumption is not even used — it works without it (it's in fact unneeded at all, since nothing will change if x==0||y==0). $\endgroup$ – Ruslan Mar 1 '16 at 5:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.