MathGroup Archive 1995

[Date Index] [Thread Index] [Author Index]

Search the Archive

Transpose, Re: The Case of the Mystery Option

>From: Jack Goldberg <jackgold at> in [mg1056] The  
Case of the Mystery Option asks some questions about transpose.
Robert Villegas responded in mg[1063].

Jack asked about uses of the second parameter in Transpose other  
than for getting the diagonal of a square matrix by  
Transpose[M,{1,1}]. I did find a use for calculating the determinant  
(mathgroup April 1992 I think). The result is given after the  
following notes.

First, let's look at the way we extract the elements of a tensor T.
Suppose that its tensor rank is 4 and that its dimensions  are   
{d1,d2,d3,d4} then we need 4 positive integers  a1,a2,a3,a4 to get  
an element :
and, of course ,we must have  ai <= di for all i.

Transpose[ T, {2,1,3,2}]] is a tensor that given  three  positive  
integers a1,a2,a3  will return the value

Transpose[ T, {2,1,3,2}] [[a1,a2,a3]]  =  T[[a2,a1,a3,a2]]

We only need three numbers because the repeated 2 uses the second  
one twice. So the tensor rank is 3  and we must have   a1 <= d2, a2  
<= d1, a2 <= d4, a3 <= d3.  In fact  Transpose  requires that  d1 =  
d4. The dimensions of the transposed tensor are  {d2,d1,d4}

	Transpose[ T, {2,1,1,2}] [[a1,a2]]  =  T[[a2,a1,a1,a2]]
and the trnsposed tensor has rank 2 and dimensions  {d2,d1}

T = Array[SequenceForm,{2,4,4,2}];
	{2, 4, 4, 2}

	{4, 2, 4}

	{4, 2}

The GENERAL PATTERN (for elements) is


Transpose[T_, p_List][[a__]] = T[[Sequence@@({a}[[p]])]]


In the examples so far the length of  p  has been equal to the  
tensor rank T (this is needed t o provide the right number of inputs  
to T. But it would be inconvenient to always have to provide  the  
full sequence so for example, for a tensor T of tensor rank 5  
Mathematica interprets
	Transpose[ T, {2,1,1}] as Transpose[ T, {2,1,1,3,4}]
(appending successive integers from the first positive integer not  
appearing in {2,1,1}).

So much for the elements.
But  how do we construct the transposed tensor itself as an array?
Well that's the job of the Array function.
Notice that we don't need to allow for extending  p  -- this is  
automatically taken care of

Trans[T_, p_List] :=
	Block[{dims, trdims},
		dims = Dimensions[T];
		trdims =
		Array[ T[[Sequence@@({##}[[p]])]]&, trdims ]
The conditions imposed by Transpose are (where the rank of T is  r  
and its dimensions are  {d1,d2,...,dr }).
 1. Union[ p]  is Range[t] for some t <= r;
 2. whenever   pj = pk then  dj = dk. (where p = {p1, p2, . . , pr}) 	

The transposed  tensor  is of rank r minus the number of repeats in  
p and, as for dimensions, if s occurs  in at position  t  in  p  
then sth dimension is  dt , otherwise the sth dimension can be  
worked out by extending  p  as  discusssed above.

Trans[T,{2,1,1,2}] === Transpose[T,{2,1,1,2}]
Trans[T,{2,1,1}] === Transpose[T,{2,1,1}]
Trans[T,{1,2,2}] === Transpose[T,{1,2,2}]

Finding the diagonal

For a square matrix  M, Transpose[M, {1,1}] is the diagonal of M  
		Transpose[M, {1,1}][[a]] = M[[a,a]]


Transpose[{{a,A},{b,B}}, {1,1}]
	{a, B}

But for non-square matrices

Transpose[{{a,A},{b,B},{c,C}}, {1,1}]
	Transpose::perm: {1, 1} is not a permutation. (*message*)
	Transpose[{{a, A}, {b, B}, {c, C}}, {1, 1}]
The message really means  that  the condition 2. is not met.

This can be allowed for by changing the definition of  trdims   in  
the code for Trans

Trans2[T_, p_] :=
		dims = Dimensions[T];
		trdims =
		Array[ T[[Sequence@@({##}[[p]])]]&, trdims ]

Now we have

Trans2[{{a,A},{b,B},{c,C}}, {1,1}]					 
	{a, B}
Trans2[Transpose[{{a,A},{b,B},{c,C}}], {1,1}]
	{a, B}
It seems resaonable to call this the diagonal of these matrices

Transpose[T,{1,1}]  seems to be the quickest way to get the  
diagonal of a square matrix.

The code for the determinant that I arrived at was

NewDet4[M_] :=
	( Signature[M]
		Append[ Transpose[#,{2,1,1}],(Signature/@#) ]&@
		Permutations[ M ]

m = Table[ StringJoin["a",ToString[i],ToString[j]], {i,1,7},{j,1,7}];
{6.58333 Second, Null}
{10.3667 Second, Null}

But be warned: I think that it was Dave Withoff who pointed out  
that the result for Det was because of internal use of Expand and,  

{0.166667 Second, Null}

(Block[{Expand},Det[m]] - Det[m])//Expand

Allan Hayes
hay at

  • Prev by Date: Re: Maximum Likelihood Estimation
  • Next by Date: Convert Products to Lists
  • Previous by thread: Books on Mma graphics
  • Next by thread: Pretty Graphics of Astroid by envolope