From cube-lovers-errors@curry.epilogue.com Thu Jun 6 23:33:56 1996
Return-Path: cube-lovers-errors@curry.epilogue.com
Received: from curry.epilogue.com (localhost [127.0.0.1]) by curry.epilogue.com (8.6.12/8.6.12) with SMTP id XAA11090 for ; Thu, 6 Jun 1996 23:33:55 -0400
Precedence: bulk
Errors-To: cube-lovers-errors@curry.epilogue.com
Date: Thu, 06 Jun 1996 12:23:05 -0500 (EST)
From: Jerry Bryan
Subject: Re: Compact Cube Representation for Shamir and Otherwise
In-Reply-To: <9606020417.AA04623@sun13.aic.nrl.navy.mil>
To: Cube-Lovers
Message-Id:
Mime-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Content-Transfer-Encoding: 7BIT
On Sun, 2 Jun 1996 hoey@aic.nrl.navy.mil wrote:
> Jerry writes of the standard S24 x S24 model, which uses 48 bytes per
> position without packing. He also has a "supplement" representation
> that uses one facelet from each edge and corner, for 20 bytes. He
> packs them into 13 bytes on tape.
>
> The way I did it the last time I worked on brute force was to
> pack eight twelve-bit fields:
The way I count it, Dan's result is twelve bytes of eight bit bytes. So
Dan has bested me by one byte (but see below).
>
> The orientations in two twelve-bit fields (2^11 and 3^7),
It looks like you are only storing the orientation of eleven of the
twelve edge cubies. The orientation of the twelfth can be inferred from
the orientation of the first eleven. On the other hand, you could store
all twelve and still fit in twelve bits. It also looks like you are only
storing the orientation of seven of the eight corner cubies. Again, the
orientation of the eighth can be inferred from the orientation of the
first seven. This time it matters. That is, 3^7 will fit in twelve
bits, but 3^8 will not, if I am counting it right.
> Also, postmultiplying by a fixed permutation can be done with table
> lookup without unpacking. I used this feature for twelve permutations
> of particular interest.
>
> I am somewhat rusty on the implications of using this representation
> in conjunction with Shamir's algorithm. I think it provides an
> ordering of the permutations that enables at least an approximation to
> the random access you need, then you unpack it and do a better job.
I think that unpacking would be required to build and traverse the "Shamir
tree", but the unpacking that would be required should be relatively easy.
I have always been reluctant to form compositions of packed formats
because it seemed both messy and slow to have to unpack both permutations
which are being composed and then to repack the result. In principle,
what you have to do is about the same as what you have to do in building
and traversing the "Shamir tree" with packed formats. But in practice it
seems a lot messier to unpack two permutations and to repack their
composition than to simply unpack one permutation as you traverse a tree.
Dan seems to have solved the problem for certain specific cases (and for
post-multiplying only) via table lookup. It is not clear to me that
table lookup could be used for the more general case of multiplying any
permutation by any other permutation.
I have always known that the supplements of the corners and edges could be
reduced by one byte each unpacked, down to seven bytes for the corners and
down to eleven bytes for the edges. The missing bytes could always be
reconstructed, but I didn't want to bother. Now, it occurs to me that
working with the supplements alone and never expanding them back to S24 x
S24, there is really no reason ever to reconstruct the missing bytes.
Hence, we can have an eighteen byte representation unpacked. The eighteen
byte representation easily packs down to twelve bytes (same as Dan's).
We could surely do better if we tried. Log2(|G|) is about 65.22, so we
should be able to represent each position in 66 bits, or in 9 bytes. But
I don't think the packing required would be worth the trouble.
66 bits is the theoretical minimum to represent |G| positions if the
positions are independent. But the positions are not really independent
(e.g., M-conjugacy). I'm not sure exactly what the best we can do
actually is.
My Shamir program is going to represent each position as a pair of indices
(i1,i2). i2 will be a single byte containing 1..48 and indexing M. i1
will be two or three bytes indexing a table of representatives of
M-conjugacy classes. The table in turn will consist of eighteen byte
supplements of permutations less the last edge and the last corner cubie.
A two-byte index will cover quarter turns through level 6 of the tree
(there are 18,395 representatives at level 6). A three-byte index will
cover quarter turns through level 9 of the tree (there are 14,956,266
representatives at level 9). I am dubious that I will even get into the
three byte index business. It will simply take too long. That is, a two
byte index through level 6 will allow level's 7 through 12 to be
calculated. It may well take too long to go any further than that.
Anyway, (i1,i2) means m'[i2]X[i1]m[i2]. The indices (i1,i2) will be
sorted according to the lexicographic order of m'[i2]X[i1]m[i2]. The
"Shamir tree" will be built with the indices as the leaf nodes. The
effective storage required for each position will be maybe five bytes or
so because you have to count the table of representatives in any honest
accounting of memory requirements. Thus, the tree structure itself will
be the biggest consumer of memory.
This approach will not require any packing or unpacking, but it will
require lots of extra multiplying of permutations. However, at most
points in the processing, it will not be necessary to multiply the entire
permutation. Multiplying a single cell of the vector will usually suffice
for comparing vectors, traversing the tree, etc.
= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =
Robert G. Bryan (Jerry Bryan) jbryan@pstcc.cc.tn.us
Pellissippi State (423) 539-7127
10915 Hardin Valley Road (423) 694-6435 (fax)
P.O. Box 22990
Knoxville, TN 37933-0990