🚀 go-pugleaf

In article <66707@sgi.sgi.com>, karsh@trifolium.esd.sgi.com (Bruce
Karsh) writes:

[ I wrote ]
|> >Algorithms have a tendency to be rediscovered independently when the time
|> >is right for them.  My field is digital signal processing.  Imagine what
|> >a horrible world we'd live in if AT&T held a monopoly on the use of the
|> >Fast Fourier Transform.

|> Well, I dabble in digital signal processing also and I thing there are a few
|> things wrong with your position here.
|> 
|> 1)  There was prior art for the Cooley-Tukey FFT algorithm, so AT&T
|>     could not have patented it.

This is an example of what I'm talking about, Bruce.  The FFT was rediscovered
many, many times.  Cooley and Tukey made it famous because they published just
at the point where the straightforward O(N^2) algorithm was too expensive but
the O(N logN) algorithm was practical.  Gauss is probably the first documented
discoverer of the FFT -- in 1805!  But prior art has not stopped people from
getting software patents; they tend to go to people in large organizations that
have legal staffs.

Note: neither Cooley nor Tukey nor Gauss INVENTED the FFT; they discovered it.
But guess what: a trivial variation of the FFT, called the Hartley Transform,
IS patented even though there is no more ground to do so than the FFT.  As
a result, people will avoid using the Hartley transform even where its use
is warranted.

|> 2)  Had the FFT been patented, there would have been a lot more
|>     investment in the whole field of fast transforms.  As it stands, it
|>     took another 20 years for new fast orthogonal transforms to be
discovered.
|>     In the intervening time, people just copied the same algorithm over and
|>     over again.

What's the use of pointless innovation?  The FFT is the most widely used
algorithm in signal processing because it's the correct algorithm to use,
and there's been plenty of research into other orthogonal transforms in
the meantime.  And no, they haven't copied the same algorithm over and
over again -- they've refined it.  There are many variants of the FFT;
some (the Winograd form, for example) are radically different from the
form described by Cooley and Tukey.  Subscribe to the IEEE's "Transactions
in Acoustics, Speech, and Signal Processing", Bruce, and you'll notice
people innovating all the time by building on each others' work.

People in real-time DSP have plenty of incentive to find faster algorithms
without patent protection for the algorithms.

|> 3)  AT&T has patents on UNIX and that has not prevented its widespread
|>     adoption and it may have even encouraged it.

Wrong, wrong, wrong.  The only patent AT&T has on UNIX is one for the setuid
bit.  This was put into the public domain in 1979, which makes it as if there
were no patent at all.  AT&T used TRADE SECRET protection, the very form
of protection you say is so harmful to innovation.

|> 4)  Had the original inventor of the FFT obtained a patent, then his
invention
|>     would not have languished in obscurity for so long.

The FFT would STILL have languished in obscurity because there were very
limited
applications for it in 1805.

> There are a lot of DSP algorithms which are currently being kept very
> secret.  Companies which depend on the competitive advantage that these
> proprietary techniques give them are keeping them trade secrets because
> they don't believe that patents are yet a secure enough protection.

No, Bruce.  It's not that they believe that patents are not a secure enough
protection; they believe that you can't patent scientific discoveries!
For example, a particular speech processing company may believe that they
have a unique insight into the way the human speech production system works.
You couldn't patent any process that exploits a particular scientific
discovery.

But most DSP companies are fooling themselves, jealously guarding rather
mundane work.

|> This is perhaps true for any invention.  But how long would it be before
|> the next rediscovery?  And who'd invest in the invention if there's no
|> patent on it?

Anyone who needed to, or who was curious about it.  You run profiling tools
on your program, find the slow parts, and do research to find ways of making
them faster.

>>If you discover an algorithm and hide it, others in your field will
>>discover it a year later.
> There's simply no evidence for this.  Simultaneous discovery does occur,
> but it's just not that common.  The existence of trade secrets is another
> reason to believe that your statement is simply false.

It doesn't always happen, but it's amazing how often it does.  Major pieces
of special relativity, for example, were discovered by several people at
about the same time (Poincare, Lorenz, as well as Einstein); if Einstein
hadn't lived, it would have held up the process of discovery only a couple
of years.

As for trade secrets, you've been working for the same company too long,
I suppose.  Often two competing companies in the same field have the same
trade secrets, but those under nondisclosure agreements can't tell them this!

|> >  That's because algorithms, for the most part,
|> >are discovered rather than invented.  This distinction was established
|> >in law by the founders of the country (you can't patent mathematical
|> >relationships or laws of nature)
|> 
|> This is clearly nonsense.  Patents weren't even invented in the US.

You clearly can't read.  I didn't say patents were invented in the US;
I said that the founders of the country established certain restrictions
in patent law, and I am correct in this.


--
Joe Buck
jbuck@galileo.berkeley.edu	 {uunet,ucbvax}!galileo.berkeley.edu!jbuck