On Tue, 26 Sep 2006 22:36:53 -0700, Richard Clark
wrote:
I've been using various flavors of E(L/Z)NEC for more than a decade
and I've never seen such dramatic cogging of the data that was not
attributable to construction (notably fractals). Your data is
stranger yet in having correlated noise on the left, and uncorrelated
noise on the right.

Please look at the ftp://space.mit.edu/pub/davis/misc/nec/swr.png
for a plot of the SWR using a spacing of 0.01 MHz. I suspect that the
noise that is showing up may be due to truncation error. I believe
that spacings of higher values, e.g., 0.2 MHz result in a different
sampling of the noise.

The version I am using (see
http://packages.debian.org/stable/hamradio/nec) contains this warning:

This version contains code which hasn't been extensively tested for
errors, which was input by hand from a report -- use with care. The
numerics are currently only SINGLE PRECISION.

If EZNEC were available for linux, I would look into it. Also, can it
be driven in "batch" mode without a GUI?

Thanks,
--John

On 27 Sep 2006 06:52:46 GMT, (John E. Davis)
wrote:

Please look at the ftp://space.mit.edu/pub/davis/misc/nec/swr.png
for a plot of the SWR using a spacing of 0.01 MHz. I suspect that the
noise that is showing up may be due to truncation error. I believe
that spacings of higher values, e.g., 0.2 MHz result in a different
sampling of the noise.

Yow! That is a lot of trash.


The version I am using (see
http://packages.debian.org/stable/hamradio/nec) contains this warning:

This version contains code which hasn't been extensively tested for
errors, which was input by hand from a report -- use with care. The
numerics are currently only SINGLE PRECISION.

Hi John,

Yes, this confirms the shift to double precision in EZNEC lowering
artifacts in the fine data.

However, I think it goes beyond simple matters of single or dual
precision math. When I was designing Fourier Analysis packages while
I was on contract to HP, I discovered there was a world of variability
in math library's transcendental functions.

Microsoft's product was abysmal, whereas Borland's was superlative. A
telling example is that for the transform of a sine wave into the
frequency domain under Microsoft math libraries, the noise floor was
at 60 to 80 dB below the fundamental peak with harmonics. When I
switched to Borland math libraries, there was a single bin response
and the noise floor plunged to 200dB down!

For others following this description, they may wonder at the
terminology of noise floor for a simple sine wave transform. Fourier
Analysis is done by parts through FFTs and this departs from classical
Fourier which is continuous. Because FFTs are discrete (bound by an
arbitrary start and stop), this injects spurious responses into the
transform. On top of that, rounding errors attributable to series
expansions of the transcendentals would give lower accuracy -
statistical (largely uncorrelated) noise. The sine wave transform was
a method of self-validation of the library used and Microsoft failed
big time (especially considering the 10:1 cost differential for the
more expensive M$).

If EZNEC were available for linux, I would look into it. Also, can it
be driven in "batch" mode without a GUI?

There have been various reports of success and failure when Windows
emulators have been used. I cannot report any personal experience
because my Linux machine is largely confined to Server development
(XAMMP/Wiki/MySQL/RubyOnRails).

73's
Richard Clark, KB7QHC

Richard Clark wrote:
On 27 Sep 2006 06:52:46 GMT, (John E. Davis)
wrote:

. . .

If EZNEC were available for linux, I would look into it. Also, can it
be driven in "batch" mode without a GUI?

There have been various reports of success and failure when Windows
emulators have been used. I cannot report any personal experience
because my Linux machine is largely confined to Server development
(XAMMP/Wiki/MySQL/RubyOnRails).

The last reports I've gotten are that Wine, the Linux Windows emulator,
isn't able to open the EZNEC manual, and has at least one other problem
with EZNEC. I've gotten several reports that the SoftWindows emulator
for the Mac runs even the professional versions of EZNEC flawlessly.
There isn't, and won't be in the foreseeable future, a native Linux or
Mac version of EZNEC; the market is simply too small.

EZNEC can be run in something resembling batch mode with MultiNec
(http://www.ac6la.com/).

The NEC-2 calculating engine in the demo and standard EZNEC program
types contains a mixture of single and double precision variables which
does considerably better than a fully single precision implementation
but with only a slightly greater memory requirement. EZNEC+ and the
professional programs also include a fully double-precision
implementation. None of these are identical to NEC-2 (of which there are
several slightly different versions in circulation); EZNEC has
consolidated scattered constant values, added protections against
numeric overflow, incorporates third-party math libraries for some
calculations, and has a few obscure bugs in the code fixed, among other
differences. Although EZNEC doesn't implement all the features of NEC-2
(patches, for example), it has features which NEC-2 doesn't, such as a
comprehensive geometry check and provision for wire insulation (and of
course the GUI). I've also found some optimizations done by various
compilers which cause errors or crashes in some cases, and of course
these are avoided when compiling EZNEC's calculating engines. In fact,
I'm just now working with a compiler manufacturer in tracking down what
looks like a bug I found in a new compiler I'm considering using for
future versions. NEC-2 is free, and people pay for EZNEC. They do get
something for it.

Roy Lewallen, W7EL

On Wed, 27 Sep 2006 00:19:53 -0700, Richard Clark
wrote:
Yow! That is a lot of trash.

For most uses, it is not so bad. For example, elsewhere in this
thread I posted the .nec file for a 2m-moxon. The corresponding plot
of the SWR may be seen at
ftp://space.mit.edu/pub/davis/misc/nec/moxon_swr.png. The noise is
still there, but the amplitude is much smaller and would not be
noticable on a coarser frequency grid.

Thanks,
--John

Richard Clark wrote:

Hi John,

Yes, this confirms the shift to double precision in EZNEC lowering
artifacts in the fine data.

However, I think it goes beyond simple matters of single or dual
precision math. When I was designing Fourier Analysis packages while
I was on contract to HP, I discovered there was a world of variability
in math library's transcendental functions.

Microsoft's product was abysmal, whereas Borland's was superlative. A
telling example is that for the transform of a sine wave into the
frequency domain under Microsoft math libraries, the noise floor was
at 60 to 80 dB below the fundamental peak with harmonics. When I
switched to Borland math libraries, there was a single bin response
and the noise floor plunged to 200dB down!

snip
73's
Richard Clark, KB7QHC

Richard

I am not surprised with your result after having used various MS
compilers over the years. Do you have any idea what the real
differences were in the libraries? Borland C always seemed better, more
robust at error handling, and more accurate.

I noticed similar problems back in the late 80's with MS C, but never
really needed the precision, and work pressure being what it was....

tom
K0TAR

On Wed, 27 Sep 2006 19:47:06 -0500, Tom Ring
wrote:

Do you have any idea what the real
differences were in the libraries?

Hi Tom,

I can only speculate from my experience coding various expansion
series before the 8087 was generally available. My guess is they went
with the first one in a cookbook - Newton's method comes to mind, but
that is of vague recollection. It is generally useful as a first pass
method. M$ became extinct in the Pascal marketplace soon after. I
also moved on into C++ in the late 80s (a local company here wrote one
of the first cross-compilers).

The M$ crowd thought they would take that one on too. In 1990 they
asked me to come in and give classes. What a fiasco. The first
question was how to do inline code. They were arrogant to the point
of wanting to call "their" version C++++ with the +s stacked in pairs
to produce #. Can anyone guess how long C-sharp took to get to market?

73's
Richard Clark, KB7QHC