[time-nuts] Re: DMTD Questions

[Attila Kinali]

Hi Chris!

On Wed, 8 Jul 2026 09:15:10 -0500
Chris Hastreiter via time-nuts <time-nuts at lists.febo.com> wrote:

> I have been working on some modifications to Corby’s DMTD design [1] in an
> effort to further improve its noise floor and had a couple questions for
> the group.  

Ah! Someone else going the DMTD route! 😃

First of all, the design that Corby came up with is superb!
When I got the first noise floor results form Jürg Kögel in 2018,
I was impressed what they had achieved. This design is on-par with
top-of-the-line commercial offerings!

Unfortunately, because the design is so good, this also makes it
difficult to improve. You have to be very careful in how you read
the plots and deducing what element might limit the performance.

If you are trying to identify sources of instability within the
circuit, I would advice against using Allan deviation and instead
switch to the modified Allan deviation. You need to be able to
tell white phase noise from flicker phase noise. You should also
look at the spectrum of the measurement to get a different view
of the signal.

Back in the discussion with Jürg I told him that I suspect the
LT1007's flicker noise to be the limiting element. As you have
replaced the LT1007 by an LT1028, which has about half the flicker
noise power I guess that should take care of this limit.
The only thing you could do to improve this, would be to build a
discrete input stage using MAT12, but even that would give you
at best a factor of around 2 in flicker noise.

What you could also do, is to replace the current 2:1:1 transformer
with a 1:1:1 transformer, to reduce the noise contribution of the
input amplifier. But even that will at best give you a factor of 2.

As I said, the design is very close to being optimal.


> I believe that much of the noise
> floor degradation at long tau is due to thermal effects in the first stage
> amplifier of the ZCD as I haven’t yet gotten the system installed into an
> enclosure and I can worsen the issue by thermally disturbing the LT1028 of
> the first stage.

Jürg's data also showed a slight bump around 10s. Yes, as you concluded
this bump hints at some thermal behaviour. So yes, you should definitely
put everything into an enclosure and temperature stabilize it.
Using thermal glue to attach some copper cubes on top of the critical
components to give them more thermal mass might also stabilize
their behaviour.

But I would not conclude that this is due to the first amplifier
after the mixer. The mixer itself has a phase shift too, which is
in the order of 1-3ps/K. Actually, a good rule of thumb is to assume
that every electronic component has 1-10ps/K phase shift/delay variation.


> My first question is in relation to mixer noise which is a noise source
> that I have not yet considered.  I am using the same SBL-1 mixer as Corby’s
> original design but [2] implies that some mixers exhibit lower flicker
> phase modulation than others and I haven’t seen any literature or
> specifications that point to any devices other than the 10514A in that
> paper.

These publications exist, but are rather hard to find unless someone
tells you where to look ðŸ˜


I can refer you to [1] and [2]. The latter contains a comparison
of various mixers you can buy today and a mixer they have built
themselves using diode connected 2N2222. Please note the choice
of 2N2222 is arbitrary. They did not evaluate different transistors
on their noise performance but took one they had at hand. I would
try a BC847D, which is the low-noise variant of the BC847 family.
There has been also discussions using the ZTX951, which potentially
could offer better noise performance due to its low base spreading
resistance. But I have never seen any measurements. I believe
they have a modern SMD variant of the ZTX951, but I can't currently
find it.


>  They also discuss adding capacitive loading to the mixer to further
> reduce noise referencing another paper that I have yet to be able to find.
> Is there any expectation that changing to a mixer different than the SBL-1
> will improve my results and should I consider capacitive loading on the
> mixer?  If so, how would I determine an appropriate amount of capacitive
> loading?  I currently have it loaded with a bridged-T filter holding 50
> Ohms of impedance across the frequency spectrum.

Quite honestly, I do not know whether replacing the mixer or changing
the mixer circuitry would help. A couple of years ago I did
extensive simulations on mixer circuits and noticed that I could
improve linearity (linearity is always good if you want low noise)
when I added a low-impedance path for the higher frequency output
components. This would align with the capacitive loading in the paper.
Unfortunately, I never had the chance to verify the impact on
noise performance by measurements.


> My second question is in regard to the common oscillator noise
> contribution, specifically for the use case of a synthesized offset
> oscillator.  The below plot shows the results I’ve gotten so far with some
> sources that I already have, and as can be seen the system noise floor
> degrades much faster with the synthesized offset than with the 10811A.


This is expected behaviour. The reason for this is that the phase
difference causes the zero crossings of the two channels to be
at different times. Which means the two zero crossings sample the
noise of the offset oscillator at different times. Which in turn means
they see slightly different noise and thus the noise does not completely
cancel out anymore. This degradation is mostly dominated by the
wide-band noise of the offset oscillator. So lower wide-band noise
means better performance when there is a phase offset between the inputs.



> However, it would be advantageous for me to be able to use a synthesized
> offset for tests where the DUTs differ slightly in frequency and the system
> will phase wrap.  The reason for this is that when using a free-running
> oscillator the phase wraps are very difficult to remove automatically
> because the divided phase wrap will differ from the expected with any
> frequency error in the offset oscillator.  Do you have any suggestions for
> synthesizers or other offset techniques that would help with this?  Using a
> 10811A and a slow PLL as a cleanup loop on the synthesizer does appear to
> work well in the below plot,

Using a loose PLL lock is the way to go here.

If you don't want to use a clean-up oscillator, to enable you
to use a wider variety of offset frequencies, you can also
set the DDS to a higher frequency and then divide this down
using a lambda-divider [3]. You don't need to use a positive
and a negative phase shift register as they did in the paper.
The important part is that the resulting waveform is not a
square wave anymore and thus has faster decaying harmonics 
(it's the harmonics that down-convert high frequency noise,
which leads to the 10log(N) in pi-dividers).

> but looking at the first plot in this email at
> least in my initial implementation there seems to be some low frequency
> oscillation in the PLL impacting the noise floor.  

Have you locked the DDS to one of the oscillators under test?
If so, then this could explain these oscillations in the ADEV.
If not, my first guess would be thermal effects in the DDS->PLL->10811 path.

I hope this helps

			Attila Kinali



[1] "Environmentlal Effects in Mixers and Frequency Distribution Systems",
by Nelson and Walls, 1992

[2] "Residual PM Noise Evaluation of Radio Frequency Mixers",
by Barnes, Hati, Nelson, and Howe, 2011

[3] "The Sampling Theorem in Pi and Lambda Digital Frequency Dividers",
by Calosso and Rubiola, 2013

-- 
The driving force behind research is the question: "Why?"
There are things we don't understand and things we always 
wonder about. And that's why we do research.
		-- Kobayashi Makoto