That'd also be a lot less smooth, right? You'd get a mountain range
instead of a safety scissor edge.

Well, you'd get linear interpolation between random values. That might be great or crap, depending on your needs. I do think it has to count as "some sort of random LFO". You could filter the results, you could also wave-shape them but a wave-shaper that scales like that might be a bit counter-productive to implement.

And SubNoise is sample rate
independent if you take the sample rate into account when you set its
rate. :D

That does assume a infinite sample-rate. The result of a shred and a Envelope/step will also be quantised to the sample-clock, but the rounding error won't carry, like it will with setting SubNoise.

Using a shred, a Step and a LPF will give you all of the advantages of a Subnoise and a LPF in the smoothness of the curve, without the timing errors, but it's also a lot more involved than a simple SubNoise.

This is the 1 Hz weirdness I noticed, by the way. The lines are noise
generated at 5 Hz, 1.1 Hz, and 1 Hz. You can see the unnaturally long
plateaus that get way out of hand at 1 Hz.

http://skitch.com/alltom/d25q8/subnoise

But the 5 Hz shape looks pretty nice.

That's strange! Those should indicate the LPF somehow ends up below the frequency of the subnoise and starts averaging too much. In any case it's not a bias in the randomness as the plateaus tend to run off towards 0.

What I can say is that at higher frequencies the frequency resolution in steps per octave of SubNoise will go down and so you will get higher rounding errors in SubNoise. If that frequency ends up above the frequency of the LPF we might see something like what you are seeing?

Yours,

Kas.