Re: Explicit measurability in the QUIC wire image (was Re: Packet number encryption)

[Kazuho Oku <[e-mail]>]

2018-02-07 21:21 GMT+09:00 Mikkel Fahnøe Jørgensen <[e-mail]>:
> Hmm a periodic truly white noise random signal ought to have an evenly
> distributed power spectrum meaning it contains all frequencies up to half
> its cycle length. At least there is a lot of solid communications theory on
> random signals that could be applied, e.g.
>
> https://ocw.mit.edu/courses/electrical-engineering-and-computer-science/6-011-introduction-to-communication-control-and-signal-processing-spring-2010/readings/MIT6_011S10_chap09.pdf
>
> http://bme.elektro.dtu.dk/31610/notes/chap_8.pdf
>
> Thus limiting the signal to a square or a sine is effectively wasting a lot
> of bandwidth, even if naive hash based signal processing is insufficient to
> deal with a proper random signal.

I mostly disagree.

The reason I am suggesting to use square or sin wave is because it has
a base frequency (i.e. wavelength) that can be used to calculate the
loss rate.

White noise does not have a "base frequency." Therefore, it becomes
harder to observe the metrics when the loss / reorder rate becomes
high.

Going back to the discussion of the approaches, I am not against
mixing in some random signals so that pattern matching can be applied
when the error rate is low. But such "mix" should be mild enough that
it would not disturb the observation of the base frequency under high
error rate.

And in fact, you do not even need to mix randomly flipping bits if you
select a sin wave, since a sin wave quantized to 1-bit will contain
many flipping bits. You can use those bits for pattern matching. OTOH,
if we select square wave, it might make sense to apply some bit
flipping so that we can apply pattern matching. There are other
possibilities as well.

However, with that said, I believe that the design decisions need to
depend on the actual use-case. What are the loss rates that we need to
work on? What are the reorder patterns that we would observe? IMO this
is the tricky part of the 1-bit approach.

> Of course there is the argument of
> simplicity, but long term I believe more information in the signal is more
> valuable.
>



-- 
Kazuho Oku