[Discuss-gnuradio] Rounding errors in PFB channelizer?

[John Ackermann N8UR]

I'm building a ridiculous flowgraph that breaks the AM broadcast band 
(540 - 1700 kHz in the U.S.) into 117 10 kHz wide channels and measures 
the energy in each.  The thing is working but I see a frequency offset 
in the output channels that is not present in the data before 
channelizing.  The output channels seem to have an offset in the range 
of 400 to 700 Hz versus the unchannelized input.

The signal chain is:

2.5 msps recording centered at 1.4e6 Hz -> xlating filter, decimation 2, 
output centered at 1.12e6 -> PFB channelizer with 117 channels, yielding 
a channel rate of 10,683.760683...... samples per second.

Looking at the spectrum at the output of the xlating filter, the carrier 
frequencies are correct.  Looking at the output of a channel, the 
carriers are offset by several hundred Hertz, always high.  (Given the 
absolute frequency is in the 1 MHz range, these offsets are parts in 
1e3, a pretty large amount.)

I wonder if the large number of PFB channels is causing a rounding error 
that results in these frequency offsets.  Or is there something else 
going on?

I can probably fudge the xlating filter frequency a bit to move the 
carriers closer to nominal, but would like to understand what's happening.

I'm attaching the (absurdly huge) .grc file.  The canvas is 4192 pixels 
tall, so the flowgraph is smaller than the screenshot. :-)

bcb_channelizer.grc
Description: application/gnuradio-grc