The answer - there is the optimum bias for each duty factor.
Effective voltage for LCD is RMS. Actual contrast depends on LCD panel and "discrimination ratio" - ratio between VrmsON and VrmsOFF.
According to http://www.nxp.com/documents/data_sheet/PCF8576C.pdf
this ratio is
D = sqrt((b^2)+m-1)/((b-2)^2 + m -1)),
where bias is 1/b and duty is 1/m.
For 1/16 duty the optimum is 1/5 bias with D = 1.291.
This is magically the same bias as stated (though not explicitly described as optimum) in Raystar datasheet :).
Discrimination for 1/4 bias is 1.277 - almost the same - that is why I did not see the difference. Transmission curve of LCD is sharp enough for 1.277 too. On the other hand a temperature drift of this curve will be visible (slightly) earlier with 1/4 bias.
(for duty factor 1/8 the optimum bias is 1/4 with D=1.446, and for 1/5 bias D is 1.414 - more considerable difference)
There is another factor - absolute voltage values. Bias 1/4 requires less V0 than 1/5 bias does:
Von(1/4)= V0*0.348, Von(1/5)= V0*0.316; (1/16 duty)
In case of RX2002 this is not an issue.
UPDATE: actually - the issue. There is no voltage margin. Vdd=4.75 V is not enough even for optimal contrast (for some view angles) using 1/5 bias. In my case an optimal VrmsON is about 1.5 V. So for 5V application of RX2002 an 1/4 bias is a practical choice.