Required reading:
http://en.wikipedia.org/wiki/Bipolar_junction_transistor#Ebers.E2.80.93Moll_modelVbe is NOT "0.7V". Obviously if it were fixed, resistors would work, and that would be the end of that. So where's the tempco? Duhh....
Right? So, obviously, calling it 0.7V, full stop, isn't right.
If you take the exponential,
I ~= Is * exp(V/Vth)
(Vth = thermal voltage, also, give or take the emission coefficient; Is = saturation current)
then you will see, for some fixed terminal current (such as for the bias diodes), the terminal voltage must go as ln(I/Is). The logarithm is a very slowly increasing function (over a factor of 10x up or down, Vf only varies about +/-0.1V), but it's nonetheless variable.
Vth is proportional to absolute temperature. Is also varies with temperature, but it's usually the less significant factor (it's not inside an exponential!).
The BJT does the same, except it does Ic ~= Is * exp(Vbe/Vth). Ib is similar, and the ratio between them is hFE, but as long as this ratio is large, we don't care about it,
and in particular, we always want to design circuits such that they are as insensitive to hFE as possible. hFE varies widely with manufacture, temperature, Ic and Vce.
The result is that, if we have a resistor-diode voltage divider (such as the bias network in the first example), we get a voltage offset that's characteristically proportional to this log-exp function, including its tempco. Connect that B-E and you get the exact same current back, give or take the differences in doping and junction area between the diode and transistor.
To further enhance stability, emitter resistors are used, which either complement the resistor in the bias chain (if it's a +V - pullup - base - diode-diode-
resistor - other base - pulldown - -V configuration), or simply buck the current mirror behavior altogether (linearizing the transistor's transfer function retains the diode's log character, making something like a Widlar current mirror, i.e., some current in, log current out).
I mentioned doping and area. Saturation current is a device constant -- more junction area means more Is, straightforward as that. If you had a power transistor that was made on the same process as a 2N3904, but which was rated for 2A instead of 0.2A, you should find Is measures 10x higher. But you're more likely to get disparate transistors (say, a TIP31C with a 3A 100V rating, clearly using a different process than the 60V rated 2N3904), which means both Vbe and Is will be different. You can very easily end up with a runaway tempco, even with correct diode compensation in place. This is most typical with excessively high amperage transistors (and is doubly a problem, because big transistors can sink big currents if they want to!).
The concept of "ratio of areas = ratio of voltages" arises in the design of the balanced bandgap voltage reference, or the intentionally unbalanced temperature sensor.
Tim