Smaller bypass caps aren't usually a problem. If you imagine the regulator having an inductive or potentially negative-resistance characteristic (i.e., a big enough or low enough ESR cap will oscillate), then you'll see that, as long as that characteristic is well damped, it'll be okay.
High ESR avoids instability, but probably doesn't offer a low enough impedance for subsequent circuits, or the output is noisy (or maybe it really is unstable with little or no load -- worse things have happened).
It follows, if you have low-ESR caps, but you dampen them with a much larger, optimal ESR cap, you can insert the damping over just the range of frequencies the regulator needs, without sacrificing supply impedance for the loads, in fact making it better than otherwise.
So, it's usually fine to have small caps bypassing things, as long as you have some damping in there to keep things happy.
Looks like a SPICE model is available for that regulator, so if nothing else, you could simulate the power supply network and see how it goes. I would suggest this:
- Instead of pure 0.1uF caps (or whatever you were intending for bypass), add maybe 0.05 ohm ESR and 2nH ESL.
- Likewise for the lossy cap, find the best estimate for ESR, ESL and C you can -- tricky for electrolytics, but almost always specified for Ta/Nb types (as you've noted!).
- Traces are very roughly 20nH/inch. Estimate how much length you'll need to route the supply, and make nodes at every branch or junction, separated by trace inductances.
- There are also pin, pad and via inductances, which might be included for even better estimates up into the 100s of MHz range. And the ESR (bonding, spreading resistance) and capacitance of ICs themselves, which... who knows. I've never even measured, say, a 74HC00's supply capacitance before, but that would be yet another thing to consider.
- If you're using ground plane or well-stitched pours, you might as well assume ground is, well, ground. If not, do the same thing for the ground return network... but note that all trace inductances will be significantly higher (maybe 2-5x??) in an unpoured design.
- When you've got the network set up, test it by applying stimulus to a given port* (e.g., an IPULSE source, or a VPULSE + 50 ohm resistor), and observing the amplitude produced at various ports.
*Port meaning, in this case, any connection point to the supply. So, IC power pins, stuff like that.
If your ground network is also implemented, don't forget to measure the voltage across the respective ports, VDD to VSS. You can also inspect common mode signals (i.e., between VDDs or VSSs of different chips), but this won't be very complete without also modeling the circuit itself (the capacitance and inductance of signal traces, the capacitance of logic inputs, and so on).
The neat part being, you can test what happens by placing different components, and different values, in different locations, or stringing them together with a different routing topology (chained, starred, branched..), or even adding trace inductance or explicit inductors to better isolate different sections. And adjusting capacitors, values and ESR to see what gives the best results.
In general, a chained power topology is going to act like a lumped transmission line, meaning you have trace inductances between multiple caps to ground. Which will tend to have a characteristic impedance (ca. sqrt(Ltotal / Ctotal)), which is the magnitude of a disturbance (e.g., a 100mA current spike on a 5 ohm transmission line will never produce more than 0.5V peak ripple, instantaneously speaking), and the approximate "best value" for damping it (a 5 ohm transmission line, consisting of 1uF total bypass, would benefit from a 5 ohm ESR, >= 5*1uF cap somewhere). Damping is also best applied at both ends (doubly terminated, essentially), but you can get by with just one. It's not usually a good idea to go without damping, otherwise that "5 ohm" transmission line will be very different from 5 ohms at certain frequencies that are hard to predict.
Accordingly, short distances between capacitors are futile: not enough inductance to matter. Typically you only need bypass caps every couple inches, but requirements become tighter with denser logic, higher speed or higher power. So, say if you have a lot of LEDs, that change intensity all at once, you can calculate how much bulk capacitance is required for some allowable voltage change, and how low the ESR must be.
Tim