I've also observed the effect that the OP mentioned: in my Honda Fit, I get a fairly consistent ~45 MPG on 101 (completely flat highway terrain), ~40 MPG on 280 (rolling hills, highway), and ~43 MPG on 84 (across the Santa Cruz Mountains; ascends 3000 ft. in a series of tight switchbacks and then descends 3000 ft. in a wavy line). My working hypothesis is that the descent on 84 is consistent enough that the engine computer can idle the engine for all of it, burning no fuel - this is consistent with the car being nearly silent on the downhill, and of the instantaneous mileage display being stuck at its max of 80 MPG. Probably also helps that speeds are much lower, so there's less air resistance.
Anyway, the point is that cars are complex enough that the simple models you learn in Physics 101 don't really hold.
It's known that you can get better fuel economy by oscillating speed on flat ground - using WOT to gain speed, coasting back down, opening up WOT again. This is because the engine is most efficient under load and the average consumption works out better. Adding in the hills would average this effect.
An engine operating at WOT but still producing a fraction of its maximum output is certainly possible (and generally more efficient than an engine producing the same output at a partial throttle setting). A diesel engine is almost always in this configuration (metering fuel but generally not restricting air with a throttle plate) and a gas engine at low RPM but WOT is also in this state.
If tuned for efficiency in that config, it can be very efficient and is how many piston airplane engines are regularly operated where fuel consumption is an important part of range (and operating economics).
But the actual situation has two novel features: (1) the engine uses up power just by being on; (2) you must maintain a minimum speed in order not to block traffic.
If you put these constraints in, you can see that it's possible for a pulsed-acceleration strategy to win. You put in a pulse of acceleration to get up the hill, turn the engine off, and coast down, maintaining enough speed to keep up with traffic.
Now sure, without the minimum-speed constraint, you'd probably be better off flat, and possibly using a pulsed-acceleration strategy.
I haven't done the math, but I think it would be reckless to say it's "physically impossible" for all values of static energy use and slopes.
Take a look at the stack exchange answer the commenter above linked to.
But you raise good points, maybe I jumped the gun on 'physically impossible'.
Even if you do downshift that doesn't necessarily mean you're being less efficient either. Being a bit higher up the rev range may actually put you in a better spot on the efficiency curve for a given power output.
http://ecomodder.com/wiki/index.php/Brake_Specific_Fuel_Cons...
I actually observed this and it is true in my case; I do get better mileage when choosing a route through a canyon than when using regular freeway.
I suspect though that the real reason is that it basically affects how I'm driving. When I ride in a canyon I can't accelerate much because there are constant turns, while on straight freeway I most likely will accelerate.
This also would explain why most of the time I get better mileage in traffic (the EV is used most of the time).
This is physically impossible. Even if the car was perfectly efficient at overcoming the force of gravity on the uphill portions (which it is not), you cannot do better than flat terrain (unless by hilly you mean 'net downhill').