Roll two fair dice and add them. Nobody needs to roll to find the chances: list the 36 equally likely pairs (first die, second die) in a 6-by-6 grid. A sum of 7 comes from (1, 6), (2, 5), (3, 4), (4, 3), (5, 2) and (6, 1), which is 6 pairs, so P(7) = 6/36 = 1/6. A sum of 2 needs (1, 1), only 1 pair, so P(2) = 1/36. The sums themselves are not equally likely, which is exactly why you count the pairs, not the 11 possible totals.
A theoretical probability is a prediction of the long run. Roll the two dice 360 times and you should see about 60 sevens, since 61×360=60. You might see 52 or 67, and that does not make the 1/6 wrong; actual results scatter around the prediction, and the scatter, measured as a share of all the rolls, shrinks as the number of rolls grows.
Some situations have no theoretical probability at all. A thumbtack can land point-up or on its side, two outcomes, but nothing about its shape makes them equally likely, so 1/2 would be a guess, not a calculation. The same goes for a free throw, a bent coin or tomorrow's weather. If you cannot argue from symmetry that the outcomes have equal chances, you need data from trials instead.
Think of it like this. Reading the blueprint instead of inspecting the building: that is theoretical probability. The plan of a fair die says each face gets one sixth of the chances, without a single roll. A real die built slightly off-plan, worn or loaded, will not match the blueprint, and only experiments can reveal that.
Core idea. When symmetry makes every outcome equally likely, you can find a probability by reasoning alone: count the outcomes in the event, divide by the number of possible outcomes, and you have the long-run share before a single trial.