A fair die has six faces, all equally likely, so before you roll it you can already say P(3)=61. That is theoretical probability: reasoning about how the object is built. Now roll it 60 times and suppose 3 comes up 14 times. Your experimental probability is 6014≈0.23, noticeably more than 61≈0.17. Nothing is wrong; 60 rolls is a small sample, and small samples wobble.
As the number of trials grows, the wobble shrinks compared with the total. In 60 rolls, 14 threes instead of the expected 10 is a difference of 4, a big share of 60. In 6,000 rolls the count of threes usually lands within about 60 of the expected 1,000, so the experimental fraction comes out within about one percentage point of 61. That settling down is the law of large numbers.
Some situations look theoretical but are not. A thumbtack can land point up or point down, two outcomes, yet the chance is not 21, because the two outcomes are not equally likely. A basketball shot either goes in or misses, but that does not make it 50%. Theoretical probability needs outcomes that are equally likely by symmetry or by design; without that, only trials can tell you the chance.
Think of it like this. Theoretical probability is like a recipe that says 'makes 24 cookies'; experimental probability is what your oven actually produced: 22 one time, 25 the next. Over many batches your average comes close to 24. If the recipe is wrong, like a loaded die, the batches settle somewhere else, and only your records show it.
Core idea. Theoretical probability is computed from equally-likely outcomes; experimental probability is the observed fraction from real trials, and they converge as trials grow.