Measure a table as 120 cm with a tape you can read to within 0.5 cm. The true length is somewhere from 119.5 to 120.5 cm, and every later calculation inherits that range. Cut the table's length into 4 equal parts and each is 30 cm, uncertain by ±0.125 cm: dividing a measured quantity by an exact number divides its uncertainty too.
Combining measurements makes uncertainty grow. Add two lengths, 50±0.5 cm and 30±0.5 cm, and in the worst case both errors point the same way: 80±1 cm. For a product it is the percentage errors that add. A rectangle measuring 10±0.5 cm by 4±0.2 cm is 5% uncertain in each direction, so its 40 cm² area is about 10% uncertain, ±4 cm². Squaring doubles the percentage, because r2 is r times r, so a 2% error in a radius becomes a 4% error in a circle's area.
None of this is about mistakes. Writing 7×8=54 is a blunder: you find it, fix it, and it's gone. Measurement uncertainty never disappears; it is part of every measured number, and the best you can do is shrink it with a better instrument and report it honestly. Significant figures are a rough version of the same idea: "120 cm" quietly suggests "to the nearest centimetre or so", while "120±0.5 cm" says exactly how far off it could be.
Think of it like this. A measured number is a slightly blurry photo: you can see it's about 120, but the edges are fuzzy by half a centimetre. Every calculation makes a copy of a copy, and adding or multiplying blurry numbers widens the blur, never sharpens it. Unlike a photo, you can state exactly how blurry the number is, and you should.
Core idea. Error analysis studies how errors arise, how big they are, and how they propagate through later steps.