Obviously for the reciprocal to be accurate, there needs to be the same amount of numbers above 1 as below 1, like there is for floats.
The big problem with fixed-point, making them a bad choice for most scientific computation (which is where the design decision for the current dominating system came from) is their lack of resolution.
You can represent only a comparatively small interval of numbers and a few division|multiplication will quickly lead to large errors in the output (even if you do not hit underflows/overflows).
>a bad choice for most scientific computation
Is weird this. Scientific computation is at most a niche.
Financial and real-world decimal calculations is the norm.
That is why a lot of people could be server better by the use of decimals or fixed point.
A significant amount of financial calculation is not in the nature of 'sum this column of numbers'. Most computation in e.g. derivative pricing has a lot more to do with scientific computing than it does with accountancy.
People doing financial computations are using fixed-point for very good reasons but, other than them, the scientific people are the only one that I know of who care deeply for the results of their numerical computations.
The numbers that you can represent with floating point are exponentially distributed. Floating point is well-behaved when you multiply things, and it maintains a good relative error. Addition works best when you add similar numbers. Overflow is almost impossible, but it's easy to get loss of significance.
The numbers that you can represent with fixed point are uniformly distributed. Fixed point is well-behaved when you add things, and it maintains a good absolute error. Loss of significance doesn't happen, but it's easy to get an overflow.
In both systems it's possible to accumulate error, but under normal circumstances the approximation error should average out. Fixed point makes it easier to reason about the error, floating point is a bit harder, but usually works well in practice (with some exceptions -- for example, if you add a huge array of small values you will run into trouble due to the loss of significance).