Unary minus precedence?

[dronir]

Not sure if this is intentional or not, but it caused me some confusion. If it's intentional, I'd be interested to hear the rationale.

The precedence of the unary minus seems to be over exponentiation such that -(x-y)^2 equals (-(x-y))^2, so for example -(1-3)^2 == 4. This differs from at least Python, where -(x-y)**2 equals -((x-y)**2).

Might be a matter of taste (as Python is my major language), but for me this latter way is more intuitive, and makes for prettier expressions. I was writing something with a gaussian function exp(-(x-a)^2) and got quite unexpected numbers out at first. Had to write it exp(-((x-a)^2)) which looks a bit cluttered.

I think this should at least be mentioned in the manual somewhere (at least I couldn't find much about operator precedence).

[JeffBezanson]

Good catch. I'd say this is a bug, because:

julia> -3^2
-9

julia> -(4-1)^2
9

The first one is the one we intended, and these clearly shouldn't be different.

[JeffBezanson]

Oh, I know why this happened. We made -(x) function call syntax, not negation of a parenthesized expression. That way you can call - using -(x,y). So this is using the same precedence as f(x)^2, where f is clearly called first. This was desired for writing definitions, such as

-(x, y) = ...

Otherwise it would have to be written as

(-)(x, y) = ...

But I think this was the wrong tradeoff, since you hardly ever want to call or define - using prefix syntax, compared to the importance of negating stuff.
@StefanKarpinski what do you think?

[StefanKarpinski]

I guess if we switch this back then methods for - will have to be written as (-)(x) = .... I guess that's ok.

[JeffBezanson]

We might be able to have both, if we're willing to assume you never want to negate a tuple.
-(x) is the same no matter how you parse it (always a 1-argument call to minus), but we could make tuples a special case so that -(x,y) by itself is 2-argument call to minus instead of negation of a tuple. The only nasty thing there is that -(x,y) would work, but -(x,y)^2 would not since that would try to square the tuple first.

[StefanKarpinski]

Seems reasonable enough to me. There will inevitably be people who try -(x,y)^2 and are flummoxed when it doesn't work, but I think that's a weird enough thing to try that having it not work is ok.