Functional style is simply better and is applicable in many languages. Modern C++ is a great example. Throw away much of OOP clunkiness, use simple functions, win! (OOP has it's place but in small quantities and with moderation).

With respect to purity I'm on the sidelines. I think what's far more important is algebraic data-types, pattern matching and value-returning conditionals (values, not statements!). Rust has them and it is great and not pure.

For various reasons, I chose to use OCaml, and from my experience, the difficulty part is only true insofar as the learning curve goes. Once past that, FP-style problem-solving is spectacular in that the vast majority of time you spend goes not into the coding, but rather (a) thinking about what exactly your problem is, and (b) how to decompose and solve it. I can't tell you how many times in Matlab, people start writing code before they've even fleshed out (a) and (b), simply because there's so much drivel and overhead that one naturally tends to feel they're "getting things done" by jumping in and writing this code.

Once I had a sufficient grasp of OCaml concepts and syntax, I would find myself getting frustrated by how little code I might get done in a day. Upon stepping back and reflecting, that was because I had not worked my way through (a) and (b), and that is where the human thinking is really required. I have since been delighted at the utility of each line of code I write and am similarly annoyed by the low signal-to-noise of something like Matlab.

With respect to slower, I think my points above address "slowness" in terms of development time. In terms of runtime, it's very much a function of your choice of language/ecosystem and paradigms. If GC and typical functional overhead are a constraint in your problem domain, one can always bring over a functional style into modern languages like Rust, where performance and lack of GC are first-order design goals. But you can't tar all of FP with the label of "slow" any more than anyone can make a blanket statement that "solving matrix problems is slow". It just doesn't make sense.

Lastly, as others have pointed out, Jon Harrop's points refer to purely functional styles and languages. In the end, striving for functional purity may not make sense. To illustrate, one can easily drop into an imperative style where needed in OCaml, for example when doing I/O. I use this quite a bit. But make no mistake that adopting a functional approach where it makes sense can pay dividends.

I think the concept of immutability confuses people bit. It really clicked for me when I stopped thinking of it in terms of things not being able to change and instead in terms of every version of things having different names.

Functional programming from that perspective is like version control for your variables. It makes explicit, not only which variable you are accessing but which version of it.

The reason it seems like you need to copy a variable each time you modify is just to give each particular mutation a different name. Note that the compiler can optimize things. It doesn't need to keep every version in memory. If it sees that you are not going to reference a particular mutation, it might just physically overwrite it with the next mutation so that in the background "var a=i, var b=a+j", might run as something like "var b = i; b+=j";

It's clear that some other aspects of FP, such as first-order functions, closures, etc are very useful in practice.

Immutability does have its advantages in that it supplies strong guarantees about what your code does. It avoids spaghetti code where everything can and does mutate everything else.

Myself, I tend to use whenever its advantages outweigh its inconveniences. Immutability is especially handy at interface boundaries.

Art - as something subjective - also means opinions differ on difficulty and slowness. That's perhaps why downvotes - the statement like this presented as an objective fact draws enough scepticism from those who have different experience in the subject.

Slower - Optimize for developer time. Same reason most people write their backends in Python/Ruby instead of C++.

Unfortunately, it does have the challenges listed in the OP, so it seems to make most sense where either (a) it is ok to leave performance on the table in the interest of clarity or (b) you create a little sandbox in which you use functional programming for clarity, and something outside your sandbox compensates for the performance cost.

React and Om are examples of (b). They let you write pure functions that return the entire DOM. Then, outside that sandbox, a DOM differ makes it so you don't force the browser to re-render the entire DOM every time.

Also, there is a lot of cases where you need a total language with all its guarantees.