Which ofc is why we can say an algorithm is incorrect; correctness is not part of the definition.
To be considered an algorithm, a method must be:
1. Finite. It must terminate after a finite number of steps. (The program, i.e. the expression of the method must also be finite.)
2. Definite. Each step must be precisely defined, requiring no 'interpretation' or ingenuity on behalf of the executor of the algorithm.
3. It must take zero or more inputs, and these inputs must be well-defined.
4. It must produce one or more well-defined outputs.
5. It must be an "effective method", which (in addition): 5a. Must always produce a result 5b. The result must always be a correct result 5c. It must always arrive at a result after a finite number of steps 5d. It must work for all defined categories of inputs
This is (almost) literally page 1, chapter 1, book 1 of Knuth. Most probably one of the first things you'll ever be taught on any respectable computer science course.
Of course, the term is used more vaguely and everyone still knows what we all mean, but the above is the widely-accepted formal definition.
In mathematics and computer science that is indeed the proper definition.
And HackerNews is definitely a mathsy/computersciencey sort of a place, so it might be nice to use the word correctly here.
Out of interest (or is it flippancy?), where do you use the word algorithm, when not talking about mathematics or computing?
Do you consider a discussion about heuristics vs algorithms (which is what this thread is above), on a site like Hacker News, to be 'general language'...? Or would you consider that this might indeed be a suitable context to comment on their proper formal definitions?
Another way of looking at it: If we have a function f(x), we can have an algorithm that computes f(x). In fact, there are an infinite number of algorithms that compute f(x). But there are also an infinite number of algorithms that compute f'(x), where f'(x) is an approximation of f(x). And such algorithms are called heuristics for f(x).
Suppose you have some algorithm for integrating a function exactly (e.g., by restricting the algorithm's domain to simple polynomials and then doing it symbolically).
Now, you can approximate this algorithm's output with the trapezoid rule. However, the trapezoid rule also looks like an algorithm to me. Is the difference that one is exact and the other is not?
Does it switch from being a heuristic to an algorithm if we switch the specification from
integrate(function, a, b) integrate_with_error_tolerance(function, a, b, tol)?(Someone will correct me if I have the facts wrong, but even if this isn't true for differential equations, there are certainly areas of mathematics where it is true.)
Another example might be the travelling salesman problem, where we know how to calculate the exact solution for every instance, but it is prohibitively expensive (unless P=NP). There are a number of heuristics that "can find solutions for extremely large problems (millions of cities) within a reasonable time which are with a high probability just 2–3% away from the optimal solution" (Source: https://en.wikipedia.org/wiki/Travelling_salesman_problem)
In that case, you can always come up with a solution, it's just not always the best solution.
If algorithm A is a heuristic for f(x), what we really mean by that is something like "A is an algorithm for g(x), where g(x) ~= f(x) with an epsilon small enough to be useful for all x" or "A is an algorithm of g(x), where g(x) = f(x) for some x (hopefully most of the values of x we care about)".