Generalising the mean of a list
30 Jan 2022
The free semigroup on a set X, denoted X^{+}, is the set of nonempty ordered lists whose items are elements of X. For example, ℤ^{+} is the set of nonempty lists of integers, such as [3, 1] and [5, 6, 12]. For all sets X, there is a “length” function Λ : X^{+} → ℕ_{>0} that calculates the length of a list. (Here, ℕ_{>0} is the set of natural numbers excluding 0.)
If (X, +) is a semigroup, then there is also a “sum” function ∑ : X^{+} → X that combines every item in a list using the semigroup operation. For example, ∑[x, y, z] = x + y + z. Because (X, +) can be any semigroup, the + symbol does not necessarily denote addition. It could also be multiplication, function composition, maximum, minimum, greatest common divisor, etc. Any associative operation will do.
Every semigroup (X, +) admits a binary function p : ℕ_{>0} × X → X, which I will denote by juxtaposition; nx, short for p(n, x), is the result of combining x with itself n times. For example, 3x = p(3, x) = x + x + x.
In summary, for any semigroup (X, +), you can calculate the length of a list in X^{+} with Λ and the “sum” of a list in X^{+} with ∑. You can also “multiply” nonzero natural numbers with the elements of X. Using these three operations, I will define a binary relation ⤚ between X^{+} and X. If a is an element of X^{+} and x is an element of X, then a ⤚ x if and only if ∑(a) = Λ(a)x. In this case, we say that x is a mean of a.
That’s a very general definition (it works for any semigroup), so here are some concrete examples:

If the semigroup in question is (ℝ, +), then a ⤚ x means that x is the arithmetic mean of the list a.

If the semigroup is (ℝ_{>0}, ·), the set of positive real numbers under multiplication, then a ⤚ x means that x is the geometric mean of the list a.

If the semigroup is (ℤ, max), then a ⤚ x means that x is the maximum of a. This is because ∑(a) calculates the maximum of a, and nx = x for all n (i.e. the semigroup is idempotent). Similarly, in (ℤ, min), a ⤚ x means that x is the minimum of a.

If the semigroup is itself a free semigroup, i.e. X^{+} for some set X, then ∑ : X^{++} → X^{+} is a function that flattens a list of lists. Unlike the previous examples, ⤚ in this case is not a total relation, or in other words, not every list of lists has a mean. Those that do have a mean are of a very special variety. For example, the mean of [[x, y], [z, x, y, z], [x, y, z]] is [x, y, z], because flattening the former gives the same result as repeating the latter 3 times.

If the semigroup is (𝔹, ∧), the set of boolean values under logical conjunction, then a ⤚ 1 if every item in a is 1; otherwise, a ⤚ 0. Conversely, in (𝔹, ∨), a ⤚ 1 if any item in a is 1.

If the semigroup operation is modular addition, then means are not always unique. For example, in the cyclic group ℤ/3ℤ, the mean of [1, 2, 0] can be either 0, 1, or 2, because 1 + 2 + 0 = 3(0) = 3(1) = 3(2). This shows why ⤚ has to be defined as a relation and not a function. A list doesn’t always have exactly one mean; it could have 0, or 3, or even an infinite number.

If the semigroup is (ℝ → ℝ, ∘), then a ⤚ f means that performing every function in the list a is equivalent to repeating f, Λ(a) times. Finding a mean of a list of functions amounts to composing every function in the list and then taking the functional root.
A semigroup (X, +) is meaningful if the ⤚ relation can be made into a function; for every a ∈ X^{+}, there is a unique μ(a) ∈ X such that a ⤚ μ(a). Instead of saying that μ(a) is a mean of a, we can now confidently say that μ(a) is the mean of a. In examples 1, 2, 3, and 5 above, the semigroup in question is meaningful, but in examples 4, 6, and 7, it is not. A semigroup that is not meaningful is called meaningless. In a meaningless semigroup, the ⤚ relation is either nontotal or multivalued, so it cannot be made into a function.
A meanpreserving homomorphism from (X, +) to (Y, ∗) is a function h : X → Y with the following properties:
 h is a semigroup homomorphism: h(x + y) = h(x) ∗ h(y).
 h respects the mean relation: a ⤚ x implies h(a) ⤚ h(x), where h(a) denotes the application of h to every item in a.
One can prove that all semigroup homomorphisms are meanpreserving. Assume h is a homomorphism from (X, +) to (Y, ∗), a ∈ X^{+}, and x ∈ X. Then a ⤚ x ⇔ ∑(a) = Λ(a)x ⇒ h(∑(a)) = h(Λ(a)x). Because h is a homomorphism, h(nx) = nh(x), and h(∑(a)) = ∑(h(a)). The latter can be demonstrated by replacing the list a with its components: h(∑(a)) = h(a_{0} + a_{1} + … a_{n}) = h(a_{0}) ∗ h(a_{1}) ∗ … h(a_{n}) = ∑(h(a)). Therefore, the equality h(∑(a)) = h(Λ(a)x) is equivalent to ∑(h(a)) = Λ(a)h(x). Applying h to a list does not change its length, so we get that ∑(h(a)) = Λ(h(a))h(x) ⇔ h(a) ⤚ h(x). In summary, for any homomorphism h, a ⤚ x implies h(a) ⤚ h(x).
I will use the term divisible semigroup for a semigroup X satisfying the following property: for every x ∈ X and n ∈ ℕ_{>0}, there is a unique y ∈ X such that ny = x. The element y is called the nth factor of x, and is written x / n. It isn’t difficult to see that all divisible semigroups are meaningful, with μ(a) = ∑(a) / Λ(a). But the reverse implication (that all meaningful semigroups are divisible) is not true.
One example of a semigroup which is meaningful but not divisible is (ℝ_{>2}, ·). This is a valid semigroup, because the product of numbers greater than 2 will also be greater than 2. It is meaningful, because the geometric mean of a list of numbers greater than 2 will also be greater than 2. But it is not divisible, because the nth root of a number greater than 2 is not necessarily greater than 2. (nth roots are the equivalent of nth factors in multiplicative semigroups.) Therefore, the concept of “meaningfulness” can be thought of as a generalisation of divisibility.
Meaningfulness does imply divisibility if the semigroup has an identity element. To find the factor x / n, you can construct a list with length n and sum x, and then calculate the mean. In a monoid, it is always possible to construct a list with these properties by appending the identity: [x], [x, e], [x, e, e], and so on. So for monoids, meaningfulness and divisibility are equivalent.
Given two meaningful semigroups (X, +) and (Y, ∗), their direct product is also meaningful. For any list of pairs c ∈ (X × Y)^{+}, we can split c into two lists a ∈ X^{+} and b ∈ Y^{+} by putting the first element of each pair in a and the second element of each pair in b. Then the pair (μ(a), µ(b)) is a mean of c, because Λ(c)(μ(a), μ(b)) = (Λ(c)μ(a), Λ(c)μ(b)) = (Λ(a)μ(a), Λ(b)μ(b)) = (Σ(a), Σ(b)) = Σ(c). It is also clearly the only mean of c, because μ(a) and μ(b) are the only means of a and b respectively. This means that the direct product of any two meaningful semigroups is meaningful.
Many other constructions also preserve meaningfulness.

Given a meaningful semigroup (X, +), the opposite semigroup (X, (x, y) ↦ y + x) is also meaningful, with μ(a) in the latter being equivalent to μ(reverse(a)) in the former.

The semigroup formed by augmenting a meaningful semigroup (X, +) with an absorbing element ∞ (such that ∞ + x = x + ∞ = ∞) is meaningful. The mean of any list not containing ∞ is already defined, and the mean of a list containing ∞ is just ∞.

If a meaningful semigroup (X, +) can be divided into two disjoint subsets Y and Z such that both are closed under the + operation and Y ∪ Z = X, then both (Y, +) and (Z, +) are meaningful. The mean of a list in Y^{+} must be in Y because Z is closed, and vice versa. For example, the meaningful semigroup (ℝ_{>0}, ·) can be divided into two subsemigroups (ℝ_{(0, 1]}, ·) and (ℝ_{>1}, ·). Both are meaningful by the preceding theorem.
One construction that does not preserve meaningfulness is the addition of an identity element. For example, (ℝ_{>1}, +) is meaningful, but (ℝ_{>1} ∪ {0}, +) is not, even though it is a valid semigroup. The arithmetic mean of [0, 2] is not contained in ℝ_{>1} ∪ {0}.