Unit 2: Operations on Fuzzy Sets: Complements, Unions & Intersections

1. Zadeh's Original Set Theoretic Operations

In 1965, Lotfi A. Zadeh defined the elementary operations on fuzzy sets $A$ and $B$ in universe $X$ point-by-point through their membership functions:

Definition 2.1 (Standard Fuzzy Operations): For all $x \in X$:

1. Fuzzy Complement ($A^c$):

1. Fuzzy Intersection ($A \cap B$):

1. Fuzzy Union ($A \cup B$):

1. Fuzzy Inclusion ($A \subseteq B$):

2. Preservation of Algebraic Lattice Properties

Under standard operations $(\min, \max, 1 - \cdot)$, the set of all fuzzy subsets $\mathcal{F}(X)$ forms a distributive, bounded pseudo-complemented lattice (a de Morgan algebra / Kleene algebra).

Satisfied Properties ($\forall A, B, C \in \mathcal{F}(X)$):

1. Involution (Double Negation): $(A^c)^c = A$

2. Idempotence: $A \cup A = A$, and $A \cap A = A$

3. Commutativity: $A \cup B = B \cup A$, and $A \cap B = B \cap A$

4. Associativity: $(A \cup B) \cup C = A \cup (B \cup C)$, and $(A \cap B) \cap C = A \cap (B \cap C)$

5. Absorption: $A \cup (A \cap B) = A$, and $A \cap (A \cup B) = A$

6. Distributivity:

7. Identity Elements: $A \cup \emptyset = A$, and $A \cap X = A$

3. De Morgan's Laws for Fuzzy Sets

Theorem 2.1 (De Morgan's Laws): For any fuzzy sets $A, B \in \mathcal{F}(X)$:

Proof:

For any $x \in X$, let $a = \mu_A(x)$ and $b = \mu_B(x)$. For the first identity:

Notice that if $a \ge b$, then $\max(a, b) = a$, so $1 - \max(a, b) = 1 - a$. At the same time, $1 - a \le 1 - b$, so $\min(1 - a, 1 - b) = 1 - a$. In general, for any two real numbers $a, b$:

Therefore:

Since this holds for all $x \in X$, $(A \cup B)^c = A^c \cap B^c$. The second identity follows symmetrically from $1 - \min(a, b) = \max(1 - a, 1 - b)$. $\blacksquare$

1. General Axioms of Fuzzy Complements

Why must a complement be $1 - a$? In 1980, mathematicians generalized the notion of complementation by defining an axiomatic system for mapping $c: [0, 1] \to [0, 1]$.

Definition 2.2 (Axiomatic Fuzzy Complement): A function $c: [0, 1] \to [0, 1]$ is a fuzzy complement if it satisfies the following two axiomatic requirements:

• Axiom C1 (Boundary Conditions): $c(0) = 1$ and $c(1) = 0$.
• Axiom C2 (Monotonicity): For all $a, b \in [0, 1]$, if $a \le b$, then $c(a) \ge c(b)$ (strictly non-increasing).

A complement is called involutive if it satisfies:

• Axiom C3 (Involution): $c(c(a)) = a$ for all $a \in [0, 1]$.
• Axiom C4 (Continuity): $c$ is a continuous function.

(Theorem: Any involutive complement satisfying C1-C2 is strictly decreasing and continuous!)

2. Parameterized Families of Fuzzy Complements

1. Sugeno's Complement Family:

For parameter $\lambda \in (-1, \infty)$:

• When $\lambda = 0$: $c_0(a) = 1 - a$ (Zadeh's standard complement).
• When $\lambda \to \infty$: $c_\infty(a) \to 0$ for all $a > 0$.
• When $\lambda \to -1$: $c_{-1}(a) \to 1$ for all $a < 1$.

2. Yager's Complement Family:

For parameter $w \in (0, \infty)$:

• When $w = 1$: $c_1(a) = 1 - a$ (standard complement).
• Every member of Yager's family is strictly involutive: $c_w(c_w(a)) = (1 - (1 - a^w))^{1/w} = a$.

3. Equilibrium Points

Definition 2.3 (Equilibrium Point): An equilibrium point of a fuzzy complement $c$ is a value $e \in [0, 1]$ that is its own complement:

Theorem 2.2 (Existence and Uniqueness of Equilibrium): Every continuous fuzzy complement $c$ possesses a unique equilibrium point $e \in (0, 1)$.

Proof:

Define the auxiliary function $g(a) = c(a) - a$ on the domain $[0, 1]$.

1. At $a = 0$: $g(0) = c(0) - 0 = 1 - 0 = 1 > 0$.
2. At $a = 1$: $g(1) = c(1) - 1 = 0 - 1 = -1 < 0$.
3. Since $c$ is continuous, $g$ is continuous on $[0, 1]$.

By the Intermediate Value Theorem, there exists at least one $e \in (0, 1)$ such that $g(e) = 0 \implies c(e) = e$. Because $c$ is strictly decreasing, $g(a) = c(a) - a$ is strictly decreasing, so the root $e$ is strictly unique! $\blacksquare$

• For Zadeh's complement $c(a) = 1 - a$: $1 - e = e \implies e = 0.5$.
• For Sugeno's complement: $\frac{1 - e}{1 + \lambda e} = e \implies \lambda e^2 + 2e - 1 = 0 \implies e = \frac{\sqrt{1 + \lambda} - 1}{\lambda}$.

1. The Axiomatization of Fuzzy Intersections

The concept of a triangular norm ($t$-norm) was introduced by Karl Menger (1942) in the study of probabilistic metric spaces and later adopted by Schweizer and Sklar to generalize fuzzy intersections.

Definition 2.4 (Triangular Norm / t-Norm): A binary operator $i: [0, 1] \times [0, 1] \to [0, 1]$ (often denoted $T(a, b)$ or $a \top b$) is a t-norm if it satisfies for all $a, b, c, d \in [0, 1]$:

1. Boundary Condition: $T(a, 1) = a$ (1 is the neutral identity element).
2. Monotonicity: If $a \le c$ and $b \le d$, then $T(a, b) \le T(c, d)$.
3. Commutativity: $T(a, b) = T(b, a)$.
4. Associativity: $T(T(a, b), c) = T(a, T(b, c))$.

Note: From boundary and monotonicity: $T(a, 0) \le T(1, 0) = T(0, 1) = 0 \implies T(a, 0) = 0$.

2. The Four Fundamental Archetypal t-Norms

1. Minimum (Standard Zadeh Intersection):

This is the largest possible t-norm: for any t-norm $T$, $T(a, b) \le \min(a, b)$. It is the unique idempotent t-norm: $T(a, a) = a \iff T = T_{\min}$.

2. Algebraic Product:

Strictly positive for all $a, b > 0$.

3. Bounded Difference (Łukasiewicz t-Norm):

Nilpotent t-norm: $T(a, b) = 0$ can occur even when $a > 0$ and $b > 0$.

4. Drastic Product:

This is the smallest possible t-norm: for any t-norm $T$, $T_D(a, b) \le T(a, b)$.

3. Universal Ordering Chain of Archetypal t-Norms

Theorem 2.3 (Ordering of Fundamental t-Norms): For all $a, b \in [0, 1]$:

Proof:

1. $T_D \le T$ is immediate from Definition 2.4.
2. For $T_{\text{Luk}} \le T_{\text{prod}}$:

Notice that $(1 - a)(1 - b) \ge 0 \implies 1 - a - b + ab \ge 0 \implies ab \ge a + b - 1$. Since $ab \ge 0$, we have $ab \ge \max(0, a + b - 1) = T_{\text{Luk}}(a, b)$.

1. For $T_{\text{prod}} \le T_{\min}$:

Since $b \le 1$, $ab \le a$. Since $a \le 1$, $ab \le b$. Thus $ab \le \min(a, b) = T_{\min}(a, b)$. $\blacksquare$

1. The Axiomatization of Fuzzy Unions

The dual operation to a $t$-norm is a triangular conorm ($t$-conorm or $s$-norm).

Definition 2.5 (Triangular Conorm / s-Norm): A binary operator $u: [0, 1] \times [0, 1] \to [0, 1]$ (denoted $S(a, b)$ or $a \bot b$) is an s-norm if it satisfies for all $a, b, c, d \in [0, 1]$:

1. Boundary Condition: $S(a, 0) = a$ (0 is the neutral identity element).
2. Monotonicity: If $a \le c$ and $b \le d$, then $S(a, b) \le S(c, d)$.
3. Commutativity: $S(a, b) = S(b, a)$.
4. Associativity: $S(S(a, b), c) = S(a, S(b, c))$.

Note: $S(a, 1) = 1$ for all $a \in [0, 1]$.

2. The Four Fundamental Archetypal s-Norms

1. Maximum (Standard Zadeh Union):

This is the smallest possible s-norm: for any s-norm $S$, $\max(a, b) \le S(a, b)$. It is the unique idempotent s-norm: $S(a, a) = a \iff S = S_{\max}$.

2. Algebraic Sum (Probabilistic Sum):

3. Bounded Sum (Łukasiewicz s-Norm):

4. Drastic Sum:

This is the largest possible s-norm: for any s-norm $S$, $S(a, b) \le S_D(a, b)$.

3. Generalized De Morgan Duality

Theorem 2.4 (Duality Theorem): Let $c$ be an involutive fuzzy complement. For every t-norm $T$, the dual operator defined by:

is an s-norm. Conversely, for every s-norm $S$:

is a t-norm. The pair $(T, S, c)$ satisfies generalized De Morgan's laws.

Universal Ordering Chain of Archetypal s-Norms:

1. The Spectrum Between Intersection and Union

Notice the strict bounds governing norm operations:

$t$-norms represent strict conjunction ("AND"), while $s$-norms represent full disjunction ("OR"). However, human decision-making frequently requires compensatory aggregation (a trade-off where a high score in one criterion compensates for a lower score in another). This motivates averaging operators $M(a, b)$ that lie strictly between min and max:

2. Generalized Means (Power Means)

For elements $a_1, a_2, \dots, a_n \in [0, 1]$ and weights $w_i \ge 0$ with $\sum w_i = 1$:

• $p \to -\infty$: $M_{-\infty} = \min(a_1, \dots, a_n)$ (pure intersection).
• $p = -1$: Harmonic Mean $M_{-1} = \frac{1}{\sum \frac{w_i}{a_i}}$.
• $p \to 0$: Geometric Mean $M_0 = \prod_{i=1}^n a_i^{w_i}$.
• $p = 1$: Arithmetic Mean $M_1 = \sum_{i=1}^n w_i a_i$.
• $p = 2$: Quadratic (RMS) Mean $M_2 = \sqrt{\sum w_i a_i^2}$.
• $p \to +\infty$: $M_{+\infty} = \max(a_1, \dots, a_n)$ (pure union).

3. Ordered Weighted Averaging (OWA) Operators

Introduced by Ronald R. Yager (1988), the OWA operator decouples weights from individual criteria and associates them with magnitudes:

Definition 2.6 (OWA Operator): An OWA operator of dimension $n$ is a mapping $F_w: [0, 1]^n \to [0, 1]$ associated with weighting vector $w = (w_1, \dots, w_n)$ where $w_i \in [0, 1]$ and $\sum w_i = 1$:

where $b_j$ is the $j$-th largest element of the collection $\{a_1, \dots, a_n\}$ ($b_1 \ge b_2 \ge \dots \ge b_n$).

Extreme Cases:

• $w = (1, 0, \dots, 0) \implies F_w(a) = b_1 = \max(a_i)$ (Pure OR).
• $w = (0, 0, \dots, 1) \implies F_w(a) = b_n = \min(a_i)$ (Pure AND).
• $w = (1/n, 1/n, \dots, 1/n) \implies F_w(a) = \frac{1}{n} \sum a_i$ (Standard Arithmetic Mean).

The degree of "orness" of an OWA operator is measured by:

• $\text{orness} = 1$ for Max; $\text{orness} = 0$ for Min; $\text{orness} = 0.5$ for Arithmetic Mean.