📌 Talk Outline
flowchart TD
A["MTL-rings
& Generalized MTL-rings"]
B["Galois rings
(commutative MTL-rings)"]
C["Cryptography
DH & ElGamal via Galois logarithm"]
D["Neural networks
Ideal function activation"]
E["Open problem solved
Perfect pseudo MTL-algebra not a chain"]
F["New hierarchy
MTL ⇒ strong semi-MTL ⇒ semi-MTL"]
A --> B
A --> C
A --> D
A --> E
A --> F
style A fill:#2c7da0, stroke:#1e4a6e, stroke-width:2px, color:#fff
style B fill:#eef7ff, stroke:#2c7da0, stroke-width:2px
style C fill:#eef7ff, stroke:#2c7da0, stroke-width:2px
style D fill:#eef7ff, stroke:#2c7da0, stroke-width:2px
style E fill:#eef7ff, stroke:#2c7da0, stroke-width:2px
style F fill:#eef7ff, stroke:#2c7da0, stroke-width:2px
This diagram illustrates the different directions explored in this talk, all stemming from the central notion of (generalized) MTL-rings.
📖 Introduction and motivation
MTL-rings are rings (not necessarily commutative) whose lattice of two-sided ideals forms an MTL algebra (or pseudo‑MTL algebra). Initiated by Belluce & Di Nola (MV‑algebras) and Esteva & Godo (monoidal t‑norm logic), they provide a deep bridge between ring theory, fuzzy logic, and universal algebra.
This work unifies:
- ✅ A non‑commutative generalization of MTL-rings.
- ✅ The use of Galois rings in cryptography via an original discrete logarithm.
- ✅ The introduction of ideal functions as an output layer for neural networks, an alternative to Softmax.
- ✅ A solution to an open problem: existence of perfect pseudo‑MTL algebras that are not chains.
- ✅ An extension of the theory to semi-MTL and strong semi-MTL rings.
🧩 Preliminaries: pseudo‑MTL algebras
Definition (Pseudo‑MTL algebra). A structure \((A,\wedge,\vee,\odot,\to,\hookrightarrow,0,1)\) such that:
- \((A,\wedge,\vee,0,1)\) is a bounded lattice;
- \((A,\odot,1)\) is a monoid;
- Residuation: \(x\odot y \le z \iff x \le y\to z \iff y \le x\hookrightarrow z\);
- Pseudo‑prelinearity: \((x\to y)\vee(y\to x) = (x\hookrightarrow y)\vee(y\hookrightarrow x) = 1\).
If \(\odot\) is commutative and \(\to = \hookrightarrow\), we obtain an
MTL algebra.
Perfect element & perfect algebra. Write \(\neg a = a\to 0\) and \(\sim a = a\hookrightarrow 0\). A pseudo‑MTL algebra is perfect if it is good, local, and for every element \(a\):
\[
\operatorname{ord}(a) < \infty \;\Longleftrightarrow\; \operatorname{ord}(\neg a) = \infty \;\Longleftrightarrow\; \operatorname{ord}(\sim a) = \infty.
\]
A chain is an algebra whose order is total.
🔗 MTL-rings and generalized MTL-rings
Let \(R\) be a unital ring (not necessarily commutative) and \(\operatorname{Id}(R)\) the set of its two‑sided ideals. Define:
\[
A \wedge B = A \cap B,\qquad A \vee B = A + B,\qquad A \odot B = A \cdot B,
\]
\[
A \to B = \{x\in R : xA \subseteq B\},\qquad A \hookrightarrow B = \{x\in R : Ax \subseteq B\}.
\]
Generalized MTL-ring. \(R\) is a generalized MTL-ring if for all two‑sided ideals \(A,B\):
\[
(A\to B) + (B\to A) = (A\hookrightarrow B) + (B\hookrightarrow A) = R.
\]
When \(R\) is commutative, we simply speak of an MTL-ring.
Theorem. \(R\) is a generalized MTL-ring if and only if the structure \(\mathcal{A}(R)=(\operatorname{Id}(R),\wedge,\vee,\odot,\to,\hookrightarrow,\{0\},R)\) is a pseudo‑MTL algebra.
Proposition (local characterization). Let \(R\) be a unital local ring. Then \(R\) is a generalized MTL-ring \(\iff\) \(R\) is a valuation ring (its ideals are totally ordered by inclusion).
Example without unit (still generalized MTL). Let \(R\) be commutative unital and \(M\) a simple \(R\)-module. The ring \(\widehat{R} = R \times M\) with multiplication
\[
(r_1,m_1)(r_2,m_2) = (r_1r_2,\; r_1m_2)
\]
is non‑commutative, local, has no identity element, but it is a generalized MTL-ring (its ideal lattice is not a chain when \(|M|\ge 2\)).
🏺 Galois rings: a family of commutative MTL-rings
Galois ring \(GR(p^k,m)\). It is a finite commutative local ring of characteristic \(p^k\) with residue field \(\mathbb{F}_{p^m}\). Construction:
\[
GR(p^k,m) = \mathbb{Z}_{p^k}[X]/(P(X)),
\]
where \(P(X)\) is a monic polynomial of degree \(m\) that is irreducible modulo \(p\) (a basic polynomial).
Examples: \(\mathbb{Z}_{p^k}=GR(p^k,1)\) and \(GR(2^2,2) \cong \mathbb{Z}_4[X]/(X^2+X+1)\).
Theorem. Every Galois ring \(GR(p^k,m)\) is a commutative MTL-ring.
Sketch. The ideal lattice of \(GR(p^k,m)\) is a chain:
\[
(0) \subset (p^{k-1}) \subset (p^{k-2}) \subset \cdots \subset (p) \subset R.
\]
For any two ideals \(I,J\), one contains the other, hence one of \((I\to J)\) or \((J\to I)\) equals \(R\); the condition \((I\to J)+(J\to I)=R\) holds automatically. Commutativity is inherited from the construction.
🔐 Cryptography: discrete logarithm in Galois rings
Let \(\mathcal{T}\) be the Teichmüller set of \(GR(p^k,m)\). Every element \(\alpha\) can be uniquely written as:
\[
\alpha = \sum_{i=0}^{k-1} \nu_i p^i,\qquad \nu_i \in \mathcal{T},\ \nu_0 \neq 0.
\]
The multiplicative group \(\mathcal{T}^* = \langle\xi\rangle\) is cyclic of order \(p^m-1\). Hence each \(\nu_i = \xi^{x_i}\) (with the convention \(\xi^{-}\) for 0). We denote \(\alpha = \xi^{(x_0,x_1,\dots,x_{k-1})}\).
Galois logarithm. The tuple \((x_0,\dots,x_{k-1})\) is the Galois logarithm of \(\alpha\) in the base \(\epsilon = \xi^{(1,1,\dots,1)}\). Write \(\log_\epsilon \alpha = (x_0,\dots,x_{k-1})\).
🔁 Diffie‑Hellman key exchange
- Private keys: \(X,Y \in \{0,\dots,p^m-2\}^k\).
- Public keys: \(A = \epsilon^{X},\; B = \epsilon^{Y}\).
- Shared secret: \(K = \epsilon^{(x_i y_i \bmod (p^m-1))}\).
🔏 ElGamal encryption
- Alice: private key \(X_a\), public key \(K_a = \epsilon^{X_a}\).
- Bob chooses a random \(Y\) and sends \(C = [\epsilon^{Y},\; M \cdot K_a^{Y}]\).
- Alice decrypts: \(M = (M \cdot K_a^{Y}) \cdot ((\epsilon^{Y})^{X_a})^{-1}\).
Concrete example in \(GR(2^2,2)\) with \(\xi^2+\xi+1=0\).
Alice picks \(X_a = (2,1)\) → \(K_a = \xi^2 + 2\xi\).
Bob wants to send \(M = \xi\) with \(Y = (1,2)\) → ciphertext is \([\xi+2\xi^2,\; 3]\).
🧠 Neural networks: ideal functions as activation
Ideal function. Let \(R\) be an MTL-ring. Define \(f : R \to \operatorname{Id}(R)\) by \(f(x) = \langle x \rangle\) (the principal ideal generated by \(x\)).
For an \(n\)-class classification task, choose \(n\) distinct primes \(p_1,\dots,p_n\) and the ring
\[
R = \bigcap_{i=1}^{n} \mathbb{Z}_{(p_i)} = \left\{ \prod_{i=1}^{n} p_i^{m_i} \cdot \frac{a}{b} \in \mathbb{Q} \;:\; m_i\in\mathbb{N},\ a,b\in\mathbb{Z},\ p_i\nmid a,b \right\}.
\]
Then \(\operatorname{Id}(R) \cong \mathbb{N}^n\) via \((m_1,\dots,m_n) \leftrightarrow \prod p_i^{m_i} R\).
\(\operatorname{Id}(R)\) is a non‑linear, local, perfect MTL algebra.
⚡ Advantages over Softmax
- Single output neuron: Softmax needs \(n\) neurons; the ideal function uses one neuron with \(n\) weights.
- Linear gradient: with quadratic loss, \(\frac{\partial \delta}{\partial w} = M + w - \bar{X}\) → no vanishing/exploding gradient.
- Very fast convergence: one step with \(\eta = 1\) gives the exact optimum in exact arithmetic.
Weight update: \(w^{(t+1)} = \big\lfloor w^{(t)} - \eta (M + w^{(t)} - \bar{X}) \big\rceil\) (round to integers). Empirical results on UCI datasets show faster convergence than sigmoid or Softmax.
🏆 Solved open problem: perfect non‑chain pseudo‑MTL algebras
Problem (long standing open). Do there exist perfect pseudo‑MTL algebras that are not chains?
Answer: YES. The following 4‑step construction (Mouchili, 2026) establishes it.
- Commutative non‑chain MTL-ring: Let \(V_1, V_2\) be two valuation domains with the same fraction field, neither containing the other (Nagata’s example). Their intersection \(R = V_1 \cap V_2\) has two incomparable maximal ideals → \(R\) is not a chain ring.
- Non‑commutative extension: Set \(\mathcal{R} = R[i]\) with \(i^2 = -1\) and multiplication \((a+ib)(a'+ib') = (aa'-bb') + i(ba'+ab')\). Ideals of \(\mathcal{R}\) are of the form \(I[i] = \{x+iy : x,y\in I\}\) for \(I\) an ideal of \(R\).
- One verifies that \(\mathcal{R}\) is a generalized MTL-ring and its ideal lattice is not a chain (because \(R\) is not).
- Perfectness: \(\mathcal{R}\) is an integral domain → every nonzero ideal has infinite order. The annihilator of any nonzero ideal is \(\{0\}\) (order 1). Hence \(\operatorname{Id}(\mathcal{R})\) is a perfect pseudo‑MTL algebra that is not a chain.
✅ This answers negatively a question that remained open in the literature of residuated t‑norm algebras.
📐 Extension: semi‑MTL and strong semi‑MTL rings
For an ideal \(I\) of a commutative ring \(R\), denote \(I^* = \operatorname{Ann}(I) = \{r\in R : rI = 0\}\) (the annihilator).
Strong semi‑MTL ring. \(R\) (commutative unital) satisfies for all ideals \(I,J\):
\[
(I^* \to J^*) + (J^* \to I^*) = R.
\]
Semi‑MTL ring. Weaker condition:
\[
\big[(I^* \to J^*) + (J^* \to I^*)\big]^* = \{0\}.
\]
Strict hierarchy:
\[
\text{MTL-ring} \;\Longrightarrow\; \text{strong semi‑MTL} \;\Longrightarrow\; \text{semi‑MTL}.
\]
Separating example (strong semi‑MTL but not MTL). \(R_1 = \mathbb{Z}_2[x,y]/(x^2,y^2,xy)\) (8 elements). Its annihilator ideals are \(\{0\}, (x,y), R\): a chain. The ring is local → strong semi‑MTL. However \((x)\) and \((y)\) are incomparable → \(R_1\) is not MTL.
Separating example (semi‑MTL but not strong). \(R_2 = \mathbb{Z}_2[x,y]/(xy)\). For \(I=(x), J=(y)\): \(I^*=(y), J^*=(x)\), \((x)\to(y)=(y)\), \((y)\to(x)=(x)\) → sum \((x,y) \neq R\) → not strong. Yet \(((x,y))^* = \{0\}\) → the semi‑MTL condition holds.
Local characterization. Let \(R\) be a commutative local ring with maximal ideal \(M\). Then \(R\) is strong semi‑MTL \(\iff\) the set of annihilator ideals is totally ordered by inclusion.
Open perspectives: preservation under homomorphisms / localizations, decomposition for Noetherian rings, subdirect representation, non‑commutative generalization.
✅ Positive result: finite direct products preserve both properties (strong semi‑MTL and semi‑MTL).
🎯 Conclusion and future work
- MTL-rings provide a coherent unification of ring theory, algebraic logic, cryptography, and deep learning.
- Galois rings form a natural family of commutative MTL-rings, and the Galois logarithm enables concrete Diffie‑Hellman and ElGamal implementations.
- The ideal function is a rigorous alternative to Softmax: single output neuron, linear gradient, accelerated convergence.
- Solution of an open problem: existence of perfect pseudo‑MTL algebras that are not chains via a non‑commutative extension of Nagata’s construction.
- Introduction of semi‑MTL and strong semi‑MTL rings, with a strict hierarchy and separating examples.
🔭 Future work:
- Deeper connections between residuated structures and deep architectures.
- Extension of ideal functions to convolutional and recurrent networks.
- Solving open problems on semi‑MTL rings (homomorphisms, localizations, subdirect representation).
- Full non‑commutative generalization of semi‑MTL rings.
« MTL-rings create an elegant bridge between algebraic logic, cryptography, and artificial intelligence. »
📚 Main references
- L. P. Belluce, A. Di Nola, Commutative rings whose ideals form an MV-algebra, Math. Log. Q. (2009).
- P. Flondor, G. Georgescu, A. Iorgulescu, Pseudo t-norms and pseudo BL-algebras, Soft Computing 5 (2001).
- F. Esteva, L. Godo, Monoidal t-norm logic, Fuzzy Sets Syst. 124 (2001).
- S. Mouchili, S. Atamewoue, S. Ndjeya, A non-commutative generalization of MTL-rings, J. Algebraic Syst. 13(3) (2025).
- S. Mouchili, Discrete Logarithm in Galois Rings, Mathematics and Statistics 6(3) (2018).
- S. Mouchili, A suggested solution to some open problems in pseudo MTL-algebras (2026).
- S. Mouchili, On semi-MTL and strong semi-MTL rings: extending the theory of MTL-rings (preprint 2026).
- M. Nagata, Theory of commutative fields, AMS Transl. (1953).