1 Introduction↩︎

1.1 The problem and its origin↩︎

The sequence of real normed division algebras — the real numbers \(\mathbb{R}\), the complex numbers \(\mathbb{C}\), the quaternions \(\mathbb{H}\), and the octonions \(\mathbb{O}\) — occupies dimensions one, two, four, and eight, a pattern that admits a structural explanation. Each algebra in the sequence arises from its predecessor by the Cayley–Dickson construction [1], which doubles the dimension at every step while sacrificing one algebraic property: \(\mathbb{R}\) is an ordered field; \(\mathbb{C}\) loses ordering but retains commutativity; \(\mathbb{H}\) loses commutativity but retains associativity; \(\mathbb{O}\) loses associativity but retains the weaker alternative law. What this construction cannot produce is an algebra of dimension three. The doubling mechanism jumps from two to four, and the gap between them is not a missing algebra in the sequence but a structural impossibility — one that Frobenius made precise.

A three-dimensional real division algebra — an object one might naturally call a trinion, by analogy with the Latin sequence singuli, bini, trini, quaterni — does not exist. The proof is short: any such candidate must contain \(\mathbb{C}\) as a subalgebra, after which any additional generator \(j\) must satisfy \(j^2 = -\lambda^2\) for some \(\lambda \in \mathbb{R}^*\), making \((j/\lambda)^2 = -1\) and producing a second copy of \(\mathbb{C}\) independent of the first. The three elements \(1\), \(i\), and \(j/\lambda\) are then linearly independent over \(\mathbb{R}\) and their product \(ij/\lambda\) contributes a fourth, driving the dimension to at least four and leaving no room for a trinion.

A self-contained presentation of this argument appears in [3]; the present paper is, in a sense, the natural sequel.

The question is simply this: Frobenius’ theorem tells us that no three-dimensional division algebra over \(\mathbb{R}\) exists, but it says nothing about three-dimensional associative algebras in general. The class of all such algebras — the near-trinions — has exactly six members, as Theorem 2 establishes.

1.2 What is and is not in the literature↩︎

The classification problem for low-dimensional associative algebras has since accumulated a substantial literature. The three-dimensional case over \(\mathbb{R}\) was first addressed by Scheffers [4] in 1891. The modern treatment — complete, rigorous, and explicitly over both \(\mathbb{R}\) and \(\mathbb{C}\) — appears in Kobayashi, Shirayanagi, Takahasi, and Tsukada [5], whose six isomorphism classes agree with Theorem 2; the correspondence is made explicit in Proposition 6.

Given this prior work, the contributions of the present paper are the following.

Framework. The near-trinion framing gives the classification a narrative coherence and a connection to the long history of the problem that the earlier treatments, which approach the question as a routine enumeration, do not provide.

Proof method. The argument in [5] is coordinate-based: structure constants are varied until all isomorphism classes are enumerated. It produces the correct list but does not explain why the list has the structure it does. The present argument proceeds through the Jacobson radical stratification and Peirce decomposition (Sections 2 and 3). This approach answers a different question: not only which algebras exist, but why the radical dimension \(d\) takes values in \(\{0,1,2\}\), why the \(d=1\) case splits on commutativity, and why no near-trinion with residue field \(\mathbb{C}\) can exist. The last point merits proof here, since it drives the entire \(d=1\) analysis. If \(A/ \mathrm{rad}(A)\cong\mathbb{C}\) and \(\mathrm{rad}(A)=\mathbb{R}\eta\) for some nonzero \(\eta\in\mathrm{rad}(A)\), then \(i\cdot\eta=c\eta\) for some \(c\in\mathbb{R}\), so applying \(i\) again gives \(i^2\cdot\eta = c^2\eta\). Since \(i^2=-1 \in \mathbb{C}\), we obtain \((-1)\eta = c^2\eta\), hence \(c^2=-1\), impossible over \(\mathbb{R}\). The \(d=1\) stratum therefore splits into exactly two near-trinions, determined entirely by whether the radical generator occupies a diagonal or off-diagonal Peirce corner subalgebra.

1.3 Road map↩︎

The proof of Theorem 2 is organized as follows. Section 2 establishes the three tools used throughout: the Jacobson radical (three equivalent characterizations and their computations for each near-trinion), the Wedderburn–Artin theorem, and Nakayama’s Lemma. Section 3 carries out the classification in three strata determined by \(d = \dim_\mathbb{R}\mathrm{rad}(A) \in \{0,1,2\}\). In the \(d = 0\) stratum, Wedderburn–Artin identifies \(\mathbb{R}^3\) and \(\mathbb{R}\times\mathbb{C}\) as the only semisimple candidates. In the \(d = 1\) stratum, the Peirce decomposition shows that the radical generator lies either in a diagonal corner (producing the commutative near-trinion NT3) or an off-diagonal corner (producing the non-commutative near-trinion NT6); the subcase \(A/\mathrm{rad}(A) \cong \mathbb{C}\) is shown to be impossible. In the \(d = 2\) stratum, whether \(\mathrm{rad}(A)^2\) vanishes distinguishes NT4 from NT5. Section 4 establishes irredundancy: the six algebras are distinguished by \(d\), nilpotency of \(\mathrm{rad}(A)^2\), commutativity, and primitive idempotent count; specifically, \(d\) alone separates the three strata; commutativity separates NT3 from NT6; and Loewy length separates NT4 from NT5. One direction for subsequent work is described in Section 5.

1.4 Conventions↩︎

Throughout, algebra means a finite-dimensional, associative, unital \(\mathbb{R}\)-algebra, unless stated otherwise. The Jacobson radical of an algebra \(A\) is denoted \(\mathrm{rad}(A)\), and we write \(\ell\ell(A)\) for the Loewy length, the least \(n\) such that \(\mathrm{rad}(A)^n = 0\). The phrase near-trinion refers specifically to any three-dimensional associative unital \(\mathbb{R}\)-algebra. For a quotient ring \(R/I\), an overbar denotes the coset of an element: \(\bar{r} = r + I \in R/I\); thus \(\bar\varepsilon \in \mathbb{R}[\varepsilon]/(\varepsilon^2)\), \(\bar x \in \mathbb{R}[x]/(x^3)\), and so on. Peirce corners are written \(e_i A e_k = \{e_i a e_k : a \in A\}\) for idempotents \(e_i, e_k\).

The proof occupies Sections 2 and 3; the invariants that establish irredundancy are computed in Section 4.

2 Preliminaries↩︎

We record here the ring-theoretic tools on which the classification rests. The reader familiar with the Wedderburn–Artin theorem and the Jacobson radical may proceed directly to Section 3.

2.1 The Jacobson radical: three characterizations↩︎

For finite-dimensional algebras over a field, the Jacobson radical admits three descriptions that are, in general, distinct objects; the content of the following proposition is that they coincide.

Recall that the Loewy length \(\ell\ell(A)\) is defined in Section 1.4 (Conventions); for \(\mathbb{R}[x]/(x^3)\) it takes the value \(3\), the maximum possible in dimension three.

2.2 The Wedderburn–Artin theorem↩︎

2.3 Comparison with Kobayashi–Shirayanagi–Takahasi–Tsukada↩︎

\(^\dagger\,U^3_{2-}\) (subscript with minus sign) denotes the algebra existing only over \(\RR\), arising from the real division algebra \(\CC\); it has no analogue over \(\CC\) and accounts for the count of six (over \(\RR\)) versus five (over \(\CC\)) unital three-dimensional algebras. The label \(U^3_2\) (no minus sign, row NT5) is a distinct algebra.

2.4 Maximal ideals of near-trinions↩︎

2.5 The Nakayama Lemma for near-trinions↩︎

For a near-trinion \(A\), set \(d = \dim_\mathbb{R}\mathrm{rad}(A)\). Since \(\dim_\mathbb{R}A = 3\) we have \(d \in \{0,1,2,3\}\) a priori, but \(d = 3\) is ruled out: if \(\mathrm{rad}(A)\) were three-dimensional it would equal \(A\), making \(A\) nilpotent. But \(1 \in A\) is not nilpotent, a contradiction, so \(1 \notin \mathrm{rad}(A)\), forcing \(d \leq 2\). The semisimple quotient \(A/\mathrm{rad}(A)\) has dimension \(3 - d\), and the three strata are:

• \(d = 0\): \(A\) semisimple; Wedderburn–Artin applies directly, yielding (NT1) and (NT2).

• \(d = 1\): \(A/\mathrm{rad}(A)\) has dimension two. Two-dimensional semisimple real algebras are either \(\mathbb{R}^2\) or \(\mathbb{C}\). The subcase \(A/\mathrm{rad}(A) \cong \mathbb{C}\) is impossible in dimension three (shown in Section 3.2). The subcase \(A/\mathrm{rad}(A) \cong \mathbb{R}^2\) yields two distinct near-trinions, (NT3) and (NT6), distinguished by whether the radical generator lies in a diagonal or off-diagonal Peirce corner.

• \(d = 2\): \(A/\mathrm{rad}(A) \cong \mathbb{R}\); \(A\) is local with two-dimensional nil radical; whether \(\mathrm{rad}(A)^2 = 0\) distinguishes (NT4) from (NT5).

3 The Classification↩︎

The three strata \(d \in \{0,1,2\}\) partition the problem exhaustively (Section 2), and within each stratum the analysis below is complete: the \(d=0\) stratum is settled by a single dimension count; the \(d=1\) stratum has two Peirce-corner sub-cases (diagonal and off-diagonal) plus one impossibility argument; the \(d=2\) stratum splits on whether \(\mathrm{rad}(A)^2=0\).

3.1 The semisimple case (\(d=0\))↩︎

When \(\mathrm{rad}(A) = 0\), Theorem 5 applies directly. The dimension constraint reads: \(\dim_\mathbb{R}M_n(D) = n^2 \dim_\mathbb{R}D\), so the only ways to write \(3\) as such a sum are \(3 = 1+1+1\) and \(3 = 1+2\). The first gives \(\mathbb{R}\times \mathbb{R}\times \mathbb{R}\); the second gives \(\mathbb{R}\times \mathbb{C}\). No other combination works: \(\dim \mathbb{H}= 4\) and \(\dim M_2(\mathbb{R}) = 4\) already exceed three.

3.2 One-dimensional radical (\(d=1\))↩︎

Here \(A/\mathrm{rad}(A)\) is a two-dimensional semisimple \(\mathbb{R}\)-algebra, hence isomorphic to either \(\mathbb{R}^2\) or \(\mathbb{C}\). The subcase \(A/\mathrm{rad}(A) \cong \mathbb{R}^2\) produces two non-isomorphic near-trinions, determined by whether the radical generator occupies a diagonal or off-diagonal Peirce corner; the subcase \(A/\mathrm{rad}(A) \cong \mathbb{C}\) is impossible over \(\mathbb{R}\), as the argument below establishes.

Subcase: \(A/\mathrm{rad}(A) \cong \mathbb{R}^2\). Choose a basis \(\{e_1, e_2, j\}\) for \(A\) where \(e_1, e_2\) are orthogonal idempotents summing to \(1\) and \(j\) generates \(\mathrm{rad}(A)\). Nilpotency forces \(j^2 = 0\). (Here \(j\) is a name for the radical generator in the \(d=1\) analysis; it is unrelated to the basis elements \(\varepsilon_1, \varepsilon_2\) used for NT4 in Section 3.) The structure is determined by how \(j\) sits in the Peirce decomposition (the decomposition of \(A\) into corner subalgebras \(e_i A e_k\) determined by the idempotents \(e_1, e_2\); see Section 2 and Conventions) with respect to \(e_1\) and \(e_2\): since \(j\) is a non-unit, \(j\) belongs to some corner \(e_i A e_k\), meaning \(e_i \cdot j = j\) and \(j \cdot e_k = j\). There are two essentially distinct cases.

Case (a): diagonal placement (\(i = k\)). Without loss of generality \(j \in e_1 A e_1\), so \(e_1 j = j e_1 = j\) and \(e_2 j = j e_2 = 0\). Here \(j\) commutes with both idempotents, and the algebra decomposes as \(A \cong (e_1 A e_1) \times (e_2 A e_2)\). Since \(e_1 A e_1 \cong \mathbb{R}[\varepsilon]/(\varepsilon^2)\) and \(e_2 A e_2 \cong \mathbb{R}\), we obtain \(A \cong \mathbb{R}\times \mathbb{R}[\varepsilon]/(\varepsilon^2)\). This algebra is commutative. \[\boxed{A \;\cong\; \mathbb{R}\times \mathbb{R}[\varepsilon]/(\varepsilon^2). }\] Hence, Case (a) yields exactly (NT3).

Case (b): off-diagonal placement (\(i \neq k\)). Without loss of generality \(j \in e_1 A e_2\), so \(e_1 j = j\), \(j e_2 = j\), while \(e_2 j = 0\) and \(j e_1 = 0\). The multiplication table on \(\{e_1, e_2, j\}\) is: \[e_1^2 = e_1,\quad e_2^2 = e_2,\quad e_1 e_2 = 0,\quad e_1 j = j,\quad j e_2 = j,\quad e_2 j = j e_1 = j^2 = 0.\] Under the identification \(e_1 = e_{11}\), \(e_2 = e_{22}\), \(j = e_{12}\), this is exactly \(T_2(\mathbb{R})\). The algebra is non-commutative: \(e_1 j = j \neq 0 = j e_1\). \[\boxed{A \;\cong\; T_2(\mathbb{R}). }\] Hence, Case (b) yields exactly (NT6).

The two cases produce non-isomorphic algebras: (NT3) is commutative and (NT6) is not. Thus the subcase \(A/\mathrm{rad}(A) \cong \mathbb{R}^2\) yields exactly the two near-trinions (NT3) and (NT6).

Subcase: \(A/\mathrm{rad}(A) \cong \mathbb{C}\): impossible in dimension three.

If \(A/\mathrm{rad}(A) \cong \mathbb{C}\) and \(\mathrm{rad}(A) = \mathbb{R}\eta\) (one-dimensional), then \(\mathrm{rad}(A)\) is a two-sided ideal, so \(i \cdot \eta \in \mathrm{rad}(A) = \mathbb{R}\eta\), giving \(i \cdot \eta = c\eta\) for some \(c \in \mathbb{R}\). But then \(i^2 \cdot \eta = c^2\eta\), and since \(i^2 = -1\) in \(\mathbb{C}\) we need \(c^2 = -1\), which has no real solution. Hence, the subcase \(A/\mathrm{rad}(A) \cong \mathbb{C}\) is impossible, and the \(d = 1\) case yields exactly (NT3) and (NT6).

3.3 Two-dimensional radical (\(d=2\))↩︎

Here \(A/\mathrm{rad}(A) \cong \mathbb{R}\) and \(A\) is local with two-dimensional nil radical \(J = \mathrm{rad}(A)\). The product \(J^2\) is either zero or one-dimensional.

Subcase: \(J^2 = 0\). Taking a basis \(\{1, \varepsilon_1, \varepsilon_2\}\) with \(\varepsilon_a \varepsilon_b = 0\) for all \(a, b\): \[\boxed{A \;\cong\; \mathbb{R}[\varepsilon_1, \varepsilon_2]\big/(\varepsilon_1^2,\, \varepsilon_2^2,\, \varepsilon_1 \varepsilon_2). }\] The isomorphism class is independent of the choice of basis for \(J\): any two bases are related by an invertible linear map \(\phi\colon J\to J\). Since \(J^2=0\), every product of two elements of \(J\) is zero, so any invertible linear change of basis \(J\to J\) fixes all products and hence extends (with \(1\mapsto 1\)) to an \(\mathbb{R}\)-algebra automorphism of \(A\).

Hence, the subcase \(J^2=0\) yields exactly (NT4).

Subcase: \(J^2 \neq 0\), so \(\dim_\mathbb{R}J^2 = 1\). Let \(x \in J \setminus J^2\). Since \(x \notin J^2 = \mathbb{R}x^2\) and \(\dim_\mathbb{R}J = 2\), the elements \(x\) and \(x^2\) are linearly independent, so \(\{x, x^2\}\) is a basis for \(J\). Since \(J^3 \subseteq J \cdot J^2 = \mathbb{R}x \cdot \mathbb{R}x^2 = \mathbb{R}x^3\), and nilpotency of \(J\) forces \(J^N = 0\) for some \(N\), we conclude \(J^3 = 0\) and hence \(x^3 = 0\). Thus \[\boxed{A \;\cong\; \mathbb{R}[x]/(x^3). }\] Any other generator \(x' = ax + bx^2\) with \(a \neq 0\) gives the same isomorphism class. Indeed, the assignment \(x \mapsto x'\) extends uniquely to an \(\mathbb{R}\)-algebra automorphism of \(\mathbb{R}[x]/(x^3)\), and \((x')^3 = 0\) since \(x^3 = 0\). Hence, the subcase \(J^2\neq 0\) yields exactly (NT5).

3.4 Multiplication tables↩︎

We now record, for each near-trinion, a concrete multiplication table on a chosen \(\mathbb{R}\)-basis \(\{v_1, v_2, v_3\}\). Off-table products are determined by \(\mathbb{R}\)-bilinearity and the unit; entries \(v_i \cdot v_j\) are read from row \(i\), column \(j\).

(NT1) \(A = \mathbb{R}^3\), basis \(\{e_1, e_2, e_3\}\): \(e_i e_j = \delta_{ij} e_i\).

(NT2) \(A = \mathbb{R}\times \mathbb{C}\), basis \(\{f,\, \mathbf{1}_\mathbb{C},\, \mathbf{i}\}\) where \(f = (1,0)\), \(\mathbf{1}_\mathbb{C}= (0,1)\), \(\mathbf{i} = (0,i)\):

(NT3) \(A = \mathbb{R}\times \mathbb{R}[\varepsilon]/(\varepsilon^2)\), basis \(\{e_1, e_2, j\}\) with \(e_1 = (1,0)\), \(e_2 = (0,1)\), \(j = (0,\bar\varepsilon)\):

Here \(j = e_2 j = j e_2\) and \(j^2 = 0\); \(j\) lies in the diagonal corner \(e_2 A e_2\), confirming Case (a) above.

(NT4) \(A = \mathbb{R}[\varepsilon_1,\varepsilon_2]/(\varepsilon_1^2, \varepsilon_2^2, \varepsilon_1 \varepsilon_2)\), basis \(\{1, \varepsilon_1, \varepsilon_2\}\):

(NT5) \(A = \mathbb{R}[x]/(x^3)\), basis \(\{1, \bar x, \bar x^2\}\):

The radical satisfies \(\mathrm{rad}(A)^2 = \mathbb{R}\bar x^2 \neq 0\), distinguishing (NT5) from (NT4).

(NT6) \(A = T_2(\mathbb{R})\), basis \(\{e_{11}, e_{22}, e_{12}\}\) with \(e_{11} + e_{22} = 1_A\):

Note \(e_{11} \cdot e_{12} = e_{12}\) while \(e_{12} \cdot e_{11} = 0\): this is the unique non-commutative near-trinion. The radical generator \(e_{12}\) lies in the off-diagonal corner \(e_{11}A e_{22}\), confirming Case (b) above. The (NT6) table is the unique one not symmetric under transposition.

4 Invariants and Irredundancy↩︎

The six near-trinions are distinguished by four invariants: \(d = \dim_\mathbb{R}\mathrm{rad}(A)\); whether \(\mathrm{rad}(A)^2 = 0\); commutativity; and the number of primitive orthogonal idempotents. Table 3 below records these together with center dimension and Loewy length. For the five commutative near-trinions, \(Z(A) = A\) so \(\dim Z(A) = 3\); for \(T_2(\mathbb{R})\) the center consists only of scalar matrices, giving \(\dim Z(A) = 1\).

5 Directions for Further Work↩︎

The classification is specific to finite-dimensional associative unital \(\mathbb{R}\)-algebras of dimension three, and three features of this setting are essential. Frobenius’ theorem restricts the Wedderburn–Artin summands over \(\mathbb{R}\) to matrix rings over \(\mathbb{R}\), \(\mathbb{C}\), or \(\mathbb{H}\); over \(\mathbb{C}\) or a finite field the decomposition differs. The effect is already visible in dimension three: \(\mathbb{R}\times\mathbb{C}\) has no analogue over \(\mathbb{C}\) since \(\mathbb{C}\) does not split as a non-trivial product of \(\mathbb{C}\)-algebras, and accordingly [5] lists five unital algebras over \(\mathbb{C}\) versus six over \(\mathbb{R}\). Finite-dimensionality underlies both the Loewy-series bound \(d\leq 2\) and Nakayama’s Lemma; neither holds in infinite dimensions. The Jacobson radical and Wedderburn–Artin theorem presuppose associativity; Lie, Jordan, and octonion algebras require different methods [1], [8], [9].

The next problem is the classification over finite fields. Over \(\mathbb{F}_q\), Wedderburn’s little theorem replaces Frobenius’ theorem in the semisimple stratum, and the cubic extension \(\mathbb{F}_{q^3}\) enters as a semisimple algebra with no real counterpart. The sign obstruction used in the \(d=1\) stratum — the equation \(c^2 = -1\) having no real solution — is genuinely specific to \(\mathbb{R}\): over \(\mathbb{F}_q\) with \(q \equiv 1 \pmod{4}\) the equation is solvable, and the obstruction, if there is one, must be found elsewhere.

Acknowledgements↩︎

The author thanks the organizers of the AMS Special Session on Quaternions at the January 2024 Joint Mathematics Meetings in San Francisco, at which the trinion presentation [3] was given; their questions, particularly regarding what occupies dimension three in the absence of a division algebra, are what set this work in motion.

References↩︎