Preprint
Article

This version is not peer-reviewed.

Matrix Theory over the Algebra of Ordered Pairs of Normalized Real Numbers

Submitted:

19 October 2026

Posted:

20 October 2026

You are already at the latest version

Abstract
We study similarity of matrices over the algebra of ordered pairs of normalized real numbers (OPNs) using its native shifted matrix cubes. For three fixed OPNs centers, the three independently given similarity classes determine the original class in every dimension if and only if all pairwise center differences are units. Three is the least uniform number of such observations. The substantive recovery step is a real matrix theorem: any three distinct real shifts suffice, whereas any two admit a six-dimensional real counterexample. The proof combines a finite-weight argument on a cubic rotation lattice, polynomial interpolation, and noncritical Jordan blocks; critical output blocks need not be used. The OPNs criterion follows through the two real characters of the coefficient algebra. As a complementary invariant, we show that labelled nullities of shifted powers through order ⌈n/2⌉, at all complex scalar parameters, determine similarity with an optimal uniform cutoff. Classical Smith–Fitting invariants and explicit higher evaluation modules place both descriptions in a self-contained algebraic framework.
Keywords: 
;  ;  ;  ;  ;  

MSC:  15A21; 15A18; 13C05; 13D07

1. Introduction

Which operations on an OPNs matrix retain its full similarity class when the results are known only up to independent changes of basis? We consider the native cubes ( A − α I ) 3 at fixed scalar centers α . Recovering a matrix from such observations is different from inverting a cubic in a supplied basis: the three observed classes carry no common eigenvalue labelling or conjugator.
Ordered pairs of normalized real numbers (OPNs) provide a two-component setting for this question. Their arithmetic is defined on ( 0 , 1 ) 2 through a continuous strictly increasing bijection φ : R → ( 0 , 1 ) satisfying φ ( u ) + φ ( − u ) = 1 [1]. Write O φ for this algebra. Its two real characters are
p ( μ , ν ) = − φ − 1 ( μ ) − φ − 1 ( ν ) , q ( μ , ν ) = φ − 1 ( μ ) − φ − 1 ( ν ) .
They identify O φ with R × R . For A ∈ M n ( O φ ) , applying the characters entrywise gives real matrices A p , A q . Here similarity is taken with respect to the full group G L n ( O φ ) ; the channel changes of basis can be independent. Write 1 ~ for the multiplicative identity. An element is a unit if it has a multiplicative inverse, equivalently if both its characters are nonzero. All matrix powers below use the original OPNs addition and multiplication, not componentwise powers of entries.
Theorem 1
(Native cubic recovery criterion). Fix α 0 , α 1 , α 2 ∈ O φ . The following conditions are equivalent:
(i)
For every n ≥ 1 and all A , B ∈ M n ( O φ ) , similarity of ( A − α i I ) 3 and ( B − α i I ) 3 for each i = 0 , 1 , 2 implies similarity of A , B .
(ii)
Each difference α i − α j with i ≠ j is a unit of O φ .
Each observed pair may use its own conjugating matrix. The least number of fixed OPNs centers guaranteeing recovery in all dimensions is three; the centers 0 1 ~ , 1 ~ , 2 1 ~ suffice.
The substantive step is Theorem 2: three distinct shifts recover the full Jordan data of an ordinary matrix, including arbitrary nonsemisimple blocks. Its proof first eliminates the missing mass at the critical fibers by summing finite weights over a cubic rotation lattice. Three-center polynomial interpolation then recovers every block multiplicity. An explicit six-dimensional real example proves failure for any prescribed pair of real centers. The OPNs statement follows by the product decomposition, and its unit-difference condition expresses exactly when both channels receive three distinct centers. It imposes no additional structure on the original algebra or its similarity group.
The ingredients and the scope should be distinguished. Jordan functional calculus is classical [2], as is interpolation by shifted powers [3]. Finite signed weights with vanishing fiber sums are related to the lightning-bolt and cycle conditions for sums of function algebras [4,5]. Those frameworks do not by themselves check the three particular cubic maps or the missing critical fibers used here. General separation of simultaneous similarity orbits by ranks of matrix pencils [6, Theorem 1.1] uses mixed observations; the present input consists instead of three separately given classes. Our contribution is the exact recovery criterion and its two-center obstruction for this observation family, not a new Jordan calculus, interpolation method, or classification of product algebras.
Section 4 supplies a complementary description using the nullities of all scalar shifts through order ⌈ n / 2 ⌉ , with no supplied characteristic or minimal polynomial. This is an order cutoff for a family indexed by all complex parameters, not a bound on the number of observations. Section 5 interprets these data through explicit tensor and Tor fibers. The classical global invariants and their complete proofs are retained to make the algebraic framework self-contained.
The algebraic and spectral ingredients have classical origins. The coefficient algebra is the split-complex, or double-number, algebra [7,8]. Related matrix decompositions over the double numbers, with particular attention to their involutions, are studied by Gutin [9, Section 2]. Smith reduction and Fitting ideals describe polynomial presentations [10,11]. Here the two polynomial factors are principal ideal domains, so the full Fitting sequence is complete. This is a special case of classical module theory, not a new classification principle for general rings.
The nullity increments are the Weyr characteristic [12,13,14]. Their relation to Jordan block sizes is the conjugacy of partitions. Rank-sequence admissibility is also discussed by Garcia, Sherman, and Weiss [15, Section 2]. Dumas, Pernet, and Saunders [16, Sections 2.1–2.2] use low-order nullities and a total-degree constraint to recover characteristic-polynomial multiplicities in a black-box setting with a supplied minimal polynomial. The shared ingredients are the recovery of small blocks and the dimension restriction on large blocks. Theorem 3 gives an exact uniform cutoff for the full similarity class under the stated, different input conditions; it is not a complexity bound for a black-box algorithm. The reconstruction uses exact ranks and should also be distinguished from regularized numerical Jordan forms [17].
All rings and modules are unital, and all coefficient rings are commutative. Left–right equivalence means G = U F V with U , V invertible of the appropriate sizes; similarity means B = P − 1 A P . Empty direct sums and unit invariant factors are allowed.
For a linear map T with ordered domain and codomain bases B , C , write M ( T , B , C ) for its matrix, abbreviated M ( T ) when the bases are understood.

2. The OPNs Algebra and Its Global Invariant

Fix a continuous strictly increasing bijection φ : R → ( 0 , 1 ) such that
φ ( u ) + φ ( − u ) = 1 ( u ∈ R ) .
In particular, φ ( 0 ) = 1 / 2 and φ − 1 ( 1 − v ) = − φ − 1 ( v ) . For α = ( μ α , ν α ) , set x α = φ − 1 ( μ α ) and y α = φ − 1 ( ν α ) . The OPNs operations of Zhou [1, Definition 3] are, in these coordinates,
c α = φ ( c x α ) , φ ( c y α ) , α + β = φ ( x α + x β ) , φ ( y α + y β ) , α β = φ ( − x α y β − y α x β ) , φ ( − x α x β − y α y β ) .
We denote ( 0 , 1 ) 2 with these operations by O φ .
Proposition 1
(Character decomposition). The maps
p ( α ) = − ( x α + y α ) , q ( α ) = x α − y α
define a unital real-algebra isomorphism
S φ : O φ ⟶ R × R , α ⟼ p ( α ) , q ( α ) ,
with inverse
S φ − 1 ( p , q ) = φ q − p 2 , φ − p + q 2 .
These are the only two unital real-algebra homomorphisms O φ → R . The only real-algebra automorphisms of O φ are the identity and
τ ( μ , ν ) = ( 1 − μ , ν ) , S φ τ ( α ) = q ( α ) , p ( α ) .
For another normalization ψ, every real-algebra isomorphism O φ → O ψ is either G φ , ψ = S ψ − 1 S φ or τ ψ G φ , ψ .
Proof. 
Addition and real scalar multiplication in (2.2) are respected by both linear forms. To check multiplication, write ( x , y ) = ( x α , y α ) and ( x ′ , y ′ ) = ( x β , y β ) . The coordinates of α β are X = − x y ′ − y x ′ and Y = − x x ′ − y y ′ , and hence
− ( X + Y ) = ( x + y ) ( x ′ + y ′ ) , X − Y = ( x − y ) ( x ′ − y ′ ) .
These are exactly p ( α ) p ( β ) and q ( α ) q ( β ) . Solving p = − ( x + y ) , q = x − y gives x = ( q − p ) / 2 , y = − ( p + q ) / 2 , proving (2.5) and bijectivity. Thus S φ transports the algebra axioms from R × R ; its unit is S φ − 1 ( 1 , 1 ) .
Write e p = S φ − 1 ( 1 , 0 ) and e q = S φ − 1 ( 0 , 1 ) . These satisfy e p 2 = e p , e q 2 = e q , e p e q = 0 ~ , and e p + e q = 1 ~ , and every α equals p ( α ) e p = q ( α ) e q . If h : O α → R is a unital real-algebra homomorphism, then h ( e p ) and h ( e q ) are idempotent real numbers with sum 1. Their values must therefore be (1,0) or (0,1) Real linearity now gives
h ( α ) = p ( α ) h ( e p ) + q ( α ) h ( e q ) ,
so h = p or h = q .
The only idempotents of R × R are ( 0 , 0 ) , ( 1 , 0 ) , ( 0 , 1 ) , ( 1 , 1 ) . An automorphism fixes zero and the unit, so it either fixes e p , e q or interchanges them. Since these two elements form a real basis, the automorphism is determined by this choice. Interchanging p , q sends ( x , y ) to ( − x , y ) ; the normalization identity then gives φ ( − x ) = 1 − φ ( x ) , proving (2.6). Finally, G φ , ψ ( φ ( x ) , φ ( y ) ) = ( ψ ( x ) , ψ ( y ) ) . For any isomorphism H : O φ → O ψ , the composite H G φ , ψ − 1 is an automorphism of O ψ , so the two possibilities just established give the asserted list of isomorphisms. □
The zero, unit, and two primitive central idempotents are
0 ˜ = ( 1 / 2 , 1 / 2 ) , 1 ~ = φ ( 0 ) , φ ( − 1 ) , e p = φ ( − 1 / 2 ) , φ ( − 1 / 2 ) , e q = φ ( 1 / 2 ) , φ ( − 1 / 2 ) .
Every element has the unique expression p e p + q e q , and e p 2 = e p , e q 2 = e q , e p e q = 0 ~ , e p + e q = 1 ~ . Real scalars are embedded as c 1 ~ . The map α ↦ − y α − x α j identifies O φ with R [ j ] / ( j 2 − 1 ) .
We retain the labels p , q throughout. Thus an OPNs matrix is uniquely written F = F p e p + F q e q . Applying τ entrywise interchanges all channel data at once. Classification without an ordering of the two distinguished factors means taking the quotient by this single global interchange, not permitting an independent exchange at each invariant factor or spectral value.
For a commutative unital real algebra B, put O B = O φ ⊗ R B = B e p ⊕ B e q . For every O B -module M, orthogonality of the two idempotents gives
M = e p M ⊕ e q M , G L r ( O B ) ≅ G L r ( B ) × G L r ( B ) .
Indeed, m = e p m + e q m for every m ∈ M . If m belongs to both summands, then e p m = m because m ∈ e p M , whereas e p m = 0 because m ∈ e q M ; hence the sum is direct. Every module map L satisfies L ( e χ m ) = e χ L ( m ) , so it preserves the summands. Kernels, images, and cokernels therefore split into their corresponding channel modules.
For matrices, multiplication is channelwise:
( U p e p + U q e q ) ( V p e p + V q e q ) = U p V p e p + U q V q e q .
Thus U is invertible precisely when both U p , U q are invertible, and its inverse is U p − 1 e p + U q − 1 e q . Projecting an equivalence or similarity gives the corresponding relations over B. Conversely, any two channel changes of basis can be recombined by this formula into a change of basis over O B , proving the converse implications as well.
Put A = O φ [ t ] . The characteristic module of A ∈ M n ( O φ ) is
M A = coker A ( t I − A ) .
In a finite presentation C a → F C b → M → 0 , the j-th Fitting ideal is generated by the ( b − j ) -minors of F. The conventions and presentation independence are given in Appendix A.
The determinant of each channel pencil t I − A χ is monic of degree n, so its normalized Smith chain h χ , 1 ∣ ⋯ ∣ h χ , n has no zero entries. The unit entries are retained.
Proposition 2
(Classical characteristic-module classification). The underlying O φ -module of M A is naturally O φ n , and multiplication by t corresponds to A. For A , B ∈ M n ( O φ ) , similarity is equivalent to M A ≅ M B as A -modules. Equivalently, the two Smith chains of t I − A and t I − B agree, or the complete Fitting sequences of their characteristic modules agree. Writing χ χ ( t ) = det ( t I − A χ ) and letting m χ ( t ) denote the monic polynomial of least degree annihilating A χ , one has
χ χ ( t ) = ∏ i = 1 n h χ , i ( t ) , m χ ( t ) = h χ , n ( t ) ,
and
Ann A ( M A ) = ( m p ) e p ⊕ ( m q ) e q .
Conversely, every pair of monic nonzero divisibility chains h χ , 1 ∣ ⋯ ∣ h χ , n satisfying
∑ i = 1 n deg h χ , i = n χ ∈ { p , q }
is realized by exactly one similarity class in M n ( O φ ) .
The polynomial evaluation map and the proof of the classification are given in Appendix B.

3. Recovery from Native Shifted Cubes

The observations in Theorem 1 use only subtraction, matrix multiplication, and similarity over the original OPNs algebra. The proof separates a scalar ambiguity from the Jordan data. A cubic can identify different eigenvalues and can shorten a block at its critical point. We first remove the ambiguity using only nonzero fibers, so that the proof never needs to invert a critical block.
We begin with finite weights and the missing critical fibers.
A finite weight on C is a function w : C → C with finite support. Sums over infinite sets below mean finite sums over this support. The nonzero fibers of z ↦ ( z − c ) 3 are the three-point orbits of rotation about c through a primitive cubic root of unity.
Lemma 1
(Elimination of critical mass). Let a , b ∈ C be distinct. Suppose a finite weight w satisfies
∑ ( z − c ) 3 = y w ( z ) = 0 ( c ∈ { a , b } , y ∈ C ∖ { 0 } ) .
Then ∑ z w ( z ) = 0 .
Proof. 
The substitution z = a + ( b − a ) u is a bijection of C and transports w to the finite weight u ↦ w ( a + ( b − a ) u ) . The two cubic outputs become ( b − a ) 3 u 3 and ( b − a ) 3 ( u − 1 ) 3 , respectively. Since b − a ≠ 0 , this preserves the hypotheses on nonzero fibers and the total weight. It therefore suffices to treat the centers 0 , 1 .
Choose a primitive cubic root of unity ω and set
L = ( 1 − ω ) Z [ ω ] , R 0 ( z ) = ω z , R 1 ( z ) = ω z + ( 1 − ω ) .
Multiplication by ω preserves L, because both ω and its inverse ω 2 belong to Z [ ω ] . Translation by 1 − ω also preserves L. Thus R 0 and R 1 restrict to permutations of L. Moreover, 0 ∈ L but 1 ∉ L : if integers r , s satisfied ( 1 − ω ) ( r + s ω ) = 1 , then ω 2 = − 1 − ω would give
( r + s ) + ( 2 s − r ) ω = 1 .
The real linear independence of 1 , ω implies r + s = 1 and 2 s − r = 0 , hence 3 s = 1 , a contradiction.
For either center c, the rotation z ↦ c + ω ( z − c ) has order three and has c as its only fixed point. Every other orbit consists of the three distinct roots of ( z − c ) 3 = y for a nonzero y. Decompose L first into R 0 -orbits. Its only singleton orbit is { 0 } , and every other orbit has weight zero by the hypothesis at c = 0 . Decompose the same set into R 1 -orbits. There is no singleton orbit because 1 ∉ L , so every orbit has weight zero by the hypothesis at c = 1 . Only finitely many orbits meet the support of w, which justifies both regroupings and yields
∑ z ∈ L w ( z ) = w ( 0 ) , ∑ z ∈ L w ( z ) = 0 .
Thus w ( 0 ) = 0 . Finally, partition all of C ∖ { 0 } into the nonzero fibers of z 3 . Each fiber sum vanishes, and finite support again permits summation, so ∑ z w ( z ) = w ( 0 ) = 0 . □
Lemma 2
(Three-center separation of finite weights). Let c 0 , c 1 , c 2 be distinct complex numbers. If a finite weight w has zero sum on every nonzero fiber of each ( z − c i ) 3 , then w = 0 .
Proof. 
Put M j = ∑ z w ( z ) z j for j ≥ 0 . Applying Lemma 1 to any two of the centers gives M 0 = 0 . Fix k ≥ 1 and a center c i . Multiply the zero sum on the fiber with output y ≠ 0 by y k , and sum over the finitely many such fibers meeting the support. On each fiber, y k = ( z − c i ) 3 k . The omitted point z = c i contributes zero because k ≥ 1 . Consequently,
∑ z w ( z ) ( z − c i ) 3 k = 0 ( i = 0 , 1 , 2 ) .
Assume inductively that M j = 0 for 0 ≤ j ≤ 3 k − 3 ; the case k = 1 uses precisely the already established identity M 0 = 0 . The binomial expansion of (3.4) is
∑ j = 0 3 k 3 k j ( − c i ) 3 k − j M j = 0 .
All terms except the final three vanish by the induction hypothesis. Hence
M 3 k − 3 k c i M 3 k − 1 + 3 k 2 c i 2 M 3 k − 2 = 0 ( i = 0 , 1 , 2 ) .
The left side, regarded as a polynomial in c i , has degree at most two and vanishes at the three distinct centers. It is therefore the zero polynomial. Since 3 k and 3 k 2 are nonzero, its coefficients give M 3 k − 2 = M 3 k − 1 = M 3 k = 0 . This extends the induction hypothesis through degree 3 k , so every moment vanishes.
If the support is empty, there is nothing to prove. Otherwise list its distinct points as z 1 , … , z s . The first s moment equations are a homogeneous linear system for w ( z 1 ) , … , w ( z s ) , whose coefficient in row j + 1 , column r, is z r j , with 0 ≤ j < s . Its Vandermonde determinant is ∏ r < t ( z t − z r ) ≠ 0 . The only solution is zero, and therefore w = 0 . □
The interpolation in this lemma is classical; it is a filtered version of the shifted-power basis property [3, Proposition 1.1]. The separate lattice argument is needed because the nonzero fibers alone do not initially supply the zeroth moment.
We next pass from finite weights to Jordan blocks and then construct the two-center obstruction.
We recall the required local matrix-function fact with its proof. It is a special case of standard Jordan functional calculus [2].
Write J j ( λ ) = λ I j + N j , where N j has ones on the superdiagonal and zeros elsewhere.
Lemma 3
(Noncritical cubic blocks). If λ ≠ c , the cube ( J j ( λ ) − c I ) 3 is similar over C to J j ( ( λ − c ) 3 ) .
Proof. 
Write J j ( λ ) = λ I + N , where N = J j ( 0 ) . Then
( ( λ − c ) I + N ) 3 − ( λ − c ) 3 I = N Q ( N ) , Q ( N ) = 3 ( λ − c ) 2 I + 3 ( λ − c ) N + N 2 .
Set a = 3 ( λ − c ) 2 ≠ 0 . We can write Q ( N ) = a ( I + H ) , where H is a polynomial in N with zero constant term. Since N j = 0 , also H j = 0 , and
Q ( N ) − 1 = a − 1 ∑ h = 0 j − 1 ( − H ) h .
In particular, Q ( N ) is invertible and commutes with N.
Put K = N Q ( N ) . Commutativity gives K r = Q ( N ) r N r for every r ≥ 1 . As Q ( N ) r is invertible, K r v = 0 holds exactly when N r v = 0 . Thus ker K r = ker N r for every r. The matrix K is nilpotent, with nilpotency index j: K j = 0 , while K j − 1 ≠ 0 when j > 1 ; the case j = 1 gives the single zero block directly. The largest nilpotent Jordan block has size equal to the nilpotency index. A block of size j already occupies the whole j-dimensional space, so K is similar to N. Adding ( λ − c ) 3 I proves the assertion. □
Theorem 2
(Three independent cubic observations). Let A , B ∈ M n ( C ) and let c 0 , c 1 , c 2 ∈ C be distinct. If
( A − c i I ) 3 ∼ ( B − c i I ) 3 ( i = 0 , 1 , 2 ) ,
then A ∼ B . Each observation may use a different conjugating matrix. For real matrices and real centers, the conclusion holds over R .
Proof. 
For each 1 ≤ j ≤ n , let w j ( λ ) be the number of size-j Jordan blocks of A at λ minus the corresponding number for B. Its support is contained in the union of their finite spectra, so w j is a finite weight.
Fix an observation i and a value y ≠ 0 . Evaluating a polynomial on a Jordan direct sum evaluates it on each block. Every input eigenvalue with ( λ − c i ) 3 = y satisfies λ ≠ c i , so Lemma 3 preserves its block sizes. Blocks coming from different input eigenvalues may have the same output eigenvalue y; they remain a direct sum of Jordan blocks, and their multiplicities add. No critical input block can contribute at y ≠ 0 . Therefore the difference between the numbers of size-j blocks in the two outputs at y is exactly
∑ ( λ − c i ) 3 = y w j ( λ ) .
Similarity of the outputs makes this difference zero. The same equation holds when y is absent from both output spectra, because then there are no contributing input blocks. We obtain
∑ ( λ − c i ) 3 = y w j ( λ ) = 0 ( i = 0 , 1 , 2 , y ≠ 0 ) .
Lemma 2 now gives w j = 0 for every 1 ≤ j ≤ n ; larger blocks cannot occur. Thus all Jordan block multiplicities of A , B agree, which proves complex similarity. This argument uses only the nonzero output eigenvalues; it does not require the lengths of critical input blocks to survive cubing.
For real A , B and real centers, the observed real similarities are also complex similarities, so the first part gives a complex invertible matrix P with A P = P B . Write P = X + i Y , where X , Y are real. Taking real and imaginary parts yields A X = X B and A Y = Y B . Consequently X + t Y intertwines A , B for every real t. The real polynomial f ( t ) = det ( X + t Y ) satisfies f ( i ) = det P ≠ 0 , so it is not the zero polynomial and has only finitely many real roots. Choose a real t outside those roots. Then X + t Y is invertible and proves similarity over R . □
Three refers to the number of independently supplied similarity classes, each containing its full Jordan information. It is neither a bound on the number of scalar invariants nor an algorithmic complexity claim. In particular, replacing an observed similarity class by its characteristic polynomial is not justified by this proof.
Proposition 3
(Failure of two real centers). For every pair of distinct real numbers u , v there exist nonsimilar six-by-six real matrices whose shifted cubes at both centers are similar. For u = 0 , v = 1 the matrices can have rational entries.
Proof. 
Set
C ( s , p ) = 0 − p 1 s , X = diag ( 2 , 2 , − 1 , − 1 , C ( 1 , 7 ) ) , Y = diag ( C ( − 2 , 4 ) , C ( 4 , 7 ) , C ( 1 , 1 ) ) .
Since C ( s , p ) has characteristic polynomial t 2 − s t + p , the characteristic polynomials of these block diagonal matrices are
p X ( t ) = ( t − 2 ) 2 ( t + 1 ) 2 ( t 2 − t + 7 ) , p Y ( t ) = ( t 2 + 2 t + 4 ) ( t 2 − 4 t + 7 ) ( t 2 − t + 1 ) , p X ( t ) − p Y ( t ) = 54 t ( 1 − t ) ≠ 0 .
Thus X ≁ Y . Each displayed quadratic has nonzero discriminant, so each quadratic block has two distinct complex eigenvalues. Together with the scalar blocks, this shows that both X and Y are diagonalizable over C .
To compute their cubic spectra exactly, the Cayley–Hamilton identity gives C 2 = s C − p I , and multiplication by C yields
C 3 = ( s 2 − p ) C − s p I , det ( t I − C 3 ) = t 2 − ( s 3 − 3 s p ) t + p 3 .
For the second formula, the trace of the first expression is s ( s 2 − p ) − 2 s p = s 3 − 3 s p , while det ( C 3 ) = ( det C ) 3 = p 3 . The same calculation applies to any two-by-two matrix with trace s and determinant p, so it also applies to C − I , whose parameters are s − 2 , p − s + 1 . For example, the first formula gives C ( − 2 , 4 ) 3 = 8 I 2 and C ( 1 , 1 ) 3 = − I 2 ; the remaining block of Y 3 has characteristic polynomial t 2 + 20 t + 343 . Applying the formulas to all displayed blocks and to their shifts gives
det ( t I − X 3 ) = det ( t I − Y 3 ) = ( t − 8 ) 2 ( t + 1 ) 2 ( t 2 + 20 t + 343 ) , det ( t I − ( X − I ) 3 ) = det ( t I − ( Y − I ) 3 ) = ( t − 1 ) 2 ( t + 8 ) 2 ( t 2 − 20 t + 343 ) .
If T = S D S − 1 with D diagonal, then f ( T ) = S f ( D ) S − 1 for every polynomial f. Thus all four observed matrices are diagonalizable, even when the polynomial identifies different input eigenvalues. For diagonalizable complex matrices, equality of characteristic polynomials means equality of the eigenvalue multisets, hence similarity by a permutation of diagonal entries. This proves similarity of both observed pairs over C . All their entries are real, so the real-intertwiner argument in Theorem 2 gives similarity over R .
For the prescribed distinct real centers u , v , put
X u , v = u I + ( v − u ) X , Y u , v = u I + ( v − u ) Y .
At the center u, the two cubes are ( v − u ) 3 X 3 and ( v − u ) 3 Y 3 . At the center v, they are ( v − u ) 3 ( X − I ) 3 and ( v − u ) 3 ( Y − I ) 3 . Each corresponding pair is similar by the conjugator already obtained. If X u , v and Y u , v were similar, subtracting u I and dividing by v − u ≠ 0 would make X , Y similar, a contradiction. The displayed matrices for u = 0 , v = 1 have rational entries; the affine construction has real entries for arbitrary real u , v . □
This proves that three is the least number of fixed real centers which works uniformly in all matrix dimensions. It does not claim that three centers are necessary in each dimension, or that six is the least dimension for a real two-center counterexample. If all eigenvalues are real, two distinct real centers already suffice: the scalar cube is injective on R , and each eigenvalue is noncritical at at least one center.
We now transfer the real recovery theorem and its obstruction to the native OPNs operations.
For α ∈ O φ , define
C α ( A ) = ( A − α I ) 3 .
All entries in this expression use exactly (2.2); the power denotes a product of three matrix factors, not entrywise cubing. Since α I is central,
C α ( A ) = A 3 − 3 α A 2 + 3 α 2 A − α 3 I , ( C α ( A ) ) χ = ( A χ − χ ( α ) I ) 3 .
An element δ is a unit precisely when both p ( δ ) , q ( δ ) are nonzero. In inverse-normalization coordinates this condition is
x δ 2 − y δ 2 ≠ 0 ; ( x α − x β ) 2 ≠ ( y α − y β ) 2 for the difference δ = α − β .
Thus the hypothesis is directly testable from the original pairs. Being distinct as pairs alone is insufficient: their difference can be a zero divisor.
Proof of Theorem 1.
Suppose first that all three pairwise differences are units. For either χ ∈ { p , q } , multiplicativity and additivity of the character imply χ ( α i − α j ) = χ ( α i ) − χ ( α j ) ≠ 0 . Thus the three real channel centers are distinct. Applying χ entrywise to a conjugating matrix for each native observation gives a real invertible conjugator for the corresponding channel observations in (3.14). Theorem 2 therefore supplies real invertible matrices P p , P q such that A χ P χ = P χ B χ . The matrix P = P p e p + P q e q satisfies A P = P B ; orthogonality of the idempotents gives its inverse explicitly as P p − 1 e p + P q − 1 e q . Hence A , B are similar over O φ .
Conversely, suppose one pairwise difference is not a unit. In some channel χ , two of the three character values coincide, so the observed centers in that channel form a set of at most two real numbers. If there are exactly two values, Proposition 3 provides real matrices U , V of dimension six that are not similar, although their cubes shifted at either value are similar. If there is just one value c, take U = c I 2 + J 2 ( 0 ) and V = c I 2 . Their shifted cubes both vanish since J 2 ( 0 ) 2 = 0 ; they are not similar because U − c I 2 has rank one and V − c I 2 has rank zero.
In either case, let n be the chosen dimension and form OPNs matrices A , B with A χ = U , B χ = V ; in the other channel set both matrices equal to 0 n . For each observation i, choose an invertible real conjugator in channel χ , which exists by the construction. In the other channel the observed matrices are identical, so take the identity as conjugator. Combining these two matrices with the idempotents gives an invertible OPNs conjugator for that observation. Thus all three observed pairs are similar. However, similarity of A , B would imply similarity of their χ -components U , V , which is impossible. Uniform recovery therefore fails.
Finally, the centers 0 ˜ , 1 ˜ , 2 1 ˜ have pairwise differences equal to nonzero real multiples of 1 ˜ , hence units, and so three observations suffice. For any prescribed one or two OPNs centers, choose a channel. Its set of character values has one or two members, and the same constructions just used give nonsimilar matrices with similar observations at all prescribed centers. Thus fewer than three centers cannot work in all dimensions, and the minimum uniform number is three. □
The channel decomposition is a proof tool and a complete description of the coefficient algebra. No shared change of basis between the two channels is imposed. The intrinsic contribution is the exact criterion for these native observations; its passage from the real matrix theorem is a direct product argument, not a new classification principle for product algebras. Transport between normalizations preserves the characters, units, shifted cubes, and this criterion.

4. Truncated Spectral Data

For A ∈ M n ( O φ ) , set
d χ , k A ( z ) = dim C ker ( A χ − z I ) k , χ ∈ { p , q } , z ∈ C , k ≥ 1 ,
and d χ , 0 A ( z ) = 0 . The channel labels are fixed.
Theorem 3
(Sharp reconstruction). Let n ≥ 1 and r n = ⌈ n / 2 ⌉ . For A , B ∈ M n ( O φ ) , the following are equivalent:
(i)
A and B are similar over O φ ;
(ii)
d χ , k A ( z ) = d χ , k B ( z ) for both channels, every z ∈ C , and 1 ≤ k ≤ r n .
For every n, r n is the smallest nonnegative uniform cutoff for which this equivalence holds. No characteristic or minimal polynomial is required as additional data.
The same order bound and optimality hold for ordinary real matrices, with a single nullity profile. The OPNs statement retains the two labelled profiles separately. The bound concerns the order of the shifted powers, not a finite set of prescribed evaluation points: the support of the nullity functions is part of the data. It does not assert that the kernels have stabilized by order r n .
The proof separates the recovered blocks from the undetermined tails. Differences of consecutive nullities give all blocks shorter than r n . If at least two blocks remain, they use the entire dimension and both lengths are forced to equal r n . If only one remains, its length is the unused dimension. This also gives necessary and sufficient conditions for a truncated profile to occur (Corollary 1). Sharpness follows from two-block matrices that move one dimension between distinct eigenvalues while preserving every lower-order nullity.
The following classical formula relates the global invariant to the local nullities [12,13,14].
Lemma 4
(Local lengths and nullities). Let ν χ , i ( z ) = ord z h χ , i , where ord z h is the multiplicity of z as a root of h. Then
d χ , k ( z ) = ∑ i = 1 n min k , ν χ , i ( z ) .
Consequently
b χ , k ( z ) : = d χ , k ( z ) − d χ , k − 1 ( z ) = # i : ν χ , i ( z ) ≥ k , c χ , k ( z ) : = b χ , k ( z ) − b χ , k + 1 ( z ) = # i : ν χ , i ( z ) = k .
The full family of nullities determines the two real Smith chains.
Proof. 
Proposition 2, followed by complex scalar extension, identifies the operator in channel χ with multiplication by t on ⨁ i C [ t ] / ( h χ , i ) . For one summand write h = ( t − z ) ν g , where g ( z ) ≠ 0 . The two factors are coprime, so Bezout’s identity gives
C [ t ] / ( h ) ≅ C [ t ] / ( ( t − z ) ν ) ⊕ C [ t ] / ( g ) .
When ν = 0 , the first summand is zero. Since t − z is coprime to g, multiplication by ( t − z ) k is invertible on the second summand and contributes no kernel. In the first summand put u = t − z . If k < ν , the kernel of multiplication by u k has basis u ν − k , … , u ν − 1 ; if k ≥ ν , the entire summand is killed. Its dimension is therefore min { k , ν } . Summing over i proves (4.2).
For every nonnegative integer ν ,
min { k , ν } − min { k − 1 , ν } = 1 , ν ≥ k , 0 , ν < k .
Taking the first and then the second difference proves (4.3). Let T z , a be multiplication by t on C [ t ] / ( ( t − z ) a ) . In the reversed power basis ( ( t − z ) a − 1 , … , 1 ) , its matrix is M ( T z , a ) = J a ( z ) . Thus b χ , k ( z ) and c χ , k ( z ) count blocks of size at least k and exactly k, respectively.
At each z, the recovered positive lengths are padded with zeros to length n and sorted in nondecreasing order. Divisibility of the Smith chain forces this to be the order ν χ , 1 ( z ) ≤ ⋯ ≤ ν χ , n ( z ) : a factor dividing the next factor cannot have a larger multiplicity at any root. Hence
h χ , i ( t ) = ∏ z ∈ C ( t − z ) ν χ , i ( z ) ( 1 ≤ i ≤ n ) .
There are only finitely many nontrivial factors. The real Smith identity remains a Smith identity over C [ t ] , so uniqueness identifies these monic products with the original real polynomials. In particular, the local lists are aligned by their nondecreasing orders; no additional choice of a matching between different roots is needed. □
Proof of Theorem 3.
Fix a channel and suppress χ . Put r = ⌈ n / 2 ⌉ . Since r ≥ 1 , the data include d 1 ; its support is exactly the set of eigenvalues, because A − z I has a nonzero kernel exactly at those points. Thus the eigenvalue labels are already known, even though no characteristic polynomial is supplied. The given nullities determine b k ( z ) for k ≤ r , and therefore c k ( z ) for k < r . All blocks shorter than r are known. Set
S = ∑ z ∈ C ∑ k = 1 r − 1 k c k ( z ) , T = ∑ z ∈ C b r ( z ) , E = n − S − r T .
The sums are finite; when r = 1 , the inner sum defining S is empty and equals zero. Here S is the total size of the recovered blocks, while b r ( z ) counts the remaining blocks at the known eigenvalue z. Thus T is their total number. Writing each remaining length as r + e , with e ≥ 0 , shows that E ≥ 0 is the sum of these excesses.
If T = 0 , no length remains unknown and S = n . If T = 1 , its label is the unique z with b r ( z ) = 1 , and its length is forced to be r + E = n − S . If T ≥ 2 , then
n = S + r T + E ≥ r T ≥ 2 r ≥ n .
All inequalities are equalities. Since r ≥ 1 , equality r T = 2 r gives T = 2 ; the other equalities give n = 2 r and S = E = 0 . Each excess is nonnegative and their sum is zero, so both remaining lengths are r. In particular, this last case cannot occur when n is odd. The complete local lengths, including their labels, and hence the Smith chain are now determined. Applying this argument to both channels and using Proposition 2 proves (ii)⇒(i). The converse follows because an invertible intertwiner carries each shifted-power kernel isomorphically onto the corresponding kernel.
For sharpness, let n ≥ 3 and s = ⌈ n / 2 ⌉ − 1 . Then s ≥ 1 and n ≥ 2 s + 1 . With J a ( z ) denoting the Jordan block of size a, set
X = J s + 1 ( 0 ) ⊕ J n − s − 1 ( 1 ) , Y = J s ( 0 ) ⊕ J n − s ( 1 ) .
All four block sizes are at least s, since n − s − 1 ≥ s . For 1 ≤ k ≤ s , each block at its own eigenvalue contributes nullity k, and the block at the other eigenvalue is invertible. Hence both nullity functions equal k at z = 0 , 1 and vanish elsewhere. The algebraic multiplicities at zero are s + 1 and s, so X and Y are not similar. Thus cutoff r − 1 = s , and therefore every smaller cutoff, fails. The OPNs matrices X e p + I n e q and Y e p + I n e q preserve this counterexample with both labels fixed. For n = 1 , 2 , a smaller nonnegative cutoff would be zero and supply no data; the zero and identity matrices are not similar. □
The counterexamples vary the characteristic polynomial. Accordingly, the optimality assertion concerns exactly the input in Theorem 3, not the different problem in which characteristic polynomials are supplied.
Corollary 1
(Realization of truncated profiles). Let r = ⌈ n / 2 ⌉ . A family of functions d χ , k : C → Z , for χ ∈ { p , q } and 0 ≤ k ≤ r , is the truncated nullity profile of an n × n OPNs matrix if and only if, in each channel:
(i)
d χ , 0 = 0 , and the increments b χ , k = d χ , k − d χ , k − 1 satisfy b χ , 1 ( z ) ≥ ⋯ ≥ b χ , r ( z ) ≥ 0 ;
(ii)
b χ , 1 has finite support, and d χ , k ( z ¯ ) = d χ , k ( z ) ;
(iii)
L χ : = ∑ z ∈ C d χ , r ( z ) ≤ n ,with L χ = n whenever b χ , r is identically zero.
Every such family is realized by exactly one real OPNs similarity class.
Proof. 
Necessity follows from (4.2): the increments count blocks of length at least k, so they are nonnegative and nonincreasing. The spectrum is finite. Since the channel matrices are real, complex conjugation maps the kernel at z bijectively onto the kernel at z ¯ , giving condition (ii). Moreover, L χ is the total size of all blocks after capping each length at r, so L χ ≤ n . If no block has length at least r, capping changes nothing and L χ = n .
For sufficiency, fix a channel and suppress χ . At each z, put c k ( z ) = b k ( z ) − b k + 1 ( z ) for k < r . Form c k ( z ) blocks of size k < r and b r ( z ) tentative blocks of size r. These multiplicities are nonnegative integers and vanish outside the finite support of b 1 . At each z, telescoping gives
d r ( z ) = ∑ j = 1 r b j ( z ) = ∑ k = 1 r − 1 k b k ( z ) − b k + 1 ( z ) + r b r ( z ) .
Consequently L = S + r T , with S , T as in (4.5). Put E = n − L ≥ 0 . If T = 0 , condition (iii) gives E = 0 . If T = 1 , the label of the only tentative block must be real: a nonreal label would have a distinct conjugate with the same positive multiplicity. Increase this block’s length by E. If T ≥ 2 , the equality n = S + r T + E and the inequalities in (4.6) force T = 2 , S = 0 , and E = 0 . Thus no extension is needed in that case. The completed lengths have total size n and retain conjugate symmetry in every case.
Sort and pad the lengths at each root to obtain n entries, and define the polynomials by (4.4). Conjugate roots have identical ordered lists, so each polynomial is real. The lists are nondecreasing, so each polynomial divides the next one. Their total degree is the sum of all completed lengths, namely n. Performing the construction in both channels and applying Proposition 2 therefore realizes these chains by an OPNs matrix. For 1 ≤ j ≤ r , the number of constructed blocks at z whose lengths are at least j is
∑ k = j r − 1 c k ( z ) + b r ( z ) = b j ( z ) .
Increasing a tentative length beyond r does not change this count. The constructed increments therefore equal the supplied increments; summing from d 0 = 0 recovers every supplied nullity. Theorem 3 gives uniqueness. □
Example 1
(Polynomial summaries and first-order data). Let J r ( z ) denote the r × r matrix with z on the diagonal, ones on the superdiagonal, and zeros elsewhere. Set
A p = J 3 ( 0 ) ⊕ J 3 ( 0 ) ⊕ J 1 ( 0 ) , B p = J 3 ( 0 ) ⊕ J 2 ( 0 ) ⊕ J 2 ( 0 ) , A = A p e p + I 7 e q , B = B p e p + I 7 e q .
Both matrices have channel characteristic polynomials ( t 7 , ( t − 1 ) 7 ) and channel minimal polynomials ( t 3 , t − 1 ) . For every z ∈ C , their order-one nullities have the same labelled pair,
3 1 z = 0 , 7 1 z = 1 .
Nevertheless, dim C ker A p 2 = 5 whereas dim C ker B p 2 = 6 , so they are not similar. First-order nullities and both polynomial summaries can therefore agree while the complete invariant differs.
The p-channel nullity sequences through order three are ( 3 , 5 , 7 ) and ( 3 , 6 , 7 ) . The corresponding Smith chains are
( 1 , 1 , 1 , 1 , t , t 3 , t 3 ) , ( 1 , 1 , 1 , 1 , t 2 , t 2 , t 3 ) .
Their characteristic products and last factors agree, but their intermediate factors do not.

5. Higher Evaluation Modules

Let
D = O φ ⊗ R C = C e p ⊕ C e q , R = D [ t ] = A C , M A C = R ⊗ A M A .
The same notation A denotes the extended matrix. For λ = z p e p + z q e q ∈ D , set
J λ , k = R / ( t − λ ) k ( k ≥ 1 ) .
The module J λ , 1 is D with t acting as λ . In what follows, a fiber type means its isomorphism class as a D-module, with the two idempotents labelled. It is determined by the two complex dimensions, not by an embedding of an eigenspace into D n .
Evaluation at λ maps R onto D, with one complex evaluation in each channel. In particular, it is not evaluation into a single residue field. A scalar parameter λ = z 1 ~ evaluates both channels at the same z, while keeping their dimensions separate.
The scalar extension is algebraic; no complex extension of the normalization function is used. Tensor products give the cokernel fibers, and the first Tor group gives the kernel fibers [18].
Proposition 4
(Natural higher fibers). Let A ∈ M n ( O φ ) . For every λ ∈ D and k ≥ 1 , there are natural R-module isomorphisms
Tor 1 R ( M A C , J λ , k ) ≅ ker D ( A − λ I ) k , M A C ⊗ R J λ , k ≅ coker D ( A − λ I ) k ,
where t acts as A on the right-hand sides. Moreover, Tor j R ( M A C , J λ , k ) = 0 for j ≥ 2 .
Proof. 
The highest nonzero coefficient of a polynomial vector f is also the leading coefficient of ( t I − A ) f . Thus t I − A is injective, and
0 ⟶ R n → t I − A R n ⟶ M A C ⟶ 0
is a free resolution, even though R has zero divisors. The cokernel is M A C because scalar extension preserves the defining cokernel of M A . By definition, the Tor groups are the homology groups obtained by tensoring this free resolution with the second module. Set u = t − λ and N = A − λ I . Then J λ , k = D [ u ] / ( u k ) , and tensoring (5.4) gives
0 ⟶ Tor 1 R ( M A C , J λ , k ) ⟶ J λ , k n → u I − N J λ , k n ⟶ M A C ⊗ R J λ , k ⟶ 0 .
Each vector in J λ , k n has a unique expansion v ( u ) = ∑ i = 0 k − 1 u i v i . Comparing its coefficients in ( u I − N ) v ( u ) = 0 gives
N v 0 = 0 , v i − 1 = N v i ( 1 ≤ i < k ) .
There is no additional coefficient equation at degree k, because u k = 0 . If w = v k − 1 , the recursion forces v i = N k − 1 − i w , and the remaining equation N v 0 = 0 is exactly N k w = 0 . Consequently the first isomorphism in (5.3) is explicitly given, in the reverse direction, by
ker D ( N k ) ⟶ ker u I − N : J λ , k n → J λ , k n , w ⟼ ∑ i = 0 k − 1 u i N k − 1 − i w .
Its inverse takes the coefficient of u k − 1 ; the same argument also covers k = 1 .
For the cokernel, define E ( v ) = ∑ i = 0 k − 1 N i v i ( mod N k D n ) . For v ( u ) = ∑ i u i v i , direct cancellation gives E ( ( u I − N ) v ) = − N k v k − 1 = 0 modulo N k D n . Thus E induces a surjective map from the cokernel to D n / N k D n , since every target class has a constant representative. In the cokernel, the relation u x = N x for a constant vector x implies inductively that [ u i x ] = [ N i x ] . Hence every class [ v ( u ) ] equals the class of the constant vector ∑ i N i v i . Every constant vector N k w is zero there, because
( u I − N ) ∑ i = 0 k − 1 u i N k − 1 − i w = − N k w in J λ , k n .
It follows that a class killed by E is already zero, which proves injectivity. We have obtained the second isomorphism,
v ( u ) ⟼ ∑ i = 0 k − 1 N i v i ∈ D n / N k D n .
For (5.7), multiplying the displayed image by u gives the image of N w : the only possible additional constant term is N k w , which is zero. For (5.8), evaluation sends multiplication by u to multiplication by N modulo N k D n . Both maps are also D-linear, and therefore R-linear, with t = u + λ acting as N + λ I = A . The kernel and image of N k are stable under A, so this action on both right-hand sides is well-defined. Finally, an intertwiner H satisfying H A = B H also satisfies H ( A − λ I ) i = ( B − λ I ) i H for every i. It commutes with both displayed constructions, proving naturality in the matrix. There are no nonzero terms of the free resolution in degrees j ≥ 2 , so all higher Tor groups vanish. □
As a D-module, either fiber is determined by its two complex dimensions. In each channel, rank–nullity for the square matrix ( A χ − z I ) k gives
dim C coker ( A χ − z I ) k = n − rank ( A χ − z I ) k = dim C ker ( A χ − z I ) k .
Thus at λ = z 1 ~ , both fibers have dimension pair ( d p , k ( z ) , d q , k ( z ) ) . Consequently Theorem 3 and Corollary 1 classify both the matrices and their realizable labelled fiber types through order ⌈ n / 2 ⌉ . Here only the D-module types are retained, although the natural identifications above preserve the full R-module structure.
Normalization transport preserves the two characters and therefore the characteristic module, its Smith–Fitting invariants, and all higher profiles. If the coefficient algebra is considered up to real-algebra isomorphism, the sole additional equivalence is one global interchange of the two channels. Independent interchanges at different roots or different invariant factors are not allowed.

6. Conclusions and Open Questions

The native OPNs operations support two exact reconstruction descriptions. Three shifted cubic similarity classes suffice precisely when their centers have pairwise unit differences, with no common basis imposed on the observations or on the channels. Two centers fail uniformly, even for semisimple real channel matrices. Separately, the full family of labelled nullity profiles needs powers only through ⌈ n / 2 ⌉ , and the cutoff is optimal in every dimension. The character decomposition, Smith–Fitting invariants, and higher evaluation fibers connect these descriptions using classical algebra.
The cubic result concerns exact classes. It gives neither a stable numerical reconstruction procedure nor a reduction of the information needed to specify each observed class. The smallest dimension of a real two-center counterexample remains to be determined here. For higher native powers, one may ask which finite center configurations eliminate critical-fiber ambiguities and give a sharp uniform recovery criterion. These questions are not resolved by the cubic proof and are not assumptions needed for any result above.

Funding

This research received no specific external funding.

Conflicts of Interest

The author declares no conflicts of interest.

Use of AI-assisted tools.

OpenAI’s ChatGPT and Codex were used to assist with the exploration and checking of mathematical arguments, manuscript drafting and restructuring, literature searches, and LaTeX editing. The author takes responsibility for the mathematical statements, proofs, references, and final text.

Appendix A Polynomial Classification

The reductions in this appendix are classical. They are included to specify the normalization and recover all presentation data. For the broader theory of elementary divisor rings, see [19,20,21]. Background on modules and determinantal ideals is given in [22,23,24]. First let P = K [ t ] for a field K.
For a matrix F ∈ M m , n ( C ) over a commutative ring C, let I k ( F ) be the ideal generated by its k × k minors, with I k ( F ) = C for k ≤ 0 and I k ( F ) = 0 for k > min ( m , n ) . Expansion of determinants gives I k + 1 ( F ) ⊆ I k ( F ) .
Lemma A5
(Finite presentations). Every submodule of P m is free of rank at most m. In particular, every finitely generated P-module is finitely presented.
Proof. 
Every nonzero ideal of P is principal: choose a nonzero member of least degree and divide any other member by it; a nonzero remainder would contradict minimality. Let N ⊆ P m and argue by induction on m, starting with m = 0 . The image of N under the first-coordinate projection is an ideal. If it is zero, use the induction hypothesis in the remaining m − 1 coordinates. Otherwise write it as ( a ) , with a ≠ 0 , and choose v ∈ N with first coordinate a. Every w ∈ N has first coordinate c a , so w − c v belongs to N 0 = N ∩ ( 0 ⊕ P m − 1 ) . Thus N = P v ⊕ N 0 : the intersection is zero because P is a domain. The induction hypothesis supplies a finite basis of N 0 , proving the assertion. Finally, the kernel of a surjection P m → M has a finite basis, which supplies a finite presentation of M. □
Lemma A6
(Polynomial Smith form). For F ∈ M m , n ( P ) , there exist invertible matrices U , V over P such that
U F V = diag ( d 1 , … , d ρ , 0 , … , 0 ) , d 1 ∣ ⋯ ∣ d ρ ,
where the nonzero entries are monic. They are uniquely determined by
I k ( F ) = ( d 1 … d k ) ( 1 ≤ k ≤ ρ ) , I k ( F ) = 0 ( k > ρ ) .
Proof. 
If F = 0 , there is nothing to prove. The degrees of nonzero entries in all matrices left–right equivalent to F form a nonempty subset of the nonnegative integers, so they have a least element. Choose an entry attaining it, move it to position ( 1 , 1 ) , and multiply its row by a nonzero constant to make it monic; call it d. If an entry f below d has division f = q d + r , subtracting q times the first row produces r in that position. Minimality forces r = 0 . Thus d divides every entry of its column, which can be cleared below d. The same division argument with column operations clears the first row to the right. These latter operations do not change the zeros below d, so the resulting matrix is diag ( d , H ) .
If an entry h of H were not divisible by d, add its row to the first row. The first entry stays d, while the entry above h becomes h. Write h = q d + r with 0 ≠ r and deg r < deg d . Subtracting q times the first column from that column puts r in the first row, contradicting minimality. Therefore every entry of H belongs to ( d ) . Apply induction on min ( m , n ) to H, extending its invertible row and column operations by the identity on the first coordinate. All entries of the transformed block remain in ( d ) , because each is a P-linear combination of its previous entries. Consequently d divides the first nonzero Smith entry of H. The induction hypothesis supplies the divisibility among the remaining entries, proving existence of the full chain.
For uniqueness, Cauchy–Binet expresses each k-minor of U F V as a P-linear combination of k-minors of F. Hence I k ( U F V ) ⊆ I k ( F ) . Applying the same argument to F = U − 1 ( U F V ) V − 1 gives the reverse inclusion. For a diagonal matrix, a nonzero k-minor is a product d i 1 ⋯ d i k with 1 ≤ i 1 < ⋯ < i k ≤ ρ . Since i j ≥ j and the entries form a divisibility chain, d j ∣ d i j for every j. Thus d 1 … d k divides each such minor and is itself a minor. There are no nonzero minors of size greater than ρ . This proves (A2).
The largest index of a nonzero minor ideal is therefore ρ . The monic generator of I k ( F ) divided by the monic generator of I k − 1 ( F ) , with I 0 ( F ) = P , is exactly d k . Both the number and the values of the normalized nonzero entries are consequently determined by F. □
If C n → F C m → M → 0 is a finite presentation, define
Fitt j C ( M ) = I m − j ( F ) ( j ≥ 0 ) .
Lemma A7
(Presentation and scalar extension). The ideals in (A3) are independent of the finite presentation. For a homomorphism u : C → C ′ ,
Fitt j C ′ ( C ′ ⊗ C M ) = C ′ u Fitt j C ( M ) .
For C = O B , they split as
Fitt j O B ( M ) = Fitt j B ( e p M ) e p ⊕ Fitt j B ( e q M ) e q .
Proof. 
Fix a surjection π : C m → M , or equivalently a finite generating list for M. Suppose the columns of F and G are two finite generating lists for ker π . Each column of either matrix is a linear combination of columns of the other, so F = G X and G = F Y for suitable matrices X , Y . Cauchy–Binet gives I k ( F ) ⊆ I k ( G ) from the first identity and the reverse inclusion from the second. Thus changing the chosen generators of the relation module changes none of the minor ideals.
Next adjoin a generator b ∈ M , and choose c ∈ C m with π ( c ) = b . The enlarged surjection is
π ′ : C m + 1 ⟶ M , π ′ ( x , z ) = π ( x ) + z b .
An element ( x , z ) lies in its kernel precisely when x + z c ∈ ker π . Writing x + z c = F y shows
( x , z ) = ( F y , 0 ) + z ( − c , 1 ) .
Therefore a relation matrix for the enlarged list is
F − c 0 1 .
Multiplication on the left by the invertible matrix I m c 0 1 changes it into diag ( F , 1 ) , without changing its minor ideals. For positive minor sizes, expansion along the last row or column gives
I m + 1 − j diag ( F , 1 ) = I m + 1 − j ( F ) + I m − j ( F ) = I m − j ( F ) .
The final equality uses I k + 1 ( F ) ⊆ I k ( F ) ; if a size is nonpositive, the same equality follows directly from the convention I k ( F ) = C . Thus adjoining a generator with its defining relation preserves every Fitting ideal.
For two arbitrary finite presentations, concatenate their generating lists in one fixed order. Starting from either presentation, adjoin the missing generators using the preceding construction, and reorder the generators if necessary by an invertible permutation matrix. Both enlarged relation matrices now generate the kernel of the same surjection onto M, so the first paragraph compares them. This proves presentation independence.
Tensor product is right exact, so applying C ′ ⊗ C − to a finite presentation gives
( C ′ ) n → u ( F ) ( C ′ ) m ⟶ C ′ ⊗ C M ⟶ 0 .
Every minor of u ( F ) is the image under u of the corresponding minor of F, because a determinant is a polynomial with integer coefficients in the matrix entries. Their generated ideals therefore satisfy (A4), with no flatness assumption.
Finally, for any ideal J of B e p ⊕ B e q , multiplication by the two idempotents gives J = e p J ⊕ e q J . Projecting a presentation onto either channel presents e χ M over B, and projection of each minor of F gives the same minor of F χ . Hence the two components of its Fitting ideal are precisely the two ideals on the right of (A5). □
These are the standard Fitting properties [11,24]. We now specialize to a ring for which the whole Fitting sequence is a complete invariant. Let K ∈ { R , C } , put P = K [ t ] , and write
A K = ( O φ ⊗ R K ) [ t ] = P e p ⊕ P e q .
A nonzero polynomial is normalized to be monic. A Smith chain includes its unit entries and places any zero entries last. Divisibility has its usual ring meaning, so every polynomial divides zero.
Theorem A4
(Smith–Fitting classification). Let s = min ( m , n ) . Every F ∈ M m , n ( A K ) is left–right equivalent to a rectangular diagonal matrix with entries
d i = d p , i e p + d q , i e q ( 1 ≤ i ≤ s ) , d χ , 1 ∣ … ∣ d χ , s ,
where each d χ , i is monic or zero. The two normalized chains are unique.
Every finitely generated A K -module admits a decomposition
M ≅ ⨁ χ ∈ { p , q } P r χ ⊕ ⨁ i = 1 ℓ χ P / ( a χ , i ) e χ , a χ , 1 ∣ ⋯ ∣ a χ , ℓ χ ,
where the a χ , i are monic and nonconstant. The integers and polynomials in this expression are unique; the isomorphism and the individual summands inside M need not be canonical.
For M χ = e χ M ,
Fitt j P ( M χ ) = 0 , j < r χ , ( a χ , 1 … a χ , r χ + ℓ χ − j ) , r χ ≤ j < r χ + ℓ χ , P , j ≥ r χ + ℓ χ .
Consequently { Fitt j A K ( M ) } j ≥ 0 determines the complete labelled isomorphism class of M.
Proof. 
Lemma A6 proves Smith existence and uniqueness over P directly by polynomial division and minors. Apply it to F p , F q , and recombine their invertible row and column transformations. The idempotent decomposition in Section 2 also proves uniqueness of the two chains.
The kernel of a surjection A K m → M splits into two submodules of P m . Lemma A5 makes both finitely generated, so M is finitely presented over A K . The same lemma gives a finite presentation of each M χ . Apply Smith reduction to these presentations. Unit diagonal entries contribute zero modules, nonunit nonzero entries contribute P / ( a χ , i ) , and unused target coordinates contribute P r χ . This proves the decomposition in (A8).
Present its χ -component by the ( r χ + ℓ χ ) × ℓ χ matrix with diagonal entries a χ , 1 , … , a χ , ℓ χ and r χ zero rows. A minor of size greater than ℓ χ is zero. For 1 ≤ k ≤ ℓ χ , the first k diagonal entries have product dividing every other nonzero k-minor, and that product is itself a minor. This proves (A9).
To reconstruct the data, first set
r χ = min j ≥ 0 : Fitt j P ( M χ ) ≠ 0 , ℓ χ = min L ≥ 0 : Fitt r χ + L P ( M χ ) = P .
Let Δ χ , j be the monic generator of Fitt j P ( M χ ) for r χ ≤ j ≤ r χ + ℓ χ , so the final generator is 1. Then
a χ , i = Δ χ , r χ + ℓ χ − i Δ χ , r χ + ℓ χ − i + 1 ( 1 ≤ i ≤ ℓ χ ) .
Lemma A7 recovers these component ideals from the OPNs Fitting ideals and makes them independent of the presentation. The formulas prove uniqueness and completeness. They also cover free modules and the zero module, with empty products interpreted as 1. □
Corollary A2
(Fixed-size presentations). For F , G ∈ M m , n ( A K ) , the following are equivalent:
(i)
F and G are left–right equivalent;
(ii)
their two normalized Smith chains agree;
(iii)
coker F ≅ coker G as A K -modules;
(iv)
Fitt j A K ( coker F ) = Fitt j A K ( coker G ) for every j ≥ 0 .
Proof. 
The theorem proves all implications except the recovery of the entire matrix chain from the cokernel. In a channel, let r χ be the free rank of the cokernel and ℓ χ its number of nonunit invariant factors. The number of nonzero Smith entries is ρ χ = m − r χ . The chain therefore consists of ρ χ − ℓ χ units, the ℓ χ recovered nonunit factors, and s − ρ χ zeros. The fixed dimensions make these numbers determined. □
The fixed-size hypothesis cannot be dropped: adjoining a generator with its defining relation preserves the cokernel but changes the size of a presentation matrix. Also, ℓ χ counts nonunit invariant factors, not composition length; for example, P / ( t 2 ) has one such factor and composition length two.

Appendix B Characteristic-module Classification

We prove Proposition 2. The characteristic-module description and its relation to rational canonical form are classical [25,26].
Proof. 
Give O φ n an A -module structure by letting t act as A. For f ( t ) = ∑ j = 0 d t j v j ∈ A n , define ε A ( f ) = ∑ j = 0 d A j v j . This map is O φ -linear and satisfies ε A ( t f ) = A ε A ( f ) , so it is A -linear for the stated action. It is surjective because every constant vector is its own image. For j ≥ 1 , telescoping the right-hand side gives
t j v − A j v = ( t I − A ) ∑ i = 0 j − 1 t j − 1 − i A i v .
Summing these identities shows that f − ε A ( f ) ∈ ( t I − A ) A n . In particular, ε A ( f ) = 0 implies f ∈ ( t I − A ) A n . Conversely, ε A ( ( t I − A ) g ) = A ε A ( g ) − A ε A ( g ) = 0 . Thus its kernel is exactly ( t I − A ) A n , and the induced map identifies M A with O φ n carrying this action of t.
Under these identifications, an A -isomorphism M B → M A is an invertible O φ -linear map L that respects multiplication by t, namely L B = A L . Its matrix P = M ( L ) consequently satisfies P B = A P , or B = P − 1 A P . Conversely, such an invertible matrix intertwines every polynomial in the two operators, and therefore defines an A -isomorphism. Corollary A2, applied to the two n × n pencils, gives the equivalent Smith and Fitting descriptions, including the retained unit entries.
In each channel the top determinantal ideal is both ( det ( t I − A χ ) ) and ( ∏ i = 1 n h χ , i ) . Two nonzero generators of the same ideal of R [ t ] differ by a nonzero real constant. Since both generators here are monic, they are equal, proving the first formula in (2.10).
For any nonzero polynomial h, a polynomial annihilates R [ t ] / ( h ) exactly when it belongs to ( h ) , as one sees by applying it to the class of 1. This includes h = 1 , whose quotient is zero and whose annihilator is all of R [ t ] . The annihilator of a direct sum is the intersection of its summands’ annihilators. Hence the channel decomposition into cyclic modules gives
Ann R [ t ] ( e χ M A ) = ⋂ i = 1 n ( h χ , i ) = ( h χ , n ) ,
where the last equality uses ( h χ , 1 ) ⊇ … ⊇ ( h χ , n ) . Under ε A , multiplication by f ( t ) is the operator f ( A χ ) , so the monic generator of this annihilator is precisely the monic polynomial of least degree annihilating A χ . This proves the formula for m χ . A polynomial in A annihilates M A precisely when its two components annihilate the two channel modules, giving (2.11).
For the converse, put N χ = ⨁ i R [ t ] / ( h χ , i ) , with t acting by multiplication. Polynomial division shows that a summand of degree d > 0 has real basis 1 , t , … , t d − 1 , while a unit factor contributes the zero space. Thus (2.12) gives dim R N χ = n . Let T χ be multiplication by t on N χ , and let A χ = M ( T χ ) in the concatenated power bases. The chosen basis identifies N χ with R n carrying the action t · v = A χ v ; the evaluation argument above therefore identifies it with coker R [ t ] ( t I − A χ ) . The unique nonunit invariant factors of this module are the prescribed nonunit h χ , i . Since the pencil has size n and nonzero determinant, its full Smith chain is obtained by padding these factors with exactly the prescribed number of units. Combining the two matrices gives A = A p e p + A q e q , with both prescribed chains. The equivalences already proved show that every other realization is similar to A. □

References

  1. Zhou, L. Ordered pair of normalized real numbers. Inf. Sci. 2020, 538, 290–313. [Google Scholar] [CrossRef]
  2. Higham, N.J. Functions of Matrices. In Handbook of Linear Algebra, 2 ed.; Hogben, L., Ed.; Chapman and Hall/CRC: Boca Raton, FL, 2014; chapter 17; pp. 17.1–17.15. [Google Scholar]
  3. Koiran, P.; Pecatte, T.; García-Marco, I. On the linear independence of shifted powers. J. Complex. 2018, 45, 67–82. [Google Scholar] [CrossRef]
  4. Marshall, D.E.; O’Farrell, A.G. Approximation by a Sum of Two Algebras. The Lightning Bolt Principle. J. Funct. Anal. 1983, 52, 353–368. [Google Scholar] [CrossRef]
  5. Asgarova, A.K.; Ismailov, V.E. On the representation by sums of algebras of continuous functions. Comptes Rendus Mathématique 2017, 355, 949–955. [Google Scholar] [CrossRef]
  6. Derksen, H.; Klep, I.; Makam, V.; Volčič, J. Ranks of linear matrix pencils separate simultaneous similarity orbits. Adv. Math. 2023, 415, 108888. [Google Scholar] [CrossRef]
  7. Cockle, J. On a new imaginary in algebra. Lond. Edinb. Dublin Philos. Mag. J. Sci. 1849, 34, 37–47. [Google Scholar] [CrossRef]
  8. Sobczyk, G. The hyperbolic number plane. Coll. Math. J. 1995, 26, 268–280. [Google Scholar] [CrossRef]
  9. Gutin, R. Unitary canonical forms over Clifford algebras, and an observed unification of some real-matrix decompositions. Linear Multilinear Algebra 2024, 72, 2288–2311. [Google Scholar] [CrossRef]
  10. Smith, H.J.S. On systems of linear indeterminate equations and congruences. Philos. Trans. R. Soc. Lond. 1861, 151, 293–326. [Google Scholar] [CrossRef]
  11. Fitting, H. Die Determinantenideale eines Moduls. Jahresber. Der Dtsch. Math.-Ver. 1936, 46, 195–228. [Google Scholar]
  12. Weyr, E. Répartition des matrices en espèces et formation de toutes les espèces. Comptes Rendus De l’Académie Des. Sci. 1885, 100, 966–969. [Google Scholar]
  13. Shapiro, H. The Weyr characteristic. Am. Math. Mon. 1999, 106, 919–929. [Google Scholar] [CrossRef]
  14. O’Meara, K.C.; Clark, J.; Vinsonhaler, C.I. Advanced Topics in Linear Algebra: Weaving Matrix Problems through the Weyr Form; Oxford University Press: New York, 2011. [Google Scholar]
  15. Garcia, S.R.; Sherman, D.; Weiss, G. On the similarity of AB and BA for normal and other matrices. Linear Algebra Its Appl. 2016, 508, 14–21. [Google Scholar] [CrossRef]
  16. Dumas, J.G.; Pernet, C.; Saunders, B.D. On finding multiplicities of characteristic polynomial factors of black-box matrices. In Proceedings of the Proceedings of the 2009 International Symposium on Symbolic and Algebraic Computation, New York, 2009; ISSAC ’09, pp. 135–142. [Google Scholar] [CrossRef]
  17. Petkov, P.H.; Konstantinov, M.M. The numerical Jordan form. Linear Algebra Its Appl. 2022, 638, 1–45. [Google Scholar] [CrossRef]
  18. Weibel, C.A. An Introduction to Homological Algebra; Vol. 38, Cambridge Studies in Advanced Mathematics, Cambridge University Press: Cambridge, 1994. [CrossRef]
  19. Kaplansky, I. Elementary divisors and modules. Trans. Am. Math. Soc. 1949, 66, 464–491. [Google Scholar] [CrossRef]
  20. Gillman, L.; Henriksen, M. Some remarks about elementary divisor rings. Trans. Am. Math. Soc. 1956, 82, 362–365. [Google Scholar] [CrossRef]
  21. Larsen, M.D.; Lewis, W.J.; Shores, T.S. Elementary divisor rings and finitely presented modules. Trans. Am. Math. Soc. 1974, 187, 231–248. [Google Scholar] [CrossRef]
  22. Atiyah, M.F.; Macdonald, I.G. Introduction to Commutative Algebra; Addison-Wesley: Reading, Massachusetts, 1969. [Google Scholar]
  23. Eisenbud, D. Commutative Algebra with a View Toward Algebraic Geometry; Vol. 150, Graduate Texts in Mathematics, Springer: New York, 1995. [CrossRef]
  24. Stacks project authors, T. The Stacks project. https://stacks.math.columbia.edu, 2026. Tags 07Z8 and 07ZA. accessed. (accessed on 9 September 2026).
  25. Gantmacher, F.R. The Theory of Matrices; Hirsch, K. A., Translator; Chelsea: New York, 1959; Vol. 1. [Google Scholar]
  26. Horn, R.A.; Johnson, C.R. Matrix Analysis, second ed.; Cambridge University Press: Cambridge, 2013. [Google Scholar] [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.