T-Rank: A Combinatorial Matrix Invariant

• T-Rank is the t-term rank of a (0,1)-matrix, defined as the maximum number of ones selectable with at most one per column and at most t per row.
• Key methodologies include adaptations of the König–Egerváry theorem, Hall-type set systems, and matroid unions to derive exact formulas and structural insights.
• The theory provides extremal results like Ryser-type formulas and joint realization of maximum ranks, highlighting concave growth and robustness under interchanges.

In combinatorial matrix theory, the $t$-term rank of an $m\times n$ $(0,1)$-matrix $A$, denoted $p_t(A)$, is the largest number of $1$s in $A$ with at most one $1$ in each column and at most $t$ $1$s in each row. The case $m\times n$0 is the ordinary term rank $m\times n$1, so $m\times n$2-term rank extends the classical matching-based notion attached to a bipartite graph or incidence matrix. The subject was developed systematically in “On the t-Term Rank of a Matrix” (Brualdi et al., 2010), which generalizes König–Egerváry- and Ryser-type results, establishes exact formulas for maxima over prescribed row- and column-sum classes, and proves that within such a class there exists a single matrix simultaneously realizing the maximum $m\times n$3-term ranks for every $m\times n$4.

1. Definition and combinatorial meaning

Let $m\times n$5 be an $m\times n$6 $m\times n$7-matrix. The ordinary term rank is

$m\times n$8

Equivalently, by the König–Egerváry theorem,

$m\times n$9

If $(0,1)$0 is the bipartite graph whose biadjacency matrix is $(0,1)$1, then $(0,1)$2 is the size of a maximum matching in $(0,1)$3 (Brualdi et al., 2010).

For a positive integer $(0,1)$4, the $(0,1)$5-term rank is defined by

$(0,1)$6

where $(0,1)$7, $(0,1)$8, and $(0,1)$9. Here $A$0 means entrywise $A$1. The identity

$A$2

places the classical theory inside the $A$3-parameter family.

This definition is combinatorial rather than linear-algebraic. The paper explicitly situates $A$4 in matching, covering, set-system, and matroid frameworks, and states that it does not systematically develop relations between $A$5 and rank over a field (Brualdi et al., 2010). A common misconception is therefore to identify $A$6-term rank with ordinary matrix rank; in this literature, it is instead a constrained packing parameter for $A$7-entries.

2. Equivalent formulations and structural interpretations

A basic equivalence is obtained by stacking $A$8 copies of $A$9. If

$p_t(A)$0

then

$p_t(A)$1

Applying König–Egerváry to $p_t(A)$2 yields the min-cover formula

$p_t(A)$3

This is the direct $p_t(A)$4-analogue of the ordinary cover characterization (Brualdi et al., 2010).

There is also a Hall-type set-system formulation. If $p_t(A)$5 is regarded as the incidence matrix of a family $p_t(A)$6 of subsets of a ground set $p_t(A)$7, then

$p_t(A)$8

For $p_t(A)$9, this reduces to the standard formula for term rank.

The matroid-theoretic interpretation is equally precise. The column sets $1$0 such that $1$1 form the independent sets of a transversal matroid $1$2, whose rank is $1$3. Then

$1$4

and its rank is $1$5. A basis of $1$6 is a set of columns $1$7 of size $1$8 that can be partitioned into $1$9 so that each $A$0 has at most one $A$1 per row and exactly one $A$2 per column. This matroid-union perspective underlies the concavity and joint-realization phenomena proved later in the paper.

Finally, in bipartite-graph language, choosing $A$3s with at most $A$4 per row and at most $A$5 per column is equivalent to a $A$6-matching in which row vertices have capacity $A$7 and column vertices have capacity $A$8. This suggests that $A$9-term rank is the natural $1$0-matching extension of ordinary matching size (Brualdi et al., 2010).

3. Growth in $1$1, strength, concavity, and interchanges

For a fixed matrix $1$2, the sequence $1$3 is nondecreasing: $1$4 If $1$5 has at least one $1$6 in each column and $1$7, then $1$8 strictly increases with $1$9 until it reaches $t$0, and then remains constant at $t$1 (Brualdi et al., 2010).

This motivates the strength $t$2, defined as the smallest positive integer $t$3 such that $t$4. Equivalently, $t$5 is the smallest $t$6 for which there is a submatrix $t$7 with exactly one $t$8 in each column and at most $t$9 $1$0s in each row. One always has

$1$1

The increments of $1$2 are concave in $1$3. Proposition 2.3 states that

$1$4

Thus the marginal gain from increasing the per-row capacity from $1$5 to $1$6 is nonincreasing. This is one of the central structural regularities of the theory.

The paper also studies interchanges, the local moves

$1$7

which preserve row and column sums. Proposition 2.4 shows that a single interchange changes $1$8 by at most $1$9: $m\times n$00 For $m\times n$01, Proposition 2.5 strengthens this: if an interchange increases the $m\times n$02-term rank, then it cannot decrease the $m\times n$03-term rank. In symbols, if

$m\times n$04

then

$m\times n$05

The examples in the paper show that different levels $m\times n$06 can respond differently to a given interchange. One example gives a matrix with

$m\times n$07

while after a specific interchange

$m\times n$08

This demonstrates that higher $m\times n$09-levels are robust but not rigidly tied to lower ones (Brualdi et al., 2010).

4. Classes $m\times n$10, semiregular matrices, and extremal behavior

A major part of the theory concerns classes of matrices with prescribed row and column sums. Let

$m\times n$11

where the entries are nonnegative integers, monotone nonincreasing, and satisfy

$m\times n$12

Then $m\times n$13 denotes the class of all $m\times n$14-matrices with row sum vector $m\times n$15 and column sum vector $m\times n$16. When nonempty, this class is governed by the classical Gale–Ryser/Ford–Fulkerson existence theory; the paper uses the equivalent Ford–Fulkerson matrix $m\times n$17, whose nonnegativity characterizes nonemptiness (Brualdi et al., 2010).

For such a class, the maximum $m\times n$18-term rank is

$m\times n$19

When $m\times n$20, Ryser’s formula gives

$m\times n$21

A particularly rigid case is the semiregular class

$m\times n$22

consisting of all $m\times n$23 $m\times n$24-matrices with exactly $m\times n$25 $m\times n$26s in each row and $m\times n$27 $m\times n$28s in each column, where $m\times n$29. Theorem 3.1 states that for any nonempty semiregular class and any positive integer $m\times n$30,

$m\times n$31

Thus in the semiregular situation the $m\times n$32-term rank depends only on $m\times n$33, not on the specific values of $m\times n$34 beyond the balancing condition $m\times n$35. When $m\times n$36, Corollary 3.2 gives

$m\times n$37

This suggests that semiregularity imposes a uniform extremal geometry on the entire class.

5. Ryser-type formula and joint realization of all maximum $m\times n$38-term ranks

The central extremal result is the $m\times n$39-analogue of Ryser’s formula. If $m\times n$40, then Theorem 4.3 states

$m\times n$41

This is exactly Ryser’s formula with $m\times n$42 replaced by $m\times n$43, reflecting the fact that in the covering interpretation each selected row has multiplicity $m\times n$44 rather than $m\times n$45 (Brualdi et al., 2010).

The proof uses Anstee’s existence theorem, a specialization giving embeddings $m\times n$46 between classes $m\times n$47, and a reduction to the full-$m\times n$48 case via enlarged matrices $m\times n$49. The paper summarizes the resulting criterion as follows: the existence of an $m\times n$50 with $m\times n$51 is equivalent to the family of inequalities

$m\times n$52

Taking the largest admissible $m\times n$53 yields the formula above.

The paper’s most distinctive theorem is the joint realization result. Let

$m\times n$54

Then Theorem 5.2 asserts that there exists a matrix $m\times n$55 and a submatrix $m\times n$56 such that:

1. $m\times n$57 for all $m\times n$58.
2. $m\times n$59 contains exactly $m\times n$60 ones.
3. All these ones lie in the leading $m\times n$61 submatrix.
4. Each of the first $m\times n$62 columns of $m\times n$63 contains exactly one $m\times n$64.
5. Each of the first $m\times n$65 rows of $m\times n$66 contains at most $m\times n$67 ones, with row counts arranged in descending blocks:
   • the first $m\times n$68 rows each containing $m\times n$69 ones,
   • the next $m\times n$70 rows each containing $m\times n$71 ones,
   • and so on down to rows containing one $m\times n$72.

The consequence is that a single matrix can simultaneously attain every extremal value $m\times n$73. The abstract describes this as a surprising result, and it generalizes Haber’s canonical form for maximum ordinary term rank to a block structure encoding all maximum $m\times n$74-term ranks at once (Brualdi et al., 2010).

6. Related notions and terminological ambiguity

Within combinatorial matrix theory, $m\times n$75-term rank is best understood as a constrained matching or covering invariant. It is the rank of the union of $m\times n$76 copies of the transversal matroid $m\times n$77, and it is equivalent to a bipartite $m\times n$78-matching parameter with row capacity $m\times n$79 and column capacity $m\times n$80 (Brualdi et al., 2010). It is not the ordinary rank of a matrix over a field.

The label “T-rank” is, however, used in other literatures for unrelated concepts. In tensor analysis based on the $m\times n$81-product and $m\times n$82-SVD, “T-rank” commonly refers to the tensor tubal rank, the number of nonzero T-singular values of a third-order tensor (Qi et al., 2021). A more recent tensor paper introduces T-phase rank, defined for sectorial third-order tensors as the number of nonzero canonical T-phases, thereby distinguishing a phase-based notion from tubal rank (Kim et al., 12 Feb 2026). In partition theory, Garvan’s higher $m\times n$83-rank denotes a family of partition statistics generalizing the crank and rank for odd $m\times n$84, with moment asymptotics developed in (Waldherr, 2012). These usages are terminologically parallel but mathematically disjoint.

A further possible source of confusion is “TFRank,” an LLM-based ranking system whose name abbreviates “Think-Free Reasoning” rather than any matrix or tensor rank notion (Fan et al., 13 Aug 2025). The combinatorial $m\times n$85-term rank of $m\times n$86-matrices is therefore a specific, classical object, and its theory is centered on matchings, covers, prescribed degree sequences, and matroid union rather than on tensor decompositions, partition statistics, or information retrieval.

In that specific sense, $m\times n$87-term rank occupies a precise place in extremal combinatorics. It extends the ordinary term rank, preserves the matching-covering duality in a modified form, admits exact optimization formulas over classes $m\times n$88, behaves concavely as a function of $m\times n$89, and supports a canonical extremal structure that simultaneously realizes all maximum levels up to a prescribed $m\times n$90 (Brualdi et al., 2010).