Subspace Polynomials in Finite Fields

• Subspace Polynomials are monic q-linearized polynomials that divide x^(q^n)-x, providing an exact algebraic representation of F_q-subspaces.
• They enable construction of cyclic subspace codes with controlled minimum distance via gap invariants and explicit families like subfields and trinomials.
• Their structure underpins applications in Reed–Solomon repair schemes, network coding, and distributed storage by ensuring efficient error correction.

In finite-field algebra, a subspace polynomial over $\mathbb{F}_{q^n}$ is a monic $q$-linearized polynomial

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$

that divides $x^{[n]}-x$; equivalently, it splits completely over $\mathbb{F}_{q^n}$ into distinct roots. This notion gives an exact algebraic representation of $\mathbb{F}_q$-subspaces of $\mathbb{F}_{q^n}$, and it is used in the construction of cyclic subspace codes for random network coding as well as in repair schemes for Reed–Solomon codes (Ben-Sasson et al., 2014, Dau et al., 2020).

1. Definition in the finite-field setting

Let $\mathbb{F}_q$ denote the finite field of order $q$, and let $\mathbb{F}_{q^n}$ be the degree-$q$0 extension field, often identified with the vector space $q$1. A $q$2-linearized polynomial, also called a $q$3-polynomial, is a polynomial of the form

$q$4

with coefficients in $q$5 and $q$6. It is monic when $q$7. A monic $q$8-linearized polynomial is a subspace polynomial if and only if it divides $q$9, or equivalently if and only if it splits completely over $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$0 into distinct linear factors (Ben-Sasson et al., 2014).

In the notation of linearized polynomials, the ordinary degree and the $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$1-degree play different roles. For subspace polynomials, the relevant degree parameter is the $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$2-degree $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$3, which matches the dimension of the associated $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$4-subspace. The same finite-field definition also appears in later work on cyclic constant-dimension codes, where a monic $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$5-polynomial $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$6 is called a subspace polynomial with respect to $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$7 exactly when it divides $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$8 (Wang et al., 23 Sep 2025).

A related formulation is used in coding-theoretic applications over $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$9 with base field $x^{[n]}-x$0. If $x^{[n]}-x$1 is an $x^{[n]}-x$2-dimensional $x^{[n]}-x$3-subspace, its subspace polynomial is

$x^{[n]}-x$4

This polynomial can be written in linearized form

$x^{[n]}-x$5

so it defines a $x^{[n]}-x$6-linear map $x^{[n]}-x$7 (Dau et al., 2020).

2. Root spaces and the representation theorem

A basic structural fact, traced in the literature to Ore and standard finite-field references, is that if a $x^{[n]}-x$8-linearized polynomial splits over $x^{[n]}-x$9, then all of its roots in $\mathbb{F}_{q^n}$0 form an $\mathbb{F}_{q^n}$1-linear subspace. Conversely, for any $\mathbb{F}_{q^n}$2-dimensional subspace $\mathbb{F}_{q^n}$3, the polynomial

$\mathbb{F}_{q^n}$4

is a monic $\mathbb{F}_{q^n}$5-linearized polynomial of $\mathbb{F}_{q^n}$6-degree $\mathbb{F}_{q^n}$7 whose roots are exactly $\mathbb{F}_{q^n}$8, each with multiplicity $\mathbb{F}_{q^n}$9 (Ben-Sasson et al., 2014).

This yields a one-to-one correspondence between subspaces and subspace polynomials. Two $\mathbb{F}_q$0-dimensional subspaces $\mathbb{F}_q$1 are equal if and only if $\mathbb{F}_q$2. Equivalently, for every $\mathbb{F}_q$3-dimensional $\mathbb{F}_q$4-subspace $\mathbb{F}_q$5 there exists a unique monic $\mathbb{F}_q$6-linearized polynomial

$\mathbb{F}_q$7

characterized by $\mathbb{F}_q$8 and $\mathbb{F}_q$9; conversely, every such divisor arises from a unique $\mathbb{F}_{q^n}$0-subspace (Ben-Sasson et al., 2014).

Several coefficient-level consequences are used repeatedly. In any subspace polynomial $\mathbb{F}_{q^n}$1, the coefficient $\mathbb{F}_{q^n}$2 is nonzero. Since $\mathbb{F}_{q^n}$3, the polynomial $\mathbb{F}_{q^n}$4 factorizes over $\mathbb{F}_{q^n}$5 into irreducible $\mathbb{F}_{q^n}$6-linearized factors whose degrees divide $\mathbb{F}_{q^n}$7 (Ben-Sasson et al., 2014).

In the Reed–Solomon repair setting, the same correspondence is expressed in kernel language. The map $\mathbb{F}_{q^n}$8 is $\mathbb{F}_{q^n}$9-linear, $\mathbb{F}_q$0, $\mathbb{F}_q$1, $\mathbb{F}_q$2, and $\mathbb{F}_q$3. All roots are simple, and $\mathbb{F}_q$4 (Dau et al., 2020).

3. Explicit families and the gap invariant

The simplest explicit family arises from subfields. If $\mathbb{F}_q$5, then the subfield $\mathbb{F}_q$6 is a $\mathbb{F}_q$7-dimensional $\mathbb{F}_q$8-subspace, and its subspace polynomial is

$\mathbb{F}_q$9

This example makes the divisor condition transparent, since $q$0 already vanishes exactly on the subfield (Ben-Sasson et al., 2014).

A second family comes from trinomials. Under the irreducibility hypothesis stated for the corresponding ordinary trinomial in $q$1, the polynomial

$q$2

is a subspace polynomial over $q$3. The resulting $q$4-subspace $q$5 yields a cyclic subspace code

$q$6

of size

$q$7

and minimum subspace distance at least $q$8. By factoring the trinomial $q$9 over $\mathbb{F}_{q^n}$0 and taking $\mathbb{F}_{q^n}$1 to be any multiple of the least common multiple of its factor-degrees, one obtains infinitely many $\mathbb{F}_{q^n}$2 for each fixed $\mathbb{F}_{q^n}$3 giving a cyclic code of size $\mathbb{F}_{q^n}$4 and distance at least $\mathbb{F}_{q^n}$5 (Ben-Sasson et al., 2014).

The coefficient pattern of a subspace polynomial is summarized by the gap parameter. If

$\mathbb{F}_{q^n}$6

then

$\mathbb{F}_{q^n}$7

If $\mathbb{F}_{q^n}$8 have gaps $\mathbb{F}_{q^n}$9 and $q$00, then

$q$01

so the subspace distance satisfies

$q$02

This algebraic bound is the basic distance estimate behind many cyclic-code constructions (Ben-Sasson et al., 2014).

Small-field instances illustrate the same mechanism. Over $q$03 with $q$04 and $q$05, if $q$06 then

$q$07

If $q$08, where $q$09 is a root of $q$10, then

$q$11

These examples exhibit explicit linearized forms for concrete root subspaces (Dau et al., 2020).

4. Cyclic subspace codes and orbit structure

For a subspace $q$12 and $q$13, the cyclic shift of $q$14 is

$q$15

The associated subspace polynomial transforms by

$q$16

Thus cyclic shifts act directly on the coefficients of the representing linearized polynomial (Ben-Sasson et al., 2014).

The orbit size of $q$17 under all nonzero scalars is

$q$18

where $q$19 is the minimal divisor of $q$20 such that every nonzero coefficient index $q$21 of $q$22 satisfies $q$23. A subspace has a full-length orbit when $q$24, in which case the size is

$q$25

This criterion turns coefficient support into an orbit-length invariant (Ben-Sasson et al., 2014).

One-orbit codes with large size and controlled distance are obtained by choosing $q$26 so that $q$27 has a nonzero $q$28-coefficient. Then $q$29, its orbit under $q$30 has size $q$31, and the gap bound gives pairwise distance at least $q$32. The irreducible-trinomial construction provides infinitely many such examples (Ben-Sasson et al., 2014).

The orbit construction can also be enlarged by Frobenius shifts. If $q$33, then under mild coefficient conditions one has

$q$34

and this union is still cyclic, has size

$q$35

and has the same minimum distance $q$36 (Ben-Sasson et al., 2014).

5. Generalized multi-orbit constructions

A later extension considers several subspace polynomials simultaneously. Fix a prime power $q$37, integers $q$38 and $q$39, a field extension $q$40, and a positive integer $q$41. For each $q$42, choose nonzero coefficients $q$43 and form

$q$44

If each $q$45 divides $q$46, then its root set

$q$47

is an $q$48-subspace of dimension exactly $q$49 (Wang et al., 23 Sep 2025).

Writing

$q$50

one has

$q$51

since $q$52 is not $q$53-linear for any $q$54. By comparing the greatest common divisors of $q$55 and $q$56, one obtains

$q$57

so the minimum distance of the single-orbit code $q$58 satisfies

$q$59

To combine several orbits, one further requires for any $q$60 and any $q$61 that

$q$62

This is enforced through the polynomial

$q$63

with

$q$64

together with a determinant/rank condition asserting that a certain $q$65 matrix $q$66 has full column rank $q$67 (Wang et al., 23 Sep 2025).

Under that matrix-rank condition, the union of orbits

$q$68

remains an $q$69-code with

$q$70

and size

$q$71

When $q$72 and $q$73, this recovers earlier trinomial-based constructions. Allowing arbitrary $q$74 and multiple distinct polynomials $q$75 gives up to $q$76 disjoint orbits rather than a single one, so the resulting codes can be much larger. An explicit example takes $q$77, $q$78, three polynomials over $q$79, embeds their root spaces in $q$80, verifies the rank conditions, and obtains a cyclic $q$81-code of size $q$82, whereas the earlier single-trinomial construction gives size $q$83 (Wang et al., 23 Sep 2025).

6. Reed–Solomon repair and the relation to trace polynomials

In distributed storage, subspace polynomials are used to construct repair checks for Reed–Solomon codes over $q$84 with base field $q$85. For an $q$86-dimensional $q$87-subspace $q$88, the polynomial

$q$89

is $q$90-linear, has kernel $q$91, degree $q$92, divides $q$93, and splits completely over $q$94 with simple roots. The associated conventional $q$95-associate is

$q$96

and under symbolic composition $q$97, the associate of $q$98 is $q$99 (Dau et al., 2020).

Trace polynomials appear as a special case. If $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$00 has dimension $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$01, then

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$02

and $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$03. In this sense, subspace polynomials generalize trace polynomials by allowing the kernel to be any $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$04-dimensional $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$05-subspace, not only a hyperplane. The Guruswami–Wootters repair scheme uses only $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$06, whereas the subspace-polynomial approach uses all subspaces $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$07 of dimension $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$08 (Dau et al., 2020).

For an $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$09 Reed–Solomon code over $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$10 with redundancy $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$11, if the symbol at $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$12 is erased, one chooses an $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$13-dimensional subspace $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$14 and defines

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$15

The resulting checks have degree at most $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$16, the values at the erased position span $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$17 over $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$18, and for each helper node at $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$19 one has

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$20

Hence the total repair bandwidth, measured in $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$21-subsymbols, is

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$22

When $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$23 and $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$24, this bandwidth is information-theoretically optimal for one erasure. For two erasures, the same bandwidth per erasure is obtained when $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$25 is a power of $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$26, and also for $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$27, $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$28 with $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$29, and for $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$30 when $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$31 is even and $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$32 is a power of two (Dau et al., 2020).

7. Distinct usage in learning theory

A separate line of work studies subspace-sparse polynomials, which are not finite-field subspace polynomials. In that setting, the ambient space is $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$33, the input distribution is standard Gaussian, and a target function is called subspace-sparse when there exists an unknown $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$34-dimensional subspace $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$35 with orthogonal projector $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$36 such that

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$37

where $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$38 is a polynomial of total degree $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$39 (Chen et al., 2024).

The learning model is a wide two-layer network represented in the mean-field limit by a probability distribution $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$40 on parameters $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$41, with network output

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$42

and population squared loss

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$43

In the vanishing-step-size and infinite-width limit, the SGD evolution satisfies a mean-field PDE, and

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$44

so the flow is a gradient descent in $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$45-Wasserstein space (Chen et al., 2024).

The conceptual overlap with finite-field subspace polynomials lies only in the shared emphasis on low-dimensional subspace structure. The actual objects are different. In the learning-theoretic setting, the main results concern a necessary condition based on a reflective property of $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$46 on a proper subspace $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$47, under which the loss stays bounded away from zero for finite time horizons, and an almost-sufficient condition under which one can design a two-stage SGD schedule achieving

$$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$48

with constants depending on $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$49 but not on the ambient dimension $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$50 (Chen et al., 2024).

This distinction matters terminologically. In finite-field coding theory, a subspace polynomial is a monic linearized divisor of $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$51 that represents an $$P(x)=x^{[k]}+a_{k-1}x^{[k-1]}+\cdots+a_1x^{[1]}+a_0x,\qquad x^{[i]}:=x^{q^i},$$52-subspace. In contemporary learning theory, a subspace-sparse polynomial is a real polynomial target that depends only on the projection of the input onto a low-dimensional subspace. The two notions share a geometric motif but belong to different algebraic and analytic frameworks (Ben-Sasson et al., 2014, Chen et al., 2024).