Almost Polynomial Growth

• Almost polynomial growth is a cluster of threshold phenomena with distinct definitions in formal languages, PI-theory, and coarse geometry.
• In formal language settings, it reduces to a strict polynomial/exponential dichotomy, while in PI-theory it identifies minimal exponential growth.
• The concept underpins sharp growth thresholds and structural classifications across automation, algebra, and geometric group theory.

Almost polynomial growth is not a single invariant but a cluster of near-threshold growth phenomena whose meaning depends strongly on context. In automata theory and some group-theoretic counting problems, it is best understood negatively: there is a sharp polynomial/exponential dichotomy, so no genuine intermediate regime exists. In PI-theory and its variants, it is a formal minimality notion: a variety has exponential codimension growth, but every proper subvariety has polynomial growth. In coarse geometry and analysis, the phrase is usually indirect and refers instead to weakly polynomial growth, strict sub-$r^{k+1}$ growth, or coarse equivalence to polynomial-growth models rather than to a separate named asymptotic class (0711.4990, Martino et al., 18 Feb 2025, Papasoglu, 2021).

1. Terminological scope

Across the papers represented here, the phrase is used in three non-equivalent ways.

Setting	Growth object	Role of “almost polynomial growth”
Regular languages, geodesic languages	word counts	usually ruled out by a dichotomy
PI-type algebraic varieties	codimension sequences	minimal non-polynomial growth
Geometric and analytic settings	ball growth, volume growth, harmonic growth	threshold or coarse-polynomial behavior

In the regular-language setting, growth is counted by exact word length, typically $g_L(m)=|L\cap \Sigma^m|$, and the main result is that regular languages are either polynomially bounded or exponential, with no subexponential-but-superpolynomial case. In PI-theory, generalized PI-theory, and trace identities, “almost polynomial growth” is a technical term for a minimal exponential-growth variety. In large-scale geometry, the more common formulations are “weakly polynomial growth,” “growth strictly less than $r^{k+1}$,” or “roughly isometric to a polynomial-growth Cayley graph,” each of which captures a different borderline form of polynomial behavior (0711.4990, Martino et al., 18 Feb 2025, Ioppolo et al., 2020, Papasoglu, 2021).

This divergence of meaning is structural rather than terminological accident. The underlying growth objects are different: exact-length counting in formal-language theory, multilinear codimensions in PI-theory, ball-volume functions in coarse geometry, and harmonic-function dimensions or random-walk displacement in analysis. Any encyclopedia treatment therefore has to distinguish these regimes rather than seek a universal definition.

2. Formal languages and sharp dichotomies

For a regular language $L\subseteq \Sigma^*$, the relevant growth function is the exact-length count

$g_L(m)=|L\cap \Sigma^m|.$

Polynomial growth means that there exists a polynomial $p(x)$ such that

$|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$

while exponential growth means that there exists $r>1$ such that

$|L\cap \Sigma^m|>r^m$

for infinitely many $m>0$. The central fact is a gap theorem: for regular languages, and more generally for context-free languages, no intermediate growth is possible. In this setting, “almost polynomial growth” is therefore not a genuine asymptotic category. A regular language is polynomially bounded or it is exponential; there is no regular analogue of $g_L(m)=|L\cap \Sigma^m|$0 or $g_L(m)=|L\cap \Sigma^m|$1 (0711.4990).

The structural reason is loop commutativity. Let $g_L(m)=|L\cap \Sigma^m|$2 be an NFA, and for each state $g_L(m)=|L\cap \Sigma^m|$3 let $g_L(m)=|L\cap \Sigma^m|$4 be the loop language at $g_L(m)=|L\cap \Sigma^m|$5. Polynomial growth holds if and only if every $g_L(m)=|L\cap \Sigma^m|$6 is commutative in the sense that $g_L(m)=|L\cap \Sigma^m|$7 for some word $g_L(m)=|L\cap \Sigma^m|$8. If some state supports two loop words $g_L(m)=|L\cap \Sigma^m|$9 with $r^{k+1}$0, then one can build exponentially many distinct accepted words of the same length, forcing exponential growth. This gives both a structural characterization and an algorithmic one: whether an NFA with $r^{k+1}$1 states and $r^{k+1}$2 transitions has polynomial or exponential growth can be decided in

$r^{k+1}$3

time, and if a DFA is already known to have polynomial growth, its exact order can be computed in

$r^{k+1}$4

time by contracting disjoint cycle components and finding the maximal number $r^{k+1}$5 of special cycle-vertices on an accepting path, which yields growth

$r^{k+1}$6

Thus, in the regular setting, “almost polynomial” collapses to “polynomial” (0711.4990).

A related dichotomy appears for geodesic growth in finitely generated virtually abelian groups. Here one fixes a finite weighted monoid generating set $r^{k+1}$7, and the cumulative geodesic growth function counts geodesic words of weight at most $r^{k+1}$8: $r^{k+1}$9 For every such $L\subseteq \Sigma^*$0, geodesic growth is either polynomial, with rational geodesic growth series, or exponential, with holonomic geodesic growth series. There is no intermediate geodesic growth in this class. The phenomenon is analogous to the regular-language gap theorem, but the counted objects are geodesic words rather than accepted words of fixed length (Bishop, 2019).

3. PI-theoretic meanings: minimal exponential growth

In associative PI-theory and its variants, “almost polynomial growth” is a precise minimality property. A variety has almost polynomial growth when its codimension sequence is not polynomially bounded—indeed exponential in the papers considered here—but every proper subvariety has polynomial growth. In this sense, the phrase does not designate a borderline asymptotic rate between polynomial and exponential; it designates a minimal obstruction to polynomial boundedness.

For finite-dimensional $L\subseteq \Sigma^*$1-algebras over characteristic $L\subseteq \Sigma^*$2, generalized codimensions are defined by

$L\subseteq \Sigma^*$3

where $L\subseteq \Sigma^*$4 is the multilinear generalized polynomial space modulo identities. The decisive criterion is that polynomial growth holds if and only if neither $L\subseteq \Sigma^*$5 nor $L\subseteq \Sigma^*$6 lies in the generated generalized variety. Consequently,

$L\subseteq \Sigma^*$7

are the only finite-dimensional $L\subseteq \Sigma^*$8-algebras generating generalized varieties of almost polynomial growth. The multiplier-algebra formalism is essential here: it reduces arbitrary $L\subseteq \Sigma^*$9-actions to an intrinsic algebraic framework and makes the classification independent of the specific structure of $g_L(m)=|L\cap \Sigma^m|.$0 (Martino et al., 18 Feb 2025).

A parallel classification exists for $g_L(m)=|L\cap \Sigma^m|.$1-graded algebras with homogeneous involution. In that setting the codimension sequence is

$g_L(m)=|L\cap \Sigma^m|.$2

and polynomial growth is characterized by the exclusion of a finite forbidden family

$g_L(m)=|L\cap \Sigma^m|.$3

together with the additional order-$g_L(m)=|L\cap \Sigma^m|.$4 family $g_L(m)=|L\cap \Sigma^m|.$5 when $g_L(m)=|L\cap \Sigma^m|.$6 is even and $g_L(m)=|L\cap \Sigma^m|.$7. The varieties of almost polynomial growth are exactly those generated, up to $g_L(m)=|L\cap \Sigma^m|.$8-equivalence, by

$g_L(m)=|L\cap \Sigma^m|.$9

Here the forbidden algebras isolate the only minimal mechanisms for exponential growth: exchange-type involution obstruction, group-algebra obstruction, and radical-bridge obstruction (Cota et al., 4 Dec 2025).

The trace setting adds a further layer. For an algebra with trace, the $p(x)$0-th trace codimension is

$p(x)$1

The decisive new feature is that commutative algebras can become minimally exponential once a trace is added. For the diagonal algebra $p(x)$2 and the commutative algebra $p(x)$3, the paper computes explicit trace codimensions: $p(x)$4 for $p(x)$5, $p(x)$6, and proves exponential growth for $p(x)$7. Polynomial growth for a finite-dimensional trace algebra is characterized by excluding the finite list

$p(x)$8

As a consequence, the only finite-dimensional algebras with trace generating varieties of almost polynomial growth are precisely the corresponding trace-decorated $p(x)$9- and $|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$0-examples, together with $|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$1 with zero trace (Ioppolo et al., 2020).

Setting	Polynomial-growth criterion	Minimal almost-polynomial examples
$|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$2-algebras	exclude $|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$3	$|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$4
$|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$5-graded with homogeneous involution	exclude all $|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$6	$|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$7
Algebras with trace	exclude $|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$8, $|L\cap \Sigma^m|<p(m)\qquad \text{for all }m\ge 0,$9- and $r>1$0-trace obstructions	$r>1$1

These classification theorems all have the same logical form. Polynomial growth is equivalent to excluding a short list of finite-dimensional obstructions, and almost polynomial growth is equivalent to being generated by one of the minimal obstructions themselves. The resulting growth dichotomy is polynomial versus exponential, not polynomial versus intermediate versus exponential.

4. Group growth, weakly polynomial growth, and coarse thresholds

In geometric group theory, the relevant growth function is usually ball volume. For a finitely generated group $r>1$2 with finite symmetric generating set $r>1$3, one writes

$r>1$4

If $r>1$5 has polynomial growth, then by Gromov’s theorem and Bass’s estimates there exists an integer $r>1$6 and constants $r>1$7 such that

$r>1$8

For Cayley graphs of such groups, the space of harmonic functions of polynomial growth of order at most $r>1$9,

$|L\cap \Sigma^m|>r^m$0

has dimension bounded by

$|L\cap \Sigma^m|>r^m$1

and this bound is asymptotically optimal. The same exponent survives on bounded-geometry graphs roughly isometric to such Cayley graphs. This shows that certain sharp polynomial-growth consequences are stable under coarse perturbation, but only when there remains a polynomial-growth Cayley graph in the background (Hua et al., 2012).

Ozawa introduced a weaker asymptotic condition for finitely generated groups: $|L\cap \Sigma^m|>r^m$2 calling it weakly polynomial growth. Through reduced first cohomology, $|L\cap \Sigma^m|>r^m$3-harmonic $|L\cap \Sigma^m|>r^m$4-cocycles, and Shalom’s property $|L\cap \Sigma^m|>r^m$5, he proved a functional-analytic route to Gromov’s theorem: a finitely generated group with weakly polynomial growth contains a nilpotent subgroup of finite index. In that paper, the decisive intermediary is slow entropy growth rather than a separate “almost polynomial” asymptotic class. This suggests that in group theory the natural borderline notion may be entropy-theoretic rather than purely volumetric (Ozawa, 2015).

A different threshold phenomenon appears for asymptotic dimension. For a connected graph with uniform growth function

$|L\cap \Sigma^m|>r^m$6

the condition

$|L\cap \Sigma^m|>r^m$7

implies

$|L\cap \Sigma^m|>r^m$8

The same conclusion holds for bounded-geometry Riemannian manifolds when the volume function satisfies $|L\cap \Sigma^m|>r^m$9. This is a genuine strict-subthreshold statement: growth strictly below $m>0$0 forces asymptotic dimension at most $m>0$1. Yet the stronger asymptotic Assouad–Nagata dimension behaves differently: there are graphs with growth

$m>0$2

and still infinite asymptotic Assouad–Nagata dimension. Thus very slow superlinear growth controls asymptotic dimension but not the stronger linear-control covering invariant (Papasoglu, 2021).

5. Random and nonsmooth threshold phenomena

Long-range percolation on $m>0$3 provides a sharp random threshold. When nearest-neighbor edges are always open and long edges appear with probability comparable to $m>0$4, graph-metric balls have polynomial growth if and only if

$m>0$5

For $m>0$6, growth is superpolynomial. For $m>0$7, chemical distance is linear and balls have Euclidean-order polynomial growth. The critical case $m>0$8 is the canonical borderline regime: there exists an exponent $m>0$9 such that

$g_L(m)=|L\cap \Sigma^m|$00

and in fact

$g_L(m)=|L\cap \Sigma^m|$01

Thus the transition at $g_L(m)=|L\cap \Sigma^m|$02 is not to an intermediate asymptotic class but to exact polynomial growth with a non-Euclidean exponent (Bäumler, 2023).

Stationary random graphs show a different limitation. If $g_L(m)=|L\cap \Sigma^m|$03 has annealed polynomial growth,

$g_L(m)=|L\cap \Sigma^m|$04

or even weakly annealed polynomial growth,

$g_L(m)=|L\cap \Sigma^m|$05

then there exists an infinite deterministic sequence of times $g_L(m)=|L\cap \Sigma^m|$06 at which the quenched mean-square displacement is at most diffusive: $g_L(m)=|L\cap \Sigma^m|$07 However, the subsequence is essential. There are stationary random graphs of almost sure polynomial growth for which the random walk is almost surely superdiffusive at an infinite subset of times. In this setting, polynomial growth does not force all-time diffusive behavior; it only guarantees good times (Ganguly et al., 2016).

Noncompact $g_L(m)=|L\cap \Sigma^m|$08 spaces with linear volume growth,

$g_L(m)=|L\cap \Sigma^m|$09

illustrate a nonsmooth rigidity phenomenon. If a slab between two level sets of a distance function has almost maximal volume compared to a cylinder, then that slab is Gromov–Hausdorff close to a product cylinder. Applied to Busemann functions, this yields

$g_L(m)=|L\cap \Sigma^m|$10

and, in the nonsplitting case,

$g_L(m)=|L\cap \Sigma^m|$11

The analytic consequence is stringent: if such a space admits a nonconstant harmonic function of polynomial growth, then it must split isomorphically as

$g_L(m)=|L\cap \Sigma^m|$12

with $g_L(m)=|L\cap \Sigma^m|$13 compact. In other words, linear volume growth plus nonsplitting excludes nonconstant polynomial-growth harmonic functions (Huang, 2017).

6. Consequences, adjacent notions, and common misconceptions

Polynomial-growth hypotheses also have dynamical and ergodic consequences that clarify what analysts often mean by a “polynomial-style” regime. For a minimal action of a finitely generated group of polynomial growth on a compact metrizable space, polynomial growth implies comparison. For free minimal actions, the small boundary property is then equivalent to almost finiteness, and this implies that the crossed product $g_L(m)=|L\cap \Sigma^m|$14 is $g_L(m)=|L\cap \Sigma^m|$15-stable and classifiable by the Elliott invariant. The proof uses the quantitative regularity of large word-metric balls, especially bounded doubling at large scales, rather than amenability alone (Naryshkin, 2021).

A complementary ergodic theorem shows that ball averages on groups of polynomial growth satisfy jump inequalities and upcrossing inequalities with exponential decay. The abstract framework uses annular decay and geometric doubling, and the polynomial-growth case enters by verifying those geometric properties for word metrics on polynomial-growth groups. This suggests that, in analytic applications, what often matters is not a named “almost polynomial growth” class but polynomial-style boundary regularity of balls (Hong et al., 2021).

A nearby but distinct notion appears in nonassociative algebra. An almost nilpotent variety is non-nilpotent but every proper subvariety is nilpotent. In the metabelian nonassociative setting there exist a countable family of almost nilpotent varieties of at most linear growth and an uncountable family of at most quadratic growth. This concerns minimal non-nilpotent behavior, not the PI-theoretic notion of almost polynomial growth, but it shows that minimality phenomena at low polynomial orders can be abundant outside the associative framework (Mishchenko et al., 2017).

Three misconceptions therefore recur. First, “almost polynomial growth” is not a universal intermediary between polynomial and exponential growth. In several fundamental settings—regular languages and geodesic growth in virtually abelian groups—it does not exist as an asymptotic regime at all. Second, in PI-theory it does not mean subexponential growth: it means minimal exponential growth. Third, in coarse geometry and analysis the interesting borderline statements are often formulated not by this phrase, but by conditions such as weakly polynomial growth, $g_L(m)=|L\cap \Sigma^m|$16 growth, annular decay, or rough isometry to a polynomial-growth model. The common theme is threshold behavior near polynomial regimes, but the mathematical content depends entirely on what is being counted.