Kamae: Theoretical and Applied Perspectives

• Kamae is a cluster of research concepts that define deterministic subsequences through zero-entropy and ergodic criteria, preserving normality in mathematical systems.
• The framework integrates algorithmic randomness reformulations and symbolic complexity invariants, providing rigorous methods to distinguish randomness from periodicity.
• In applied contexts, Kamae’s name designates both a standard astrophysical cross-section model and an open-source Python library bridging Spark preprocessing with Keras models.

Kamae denotes a cluster of research concepts rather than a single object. In mathematics, it refers primarily to the work of Teturo Kamae on subsequences of normal numbers, entropy-theoretic determinism, symbolic complexity, recurrence, and ergodic theory; in astroparticle physics, it names the widely used $pp\to\gamma$ production parameterization of Kamae et al. (2006); and in machine-learning systems, it names an open-source Python library that translates Spark preprocessing pipelines into Keras models [(Takahashi, 2012); (Shen et al., 2019); (Barrowclough et al., 8 Jul 2025)].

1. Kamae and deterministic subsequences of normal numbers

A central Kamae theme is the study of coordinate-selection maps on infinite binary sequences. For $x=x_1x_2\cdots$ and $y=y_1y_2\cdots$, if $\tau$ enumerates the positions where $y_i=1$, then the selected subsequence is

$x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$

In the classical setting, $x$ is normal if every finite binary block appears with the correct limiting frequency, and the problem is to characterize those selectors $y$ for which normality is preserved under $x\mapsto x/y$ [(Takahashi, 2012); (Takahashi, 2011)].

The theorem presented in later work as the Kamae–Weiss theorem states that, under the density condition

$\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$

the following are equivalent: $x=x_1x_2\cdots$0, where $x=x_1x_2\cdots$1 is Kamae entropy, and

$x=x_1x_2\cdots$2

where $x=x_1x_2\cdots$3 is the set of binary normal numbers (Takahashi, 2012). Here $x=x_1x_2\cdots$4 is defined through cluster points of empirical block frequencies; when $x=x_1x_2\cdots$5, the sequence is called completely deterministic [(Takahashi, 2012); (Takahashi, 2011)].

Subsequent work emphasizes that this theorem is foundational because it characterizes exactly which selection sequences preserve normality of every normal input. A deterministic selector, in Kamae’s sense, is “good” precisely when its empirical symbolic statistics have zero entropy (Takahashi, 2012). The 2020 amenable-semigroup generalization preserves this logic: for subsets $x=x_1x_2\cdots$6 of a countable cancellative amenable semigroup with positive lower $x=x_1x_2\cdots$7-density, normality preservation, $x=x_1x_2\cdots$8-determinism, and subexponential complexity are equivalent (2004.02811).

A common misconception is that the Kamae–Weiss preservation phenomenon extends unchanged to all positive-entropy systems. Later work shows a sharper picture. For i.i.d. measures, deterministic sets of positive lower density preserve $x=x_1x_2\cdots$9-normality, but for non-i.i.d. measures with completely positive entropy, deterministic sets—except superficial ones—destroy $y=y_1y_2\cdots$0-normality (Abrams et al., 2022). The same paper also clarifies a remark attributed to Kamae: disjointness from zero-entropy systems preserves a weaker property, simple $y=y_1y_2\cdots$1-normality, meaning preservation of symbol frequencies rather than full block statistics (Abrams et al., 2022).

2. Algorithmic-randomness reformulations of Kamae’s theorem

A second Kamae line concerns algorithmic analogies to the subsequence-selection theorem. Takahashi’s work reformulates the classical entropy-theoretic criterion in terms of Martin-Löf randomness and Kolmogorov-complexity rate, explicitly presenting these results as algorithmic versions of Kamae–Weiss [(Takahashi, 2012); (Takahashi, 2011)].

In one formulation, if $y=y_1y_2\cdots$2 is Martin-Löf random with respect to some computable probability $y=y_1y_2\cdots$3 and

$y=y_1y_2\cdots$4

then the following are equivalent:

1. $y=y_1y_2\cdots$5 is computable;

2.

$y=y_1y_2\cdots$6

where $y=y_1y_2\cdots$7 is the set of Martin-Löf random sequences for the fair-coin measure and $y=y_1y_2\cdots$8 denotes randomness relative to oracle $y=y_1y_2\cdots$9 (Takahashi, 2012).

A second, closer analogue uses maximal complexity rate. If $\tau$0 has maximal complexity rate with respect to a computable probability and

$\tau$1

then the following are equivalent: $\tau$2 and

$\tau$3

This replaces zero Kamae entropy by zero asymptotic information density in the selector (Takahashi, 2012).

These papers are explicit that the algorithmic statements are analogies, not literal restatements. They “neither prove nor disprove the conjecture of van Lambalgen,” which proposed a direct equivalence between zero prefix-complexity rate of $\tau$4 and preservation of Martin-Löf randomness under selection (Takahashi, 2012). The distinction is mathematically significant: the Champernowne sequence satisfies the complexity-rate condition

$\tau$5

but its Kamae entropy is not zero (Takahashi, 2012). This suggests that zero algorithmic information rate and zero Kamae entropy are related but nonidentical notions of determinism.

3. Complexity invariants in symbolic dynamics

Kamae’s name is also attached to several symbolic-complexity invariants. The most prominent is maximal pattern complexity, introduced by Kamae and Zamboni. For a window $\tau$6 of cardinality $\tau$7, the $\tau$8-language $\tau$9 consists of all patterns seen by translating $y_i=1$0 along a sequence $y_i=1$1, and the maximal pattern complexity is

$y_i=1$2

Kamae and Zamboni proved the analogue of Morse–Hedlund: if $y_i=1$3 is not eventually periodic, then

$y_i=1$4

(Le et al., 19 Aug 2025).

They also defined a binary sequence to be pattern Sturmian if

$y_i=1$5

Recent work resolves the question they posed about recurrent pattern Sturmian sequences: a recurrent binary sequence is pattern Sturmian if and only if it is either a recurrent simple circle rotation coding sequence or a sequence in a nearly simple Toeplitz subshift (Le et al., 19 Aug 2025). The same paper shows that nonrecurrent pattern Sturmian sequences are either almost constant or nonrecurrent simple circle rotation coding sequences (Le et al., 19 Aug 2025). A parallel 2025 result over larger alphabets proves that the minimal maximal pattern complexity of an aperiodic sequence using all $y_i=1$6 letters is

$y_i=1$7

and classifies extremal examples as decompositions with one binary pattern Sturmian residue and constant residues elsewhere (Schlortt, 8 May 2025).

A different Kamae-associated invariant is the Kamae–Xue complexity function

$y_i=1$8

the sum of squares of factor-occurrence counts. The 2014 characterization of eventual periodicity shows that an infinite binary word is eventually periodic if and only if

$y_i=1$9

exists and is positive (Kamae et al., 2014). More precisely, if the minimal period is $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$0, then

$x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$1

(Kamae et al., 2014). This places the Kamae–Xue function at the opposite pole from its original “criterion of randomness”: small $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$2 signals balanced block counts, while a positive cubic-order limit signals eventual periodicity (Kamae et al., 2014).

Taken together, these developments show that Kamae-associated complexity invariants operate at both ends of the order–randomness spectrum: maximal pattern complexity isolates low-complexity nonperiodic structure, while $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$3-asymptotics separate random-like quadratic growth from periodic cubic growth [(Le et al., 19 Aug 2025); (Kamae et al., 2014)].

4. Generic points, joinings, and nonstandard ergodic theory

A further Kamae contribution concerns generic points and ergodic theorems. Downarowicz and Weiss generalize a theorem of T. Kamae from 1973. In Kamae’s original full-shift setting, if $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$4 is a joining of invariant measures $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$5 and $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$6, and $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$7 is quasi-generic for $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$8, then there exists $x/y := x_{\tau(1)}x_{\tau(2)}\cdots.$9 such that $x$0 generates $x$1; if $x$2 is ergodic, $x$3 can be chosen generic for $x$4 (Downarowicz et al., 2023). The generalization replaces full shifts by topological systems $x$5 with the weak specification property: if $x$6 has weak specification, $x$7 is ergodic, and $x$8 is quasi-generic for $x$9, then there exists $y$0, generic for $y$1, such that $y$2 is quasi-generic for $y$3 (Downarowicz et al., 2023).

The same paper makes clear what is gained and lost relative to Kamae’s original theorem. The gain is that $y$4 need not be symbolic and $y$5 is produced inside the prescribed system $y$6; the loss is that $y$7 is generally only quasi-generic, even when $y$8 is generic along all of $y$9 (Downarowicz et al., 2023). The proof is explicitly described as following “the framework of the original proof,” with weak specification replacing exact symbolic concatenation (Downarowicz et al., 2023).

Kamae’s name also appears in nonstandard proofs of the ergodic theorem. De Piro’s paper follows Teturo Kamae’s 1982 “A Simple Proof of the Ergodic Theorem Using Nonstandard Analysis” and gives a rigorous proof via Loeb measure, internal approximation, and hyperfinite models (Piro, 2014). A related 2011 paper explains that the proof of the Birkhoff Ergodic Theorem for hyperfinite Loeb spaces, suggested by T. Kamae, “works, actually, for arbitrary probability spaces,” as shown by Katznelson and Weiss, and develops the notion of a hyperfinite approximation of a dynamical system (Glebsky et al., 2011). In that framework, every Lebesgue dynamical system is a homomorphic image of an appropriate transitive Loeb dynamical system (Glebsky et al., 2011).

These strands suggest a coherent Kamae program in ergodic theory: symbolic constructions of generic points, entropy-based determinism, and nonstandard models are all used to convert asymptotic statistical properties into structurally rigid dynamical statements (Downarowicz et al., 2023, Piro, 2014).

5. Intersective polynomials and the Furstenberg–Sárközy phenomenon

In number-theoretic recurrence, the paper on polynomial actions of rings of integers of global fields identifies the modern Furstenberg–Sárközy theorem with Kamae and Mendès France. In that formulation, if $x\mapsto x/y$0 has positive density and $x\mapsto x/y$1 is intersective—meaning it has a root mod $x\mapsto x/y$2 for every $x\mapsto x/y$3—then there exist distinct $x\mapsto x/y$4 such that

$x\mapsto x/y$5

for some $x\mapsto x/y$6 (Ackelsberg et al., 22 Sep 2025). The same paper states that the key observation due to Kamae and Mendès France is that the value set of an intersective polynomial is a van der Corput set (Ackelsberg et al., 22 Sep 2025).

The global-field extension replaces congruence solvability modulo every integer by solvability modulo every nonzero ideal. For a global field $x\mapsto x/y$7 with ring of integers $x\mapsto x/y$8, the following are equivalent for $x\mapsto x/y$9: for every nonzero ideal $\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$0, there exists $\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$1 with $\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$2; recurrence in every $\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$3-measure-preserving system; a syndetic near-optimal recurrence set; and a density-difference statement for every $\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$4 with $\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$5 (Ackelsberg et al., 22 Sep 2025). This is presented as an extension of the result of Kamae and Mendès France (Ackelsberg et al., 22 Sep 2025).

A complementary refinement appears in $\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$6. The 2013 paper on prime powers, recurrence, and van der Corput sets explicitly says that it refines and unifies results of Sárkőzy, Furstenberg, Kamae and Mendes France, and Bergelson–Lesigne (Bergelson et al., 2013). Its central arithmetic sets are

$\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$7

and

$\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$8

where the $\liminf_{n\to\infty}\frac1n\sum_{i=1}^n y_i>0,$9 are positive integers and the $x=x_1x_2\cdots$00 are positive non-integers. These are proved to be nice $x=x_1x_2\cdots$01 sets in $x=x_1x_2\cdots$02, hence van der Corput sets and sets of nice recurrence (Bergelson et al., 2013). The paper also remarks that Kamae and Mendes France showed $x=x_1x_2\cdots$03 is a vdC set if and only if $x=x_1x_2\cdots$04, which explains the special role of prime shifts $x=x_1x_2\cdots$05 (Bergelson et al., 2013).

This branch of the literature fixes a second major Kamae legacy: local solvability, expressed as intersectivity, is the exact criterion connecting difference theorems, recurrence, and uniform distribution [(Ackelsberg et al., 22 Sep 2025); (Bergelson et al., 2013)].

6. Later technical uses of the name

Outside pure mathematics, “Kamae” labels two unrelated technical objects.

In astroparticle physics, the “Kamae et al. (2006)” cross-section model is one of the standard analytic parameterizations for secondary production in proton–proton interactions. A 2019 Fermi-LAT study of a mid-latitude region in the third Galactic quadrant uses interstellar $x=x_1x_2\cdots$06-ray emissivity to infer the local cosmic-ray proton spectrum and compares two hadronic prescriptions: the Kamae et al. (2006) model and Kafexhiu et al. (2014). In that analysis, the Kamae implementation is the “KK06 model,” using the $x=x_1x_2\cdots$07 cross section plus a constant nuclear enhancement factor of $x=x_1x_2\cdots$08; it yields

$x=x_1x_2\cdots$09

with best-fit proton-spectrum parameters

$x=x_1x_2\cdots$10

and a spectrum within $x=x_1x_2\cdots$11 of AMS-02 above $x=x_1x_2\cdots$12 (Shen et al., 2019). The same paper stresses that both Kamae and Kafexhiu fit the emissivity well but imply rather different proton spectra, so cross-section uncertainty remains a dominant bottleneck (Shen et al., 2019). A 2021 comparison with AAfrag identifies the Kamae et al. parameterization as a pre-LHC model based on Pythia 6.2, with photon-spectrum normalization differences reaching $x=x_1x_2\cdots$13–$x=x_1x_2\cdots$14 at intermediate transferred energy fraction $x=x_1x_2\cdots$15 and up to a factor of two as $x=x_1x_2\cdots$16, while still noting its usefulness as a low-energy complement below $x=x_1x_2\cdots$17 in the public package aafragpy (Koldobskiy et al., 2021).

In machine-learning systems, “Kamae” is the name of an open-source Python library for bridging Apache Spark preprocessing pipelines and Keras/TensorFlow inference models. The library defines Spark transformers and estimators, maps each to an equivalent Keras layer, extends Spark’s pipeline API for one-to-one conversion into a Keras model, and currently supports only the TensorFlow backend (Barrowclough et al., 8 Jul 2025). Its stated workflow is: build preprocessing in Spark, fit the pipeline, apply it to distributed training data, export the fitted pipeline to Keras, optionally fuse the preprocessing model with the trained neural model, and serve the resulting Keras model (Barrowclough et al., 8 Jul 2025). Supported transformation families include mathematical, string, date, geographical, logical, array, list, and conditional operations, as well as estimators for string indexing, hash indexing, bloom encoding/indexing, shared indexing, standard scaling, and imputation (Barrowclough et al., 8 Jul 2025). In Expedia’s learning-to-rank search-filters pipeline, the fitted preprocessing model was fused with the model and deployed in production inside a Java chassis serving at an average rate of 200 requests per second; after switching from MLeap to a Keras model plus TensorFlow Java, the reported operational changes were a $x=x_1x_2\cdots$18 decrease in service latency and a $x=x_1x_2\cdots$19 reduction in service costs (Barrowclough et al., 8 Jul 2025).

These later usages are historically unrelated to Teturo Kamae’s mathematical program. Their coexistence under the same name is terminological rather than conceptual.