Takagi Function: Theory and Applications

• Takagi function is a continuous, nowhere differentiable function defined by a dyadic tent-series that exhibits exact self-similarity and a fractal structure.
• Multiple equivalent representations, such as binary expansion and iterated tent-maps, provide deep insights into its combinatorial and level-set properties.
• Generalizations of the Takagi function serve as fundamental models in real analysis, probability, and dynamics, bridging theory with practical computation.

The Takagi function is the classical continuous nowhere differentiable function on $[0,1]$ defined by a dyadic tent-series. In modern notation,

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$

and it extends to a $1$-periodic function on $\mathbb{R}$. Introduced by Teiji Takagi in 1903, it has become a canonical object in real analysis, fractal geometry, ergodic theory, probability, and number theory. Its importance lies not only in being an explicit nowhere differentiable function, but also in the exact dyadic self-similarity that makes many of its fine properties analyzable in closed form (Lagarias, 2011).

1. Definition and canonical representations

On $[0,1]$, the kernel $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$ is the symmetric tent profile

$\phi(x)=\min\{x,1-x\},$

extended periodically to $\mathbb{R}$. The Takagi series

$T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$

converges uniformly, so $T$ is continuous and $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$0-periodic. A related normalization uses the unit-height tent map

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$1

for which $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$2 on $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$3 (Lagarias, 2011).

Several equivalent representations are fundamental. If

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$4

is a binary expansion, and

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$5

then

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$6

Writing

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$7

the integers $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$8 record the dyadic digit imbalance and recur throughout the theory. There is also an iterated tent-map formula

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$9

and a Faber–Schauder interpretation in which the function appears as a particularly rigid dyadic tent expansion (Lagarias, 2011).

The function also admits classical harmonic and integral descriptions. Its Fourier series has coefficients decaying like $1$0 on appropriate frequencies, and

$1$1

These representations place the function simultaneously in the contexts of dyadic expansions, dynamical systems, and harmonic analysis (Lagarias, 2011).

2. Dyadic self-similarity and combinatorial structure

Two functional equations organize most of the geometry: $1$2 Equivalently,

$1$3

Continuity together with these identities determines the function uniquely on $1$4 (Lagarias, 2011).

The dyadic structure becomes explicit on each interval of length $1$5. If $1$6 and $1$7 with $1$8, then

$1$9

When $\mathbb{R}$0, the affine tilt disappears: $\mathbb{R}$1 Thus balanced dyadic points generate exact miniature copies of the full graph. This self-affinity is the source of the graph’s recursive “hump” decomposition and of many level-set constructions (Lagarias, 2011).

The piecewise linear approximants

$\mathbb{R}$2

are linear on each dyadic interval $\mathbb{R}$3, with slope $\mathbb{R}$4. They satisfy the uniform tail bound

$\mathbb{R}$5

and for dyadic rationals $\mathbb{R}$6 the series stabilizes exactly at stage $\mathbb{R}$7 (Lagarias et al., 2010). This makes the function unusually computable at dyadic points and explains why binary combinatorics permeate its analysis.

A recent dyadic refinement connects these values to Hamming weights and divide-and-conquer trees. For $\mathbb{R}$8 with $\mathbb{R}$9, the identity

$[0,1]$0

relates dyadic values of the Takagi function to the number of unbalanced interior nodes in the unique divide-and-conquer full binary tree on $[0,1]$1 leaves, and yields exact recurrences for both $[0,1]$2 and binary digit sums (Monroe, 2021).

3. Regularity and nonsmooth analysis

The Takagi function is continuous everywhere and nowhere differentiable. Cater’s strengthening, recorded in the survey literature, shows that finite one-sided derivatives do not exist either. More sharply, the sets

$[0,1]$3

are both dense in $[0,1]$4 and each has Hausdorff dimension $[0,1]$5 (Lagarias, 2011).

Its modulus of continuity is of Zygmund type rather than Lipschitz. For $[0,1]$6,

$[0,1]$7

and this is essentially sharp. At the same time, $[0,1]$8 belongs to every Hölder class $[0,1]$9 with $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$0. The graph has infinite length on every open subinterval, so the function is not of bounded variation and is not absolutely continuous (Lagarias, 2011).

The failure of smoothness persists under finer notions of differentiability. The function is nowhere approximately derivable: there is no point at which the difference quotient converges along a density-one set of nearby points. The proof exploits the dyadic geometry of the approximants and the oscillation of their integer slopes across arbitrarily small scales (Ferrera et al., 2018).

Nonsmooth analysis gives a more selective description of tangential behavior. The Fréchet subdifferential is empty off the dyadic rationals and equals $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$1 at dyadic points, while the superdifferential is nonempty only on a thin fractal set determined by eventual binary alternation. If

$\phi(x)=\operatorname{dist}(x,\mathbb{Z})$2

then $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$3 is uncountable, has Lebesgue measure zero, and has Hausdorff dimension $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$4 (Ferrera et al., 2019). This sharpens the classical statement “nowhere differentiable” into a detailed stratification of supporting slopes.

4. Extrema, level sets, and local level sets

The range of $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$5 on $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$6 is exactly $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$7. The minimum value $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$8 occurs only at $\phi(x)=\operatorname{dist}(x,\mathbb{Z})$9 and $\phi(x)=\min\{x,1-x\},$0, while

$\phi(x)=\min\{x,1-x\},$1

The top level set $\phi(x)=\min\{x,1-x\},$2 is uncountable and has Hausdorff dimension $\phi(x)=\min\{x,1-x\},$3; for example, $\phi(x)=\min\{x,1-x\},$4. Local minima are exactly the dyadic rationals, and local maxima are characterized by eventual balance conditions on the digit-imbalance sequence $\phi(x)=\min\{x,1-x\},$5 (Lagarias, 2011).

For

$\phi(x)=\min\{x,1-x\},$6

all three cardinality types occur: finite, countably infinite, and uncountable. Almost every level set is finite with respect to Lebesgue measure on ordinates, yet

$\phi(x)=\min\{x,1-x\},$7

so the expected cardinality of a uniformly sampled ordinate level is infinite (Lagarias et al., 2010). A complementary category-theoretic statement goes in the opposite direction: the set of $\phi(x)=\min\{x,1-x\},$8 for which $\phi(x)=\min\{x,1-x\},$9 is uncountable is residual in $\mathbb{R}$0, and the subset of ordinates with countably infinite level sets is dense (Allaart, 2011).

A finer decomposition uses local level sets. Two points are locally equivalent when the absolute digit-balance profile agrees at every depth, and each global level set partitions into disjoint local level sets. Every local level set is either finite or a Cantor set. The expected number of local level sets contained in $\mathbb{R}$1, for $\mathbb{R}$2 uniformly distributed on $\mathbb{R}$3, is exactly

$\mathbb{R}$4

This value arises from the Catalan-number enumeration of leading humps and their truncated projections (Lagarias et al., 2010).

The geometric size of level sets is also sharply understood. For every ordinate,

$\mathbb{R}$5

and this bound is optimal: the set of levels with Hausdorff dimension $\mathbb{R}$6 is dense, and $\mathbb{R}$7 itself has dimension $\mathbb{R}$8 (Lagarias, 2011). More generally, the set

$\mathbb{R}$9

has Hausdorff dimension $T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$0 but Lebesgue measure $T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$1 (Lagarias et al., 2010). On the cardinality side, the most common finite cardinality is $T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$2; if

$T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$3

then

$T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$4

for Lebesgue measure on the ordinate axis (Allaart, 2011).

5. Generalizations and the Takagi class

The classical function is the basic member of several larger families built from the same dyadic tent mechanism. One natural extension is the signed family

$T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$5

These functions are continuous, symmetric about $T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$6, and nowhere differentiable. Their graph height depends explicitly on the first hitting times of the sign-walk, and the average number of local level sets in a level set equals the reciprocal of that height, hence always lies between $T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$7 and $T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$8 (Allaart, 2012).

A broader variable-sign class is

$T(x)=\sum_{n=0}^{\infty}2^{-n}\phi(2^n x)$9

where each $T$0 is constant on dyadic intervals of level $T$1. This includes the classical Takagi function, the alternating Takagi function, the Gray Takagi function, and Kawamura’s function $T$2. Every member is continuous and nowhere differentiable, and every graph has Hausdorff dimension $T$3. For the constant-sign subclass $T$4 one has the sharp uniform level-set bound

$T$5

while for the full variable-sign class $T$6 the optimal value is

$T$7

(Allaart, 2013).

Base-$T$8 analogues,

$T$9

form the Takagi–van der Waerden family. For even integers $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$00, a generalized local-level-set theory exists, with exact expectation

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$01

for a uniformly chosen ordinate $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$02 in the range of $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$03. At the same time,

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$04

despite almost-everywhere finiteness of level sets (Jiang et al., 29 Aug 2025).

The term Takagi class is also used for expansions

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$05

with $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$06. In the borderline regime $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$07, the graph has Assouad dimension $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$08; in particular,

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$09

for

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$10

(Jiang, 3 Feb 2025). Other recent extensions include matrix-parametric derivatives of singular functions, which recover the classical Takagi function as a parameter derivative of the Lebesgue singular function (Okamura, 2015), and beta-expansion analogues for $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$11, which remain pointwise $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$12-Hölder for every $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$13 but are not pointwise Lipschitz outside a Lebesgue-null set (Suzuki, 20 Apr 2026).

6. Interfaces with analysis, probability, number theory, and dynamics

The Takagi function appears naturally in digital-sum asymptotics. If $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$14 denotes the summatory base-$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$15 digit sum, then the binary case of the Trollope–Delange theory expresses the error term through $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$16. For integers $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$17 with $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$18, writing $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$19 with $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$20,

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$21

There is also a Farey-sum criterion equivalent to the Riemann hypothesis: $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$22 (Lagarias, 2011).

In probability, $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$23 occurs as the derivative at the symmetric point of a family of singular Bernoulli-convolution distribution functions: $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$24 More generally, probabilistic limit theorems for Takagi-class tails have been developed via reverse-martingale methods. These yield law-of-large-numbers, central-limit-theorem, and law-of-the-iterated-logarithm regimes for generalized coefficient sequences, while the classical geometric-coefficient case shows that the Takagi tail does not satisfy the law of large numbers in the usual pointwise sense [(Lagarias, 2011); (Osaka et al., 2019)].

Recent work has added genuinely dynamical and fractal-slicing perspectives. For the discrete dynamical system generated by iteration of $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$25, almost every orbit converges to the fixed point $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$26, and the map has the shadowing property; however, scaled versions

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$27

can fail shadowing for infinitely many parameters $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$28 (Buczolich et al., 23 Mar 2026). On the geometric side, for the graph $$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$29 of the generalized Takagi function

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$30

the maximal Hausdorff dimension of a slice is exactly

$$T(x)=\sum_{n=0}^{\infty}2^{-n}\,\phi(2^n x),\qquad \phi(x)=\operatorname{dist}(x,\mathbb{Z}),$$31

and this bound is sharp (Anttila et al., 2023). This suggests that the Takagi graph is not merely a pathological curve, but a tractable self-affine object whose slices, projections, and tangents encode nontrivial dimension theory.

Across these settings, the Takagi function serves as a rare example in which extreme irregularity coexists with exact symbolic structure. Its dyadic self-similarity makes it a prototype for continuous nonsmooth functions, while its level sets, graph geometry, and arithmetic identities continue to generate new results well beyond the classical theory (Lagarias, 2011).