Hölder Continuity and Non-Differentiability¶

In Brownian Motion Foundations we saw that Brownian increments satisfy \(\mathbb{E}[(W_{t+h}-W_t)^2] = h\), so a typical increment has size \(\sqrt{h}\) rather than \(h\). This \(\sqrt{h}\) scaling is the fingerprint of an unusual regularity class: the paths are continuous — in fact Hölder continuous of every order \(\alpha < \tfrac{1}{2}\) — yet they are nowhere differentiable and have infinite total variation on every interval.

Understanding this dichotomy is essential before building stochastic calculus. Hölder continuity explains why paths exist and are well-behaved enough to study. Nowhere differentiability and infinite total variation explain why the Riemann–Stieltjes integral \(\int f(t)\,dW_t\) cannot be defined in the classical sense and why a new calculus (Itô's calculus) is required.

Hölder Continuity¶

Definition¶

Examples:

• \(f(t) = t^{1/2}\) is Hölder-\(\tfrac{1}{2}\) but not Hölder-\(\alpha\) for any \(\alpha > \tfrac{1}{2}\).
• Every \(C^1\) function is Hölder-\(1\) (Lipschitz).
• The Cantor function is Hölder-\((\log 2/\log 3)\) but nowhere differentiable.

Hölder continuity with \(\alpha > 0\) implies uniform continuity, but for \(\alpha < 1\) it allows the function to oscillate in ways that preclude differentiability.

Kolmogorov's Continuity Criterion¶

The tool used to establish path regularity for stochastic processes is the following general theorem.

Proof sketch. One bounds the oscillation of \(X\) over dyadic intervals and applies the Borel–Cantelli lemma to show that the modulus of continuity decays at the required rate. The details use the Markov inequality and a chaining argument; see Karatzas & Shreve (1991), Theorem 2.2.8 for the full proof. \(\square\)

Brownian Motion is Hölder-\(\alpha\) for Every \(\alpha < \tfrac{1}{2}\)¶

Theorem (Hölder Regularity of Brownian Motion)

Brownian motion \(\{W_t\}_{t \in [0,T]}\) has a continuous modification that is almost surely Hölder continuous of order \(\alpha\) for every \(\alpha < \tfrac{1}{2}\).

Proof.

Since \(W_t - W_s \sim \mathcal{N}(0, t-s)\), the \(p\)-th moment of a Gaussian increment is:

For the \(p\)-th absolute moment of a standard normal, \(\mathbb{E}[|Z|^p] = \frac{2^{p/2}\,\Gamma((p+1)/2)}{\sqrt{\pi}} < \infty\), so set \(C_p = \mathbb{E}[|Z|^p]\):

We want to apply Kolmogorov's criterion with exponent \(1+\beta\). Matching:

For \(\beta > 0\) we need \(p > 2\). With \(p > 2\), the criterion gives Hölder continuity of every order \(\alpha < \beta/p = (p/2 - 1)/p = \tfrac{1}{2} - \tfrac{1}{p}\).

Taking \(p \to \infty\), the Hölder exponent approaches \(\tfrac{1}{2}\) from below. Since this works for every \(p > 2\), the paths are almost surely Hölder-\(\alpha\) for every \(\alpha < \tfrac{1}{2}\). \(\square\)

Nowhere Differentiability¶

Statement¶

This is a striking contrast with the Hölder continuity result: the paths are continuous everywhere but differentiable nowhere — much like the Weierstrass function, but arising naturally from probabilistic axioms.

Proof¶

We give a clean argument using the quadratic variation.

Suppose for contradiction that \(W_t(\omega)\) is differentiable at some \(t_0\), with derivative \(W'_{t_0} = L(\omega)\). Then for any \(\varepsilon > 0\) there exists \(\delta > 0\) such that

Fix a partition \(\Pi_n\) of a small interval \([t_0, t_0+\eta]\) with \(\eta < \delta\) and mesh \(\|\Pi_n\| \to 0\). Then

The first factor is bounded by \((|L|+\varepsilon)\|\Pi_n\| \to 0\). The second factor is the total variation of \(W\) over \([t_0, t_0+\eta]\), which is finite (since the increments are bounded by \((|L|+\varepsilon)\eta\)). Therefore the left side converges to 0.

But by the quadratic variation theorem (Section: Quadratic Variation of Brownian Motion), for any interval \([t_0, t_0+\eta]\):

\(L^2\) convergence implies convergence along a subsequence of partitions for a.e. \(\omega\), so for a.e. \(\omega\) there exists a subsequence along which \(\sum_i (W_{t_{i+1}} - W_{t_i})^2(\omega) \to \eta > 0\) — contradicting the deterministic bound above which forces the same sum to \(0\) for any \(\omega\) where \(W\) is differentiable at \(t_0\).

This contradiction shows that \(W_t\) cannot be differentiable at \(t_0\). Since \(t_0\) was arbitrary and the exceptional null set can be chosen uniformly over all rational \(t_0\) (using a countable union argument), the paths are almost surely differentiable nowhere. \(\square\)

Heuristic Explanation¶

The \(\sqrt{\Delta t}\) scaling of increments is the immediate intuition:

The difference quotient diverges in magnitude (though not in a fixed direction, since increments are symmetric about zero). This is why no limit exists.

flowchart LR
    A["Increment: W_{t+h} - W_t ~ N(0,h)"] --> B["Size ≈ sqrt(h)"]
    B --> C["Difference quotient ≈ 1/sqrt(h)"]
    C --> D["→ ∞ as h → 0"]
    D --> E["No derivative exists"]

Infinite Total Variation¶

Definition and Smooth Comparison¶

For a \(C^1\) function, \(V_1(f) = \int_0^T |f'(t)|\,dt < \infty\).

Why finite total variation matters. The Riemann–Stieltjes integral \(\int_0^T g(t)\,df(t)\) is well-defined for continuous \(g\) whenever \(f\) has finite total variation (the Jordan decomposition applies). Classical integration theory relies entirely on this fact.

Brownian Motion Has Infinite Total Variation¶

Proof.

For any partition \(\Pi\) of \([0,T]\), bound each squared increment by the maximum:

Rearranging:

As \(\|\Pi\| \to 0\): the numerator converges in \(L^2\) to \(T > 0\) (the quadratic variation result), while the denominator converges to 0 in probability (since \(\max_i |\Delta W_i| \leq \|\Pi\|^{1/2-\varepsilon}\) eventually, by Hölder continuity). Therefore \(V_1(W, \Pi) \to \infty\) in probability, and hence almost surely along a suitable subsequence. Since the total variation is a supremum over partitions, it equals \(+\infty\) a.s. \(\square\)

Consequence for Integration Theory¶

The failure of finite total variation means:

Specifically: the Riemann–Stieltjes construction requires approximating \(\sum_i f(t_i^*)(W_{t_{i+1}} - W_{t_i})\), and the limit fails to exist for general adapted integrands \(f\) because the accumulated oscillations of \(W\) are too large.

This motivates the construction of the Itô integral, which handles this by working in \(L^2\) and exploiting the martingale structure of Brownian motion rather than pathwise control. The key idea is that the variance of increments sums to \(T\) even though the total variation is infinite.

Summary Table¶

The last row is the decisive one for stochastic calculus: Brownian motion has finite, nonzero quadratic variation, whereas smooth functions have zero quadratic variation. This is why \(dW_t^2 = dt\) survives in Itô's formula while \((dt)^2 = 0\).

Python: Observing Path Roughness¶

The following simulation illustrates the Hölder regularity and infinite total variation by computing the \(p\)-th variation \(\sum_i |\Delta W_i|^p\) for several values of \(p\).

For a Hölder-\(\alpha\) function and a uniform partition with \(n\) steps:

• \(p > 1/\alpha\): \(p\)-variation \(\to \infty\) as \(n \to \infty\)
• \(p < 1/\alpha\): \(p\)-variation \(\to 0\)
• \(p = 1/\alpha\): \(p\)-variation converges to a finite limit

For Brownian motion (\(\alpha \approx 1/2\)), the critical \(p\) is \(1/\alpha = 2\), so total variation (\(p=1\)) diverges and quadratic variation (\(p=2\)) converges to \(T\).

```python import numpy as np import matplotlib.pyplot as plt

np.random.seed(42)

T = 1.0 p_values = [1.0, 1.5, 2.0, 2.5, 3.0] # p = 1 is total variation, p = 2 is quadratic n_values = [10, 50, 100, 500, 1000, 5000] n_paths = 200

results = {p: [] for p in p_values}

for n in n_values: dt = T / n dW = np.random.normal(0, np.sqrt(dt), size=(n_paths, n)) for p in p_values: pvar = np.mean(np.sum(np.abs(dW)**p, axis=1)) results[p].append(pvar)

fig, ax = plt.subplots(figsize=(10, 5)) for p in p_values: label = f'\(p = {p}\)' + (' (total var)' if p == 1.0 else ' (quad var)' if p == 2.0 else '') ax.semilogx(n_values, results[p], 'o-', label=label)

ax.axhline(T, color='black', linestyle='--', linewidth=1.5, label=f'\(T = {T}\) (quad var limit)') ax.set_xlabel('Number of partition steps \(n\)', fontsize=12) ax.set_ylabel('\(p\)-variation', fontsize=12) ax.set_title('\(p\)-Variation of Brownian Motion vs. Partition Size', fontsize=13) ax.legend(fontsize=10) ax.grid(alpha=0.3) plt.tight_layout() plt.savefig('figures/fig_holder_pvar.png', dpi=150, bbox_inches='tight') plt.show() ```

Interpretation:

• \(p = 1\) (total variation): grows without bound as \(n \to \infty\) — confirming infinite total variation.
• \(p = 2\) (quadratic variation): converges to \(T = 1\) — the quadratic variation theorem.
• \(p < 2\): diverges; \(p > 2\): converges to 0. The critical exponent is exactly \(p = 2\).

Summary¶

Looking ahead:

• Quadratic Variation (next section): Proves \([W]_T = T\) rigorously in \(L^2\).
• Itô's Formula (Chapter 3): The correction term \(\tfrac{1}{2}f''(W_t)\,dt\) is a direct consequence of the nowhere differentiability forcing \(dW_t^2 = dt\).

Exercises¶

1. Show that if \(f:[0,T]\to\mathbb{R}\) is Hölder-\(\alpha\) with \(\alpha > 1\), then \(f\) must be constant. (Hint: For any \(t, s\), write \(|f(t)-f(s)| \leq C|t-s|^\alpha\) and let \(s \to t\).)

1. Let \(W_t\) be Brownian motion. Use Kolmogorov's criterion with \(p = 4\) to show that \(W\) is almost surely Hölder-\(\alpha\) for \(\alpha < 1/4\). (This is a weaker bound; the optimal \(1/2\) requires \(p \to \infty\).)

1. Show that the process \(X_t = t\) has zero quadratic variation on \([0,T]\) but finite (in fact linear) total variation. Contrast with Brownian motion.

1. Let \(f\) be Hölder-\(\alpha\) with \(\alpha > 1/2\). Show that \(\sum_i (f(t_{i+1}) - f(t_i))^2 \to 0\) as the mesh \(\|\Pi\| \to 0\), confirming zero quadratic variation for regular functions.

1. Verify the infinite total variation of Brownian motion numerically: for \(T = 1\) and increasing \(n\), plot \(\sum_{i=0}^{n-1}|W_{(i+1)/n} - W_{i/n}|\) as a function of \(n\) and confirm divergence.

1. Suppose \(f:[0,T]\to\mathbb{R}\) is differentiable at \(t_0\) with \(f'(t_0) = L\). Show directly that for any partition \(\Pi\) of \([t_0, t_0+\eta]\) with small \(\eta\), \(\sum_i (f(t_{i+1}) - f(t_i))^2 \to 0\) as \(\|\Pi\| \to 0\).

References¶

• Karatzas, I., & Shreve, S. E. (1991). Brownian Motion and Stochastic Calculus, 2nd ed. Springer. (Theorem 2.2.8 for Kolmogorov criterion; Chapter 1 for path properties)
• Mörters, P., & Peres, Y. (2010). Brownian Motion. Cambridge University Press. (Chapter 1 for Hölder regularity and non-differentiability)
• Revuz, D., & Yor, M. (1999). Continuous Martingales and Brownian Motion, 3rd ed. Springer. (Chapter I for Kolmogorov criterion)
• Billingsley, P. (1995). Probability and Measure, 3rd ed. Wiley. (Weak convergence and tightness)