# Applicability of Glivenko-Cantelli theorem to functions other than cdf

Glivenko-Cantelli theorem states that approximation function $$F_n(t):=\frac{1}{n} \sum_{i=1}^{n} [X_i < t]$$ uniformly converges to actual cumulative distribution function $$F$$. That is, $$\mathrm{sup}_{t} ||F_n(t) - F(t)|| \to 0$$, a.s.

Reading through the proof of the theorem, I found that the proof uses only the three properties of $$F_n, F$$ pair:

Property 1: $$\forall t, F_n(t) \to F(t)$$, a.s.

Property 2: $$F$$ is bounded function

Property 3: $$F$$ and $$F_n$$ is monotonically increasing function

Thus, I think any function-approx-function-pair $$(f_n, f)$$ satisfies the all three properties, then Glivenko-Cantelli theorem can also be applied to that pair, and conclude that $$\mathrm{sup}_t||f_n(t) - f(t)|| \to 0$$, a.s.

Is this correct? Or, is there other important properties that $$(f, f_n)$$ must have?

### Proof of Glivenko-Cantelli

For your information, here is that proof of Glivenko-Cantelli fetched from Thm. 19.1 of [1]

*By the strong law of large numbers, both $$F_{n} \to F$$, a.s., and $$F_n(t-) \to F(t-)$$, a.s for given $$t$$. (Property 1 used). Given a fixed $$\epsilon > 0$$, there exists a partition $$-\infty = t_0 < t_1 < \cdots < t_k = \infty$$ such that $$F(t_i-)-F(t_{i-1}) < \epsilon$$ for every $$i$$. (Property 2 and 3 used) Now for $$t_{i-1}\leq t < t_i$$,

$$F_n(t) - F(t) \leq F_n(t_i -) - F(t_i -) + \epsilon$$ (Property 3 used)

$$F_n(t) - F(t) \geq F_n(t_{i-1}) - F(t_{i-1}) - \epsilon$$ (Property 3 used)

The onvergene of $$F_n(t)$$ and $$F_n(t-)$$ for every fixed $$t$$ is certainly uniform for $$t$$ in the finite set $$\{t_1,\ldots, t_n\}$$. Conclude that $$\mathrm{limsup} (\mathrm{sup}_{t}||F_n(t) - F(t)|) \leq \epsilon$$ almost surely. This is true for every $$\epsilon$$ and hence the limit superior is zero.*

[1] Van der Vaart, Aad W. Asymptotic statistics. Vol. 3. Cambridge university press, 2000.

$$F_n$$ is also bounded between 0 and 1 by construction for all $$n$$, and also attains these values for $$x<\min\{x_i\}$$ and $$x>\max\{x_i\}$$, so you would need your $$f_n$$ to behave like that also. The important thing here is the strict monotonicity of $$F$$ that implies that $$F$$ can only be discontinuous at at most countable number of points. This means that we can find a finite partition of $$[-\infty, \infty]$$ with the properties in van der Vaart's proof of Thm 19.1.. In general, one has to check that an exponential inequality is satisfied, so the Borel Cantelli lemma applies and yields uniform convergence. See Pollard, Convergence of Stochastic Processes, (especially p. 13-16), and the book of Van der Vaart and Wellner, Weak Convergence and Empirical Processes with Applications to Statistics for uniform convergence of empirical processes.
I think you need slightly more than this, just to control the behaviour of $$F_n$$ at $$\pm \infty$$; the proof also uses that $$F_n(\pm \infty) \to F(\pm \infty)$$, which for CDFs is trivial. Given that, yes, the same result holds for any monotone $$F_n$$ converging pointwise to a bounded monotone $$F$$. You don't even need a probability space, it's just a result about convergence of functions.