# Why is $k = \sqrt{N}$ a good solution of the number of neighbors to consider?

In $$k$$-NN it is often stated that a good starting number of neighbors to select is $$k = \sqrt{N}$$ , where $$N$$ is the total number of points. But why is this so?

Examples:

• It might help here to include the source(s) where you have read this. Because that may clarify whether setting $k = \sqrt{N}$ is merely an off-the-shelf programming hack/heuristic, or if it is grounded in theoretical principles (e.g. asymptotics or finite sample results). There are some fairly well known references on the latter, if that is what you are soliciting? Jul 18, 2021 at 17:27
• $N^{1/d}$ would be my choice, where $d$ is number of dimensions Jul 19, 2021 at 1:49
• Hi @microhaus, good point. I've added 3 examples. From my understanding, this seems like a ad-hoc starting value, but I'd still like to know why.
– Sos
Jul 19, 2021 at 10:24
• +1 for adding sources, it happens to confirm that the recommendation on $k$ was made as a rule of thumb. You've stated that you would like to know why the heuristic makes sense. The theorem and references I listed in the answer below, in my view, supply some insight on why that heuristic makes sense, as well as its limitations. Whilst the result has a standard statistical learning theory flavour to it, it might use quantities which seem unclear or specialised. If that is the case, do say and I can unpack it using more conventional probabilistic language. Jul 19, 2021 at 17:37

## 1 Answer

There are a number of quantitative finite-sample results, and also asymptotic arguments, in support of using the heuristic $$k = \sqrt{n}$$, where $$n$$ is the sample size. However, in practice, it would seem that this heuristic really should only be a starting point for selecting $$k$$ using data-dependent methods.

Theoretical arguments for the use of such a heuristic.

Adapting from Devroye, Györfi and Lugosi (1996), theoretical performance of the $$k$$-nearest neighbour classifier can be organised, albeit not exclusively, along the following lines:

• $$k$$ is determined to be a finite, fixed constant, whilst the sample size $$n \rightarrow \infty$$. The determination of $$k$$ is a priori, i.e. selected in advance. That is, using prior knowledge, after exploratory data analysis, or using a heuristic.

• $$k \rightarrow \infty$$ whilst $$k / n \rightarrow 0$$. Similar to above in a priori determination of $$k$$, but now $$k$$ is not fixed relative to sample size $$n$$.

• Data dependent methods for determining $$k$$ e.g. using a training set, test set and selecting $$k$$ to minimise the estimated classification error rate, or using cross-validation.

The heuristic $$k = \lfloor \sqrt{n} \rfloor$$, where $$\lfloor \cdot \rfloor$$ is the floor function, would fall under the second category. From the above book, the following is a quantitative, finite-sample probabilistic bound on the excess risk $$L_n - L^*$$, which in turn implies an asymptotic result:

Theorem 11.1. (Devroye and Györfi (1985), Zhao (1987)). Assume that each $$\mu$$ has a density. If $$k \rightarrow \infty$$ and $$k / n \rightarrow 0$$ then for every $$\epsilon > 0$$ there is an $$n_0$$ such that for $$n > n_0$$,

$$P \left( L_n - L^* > \epsilon \right) \leq 2e^{-n \epsilon^2 / (72 \gamma_d^2)},$$

where the $$\gamma_d$$ is the minimal number of cones centered at the origin of angle $$\pi / 6$$ that cover $$\mathbb{R}^d$$. (For the definition of a cone, see Chapter 5). Thus, the $$k$$-NN rule is strongly consistent.

Supplying context on the terms not defined in the extract, $$L_n = L_n({g}_n) = P({g}_n(X) \neq Y)$$ is the risk of the $$k$$-nearest neighbour classifier $${g}_n(X)$$, where $${g}_n$$ is estimated from a sample of size $$n$$. $$L^* = L(g^*) = \inf_{g \in \mathcal{G}} P(g(X) \neq Y)$$, is the Bayes-optimal classification risk, or Bayes error rate, that is, the risk of the Bayes classifier $$g^*$$. Glossing over the measure theoretic technicalities, that the measure $$\mu$$ has a density just means that $$X$$ has a density.

Parsing the theorem, the main condition requires that $$k \rightarrow \infty$$ as the sample size $$n \rightarrow \infty$$ in such a way that $$k / n \rightarrow 0$$. Your heuristic satisifies this condition because $$k = \lfloor \sqrt{n} \rfloor \rightarrow \infty$$ and $$k / n = \lfloor \sqrt{n} \rfloor / n \approx (1 / \sqrt{n}) \rightarrow 0$$. Taking the theorem as an asymptotic result, the $$k$$-nearest neighbour classifier is strongly consistent, in the sense of

$$L_n \overset{a.s.}{\longrightarrow} L^* \iff P\left( \lim_{n \rightarrow \infty} L_n = L^* \right) = 1.$$

That is, as you collect more observations $$n \rightarrow \infty$$, the classification error rate $$L_n$$ of the $$k$$-nearest neighbour classifier $$g_n$$ will converge almost surely to the minimal classification error rate you can possibly hope to achieve, $$L^*$$. And that this converges exponentially quickly.

Furthermore, the result is non-asymptotic in that for finite $$n$$, it bounds the probability that $$L_n$$ deviates from $$L^*$$ by more than $$\epsilon$$ in terms of finite constants.

On the use of data-dependent methods for selecting $$k$$.

The utility of the above theoretical result then is that it supplies insight on heuristics like the one you have outlined. Its limitations, like many results in statistical learning theory, is that the constant $$\gamma_d$$ may be difficult to compute, or in the case that it is computable, renders the bound too loose to give any practical prescriptions.

Echoing the sentiment expressed in the following linked question, the authors advocate the use of data-dependent means of selecting $$k$$ in practice:

Consistency by itself may be obtained by choosing $$k = \lfloor \sqrt{n} \rfloor$$, but few --if any-- users will want to blindly use such recipes. Instead, a healthy dose of feedback from the data is preferable.

Similar consistency results for the use of a test set to select $$k$$ based on minimising a holdout estimate of the classification error rate are supplied therein.

There seems to be some consensus that the kind of result listed above is a continuation of a line of work in the spirit of Stone (1977). A more recent, specialised treatment is by Chaudhuri and Dasgupta (2014). Further details can be found in the references below.

References.

Devroye, L., Györfi, L., & Lugosi, G. (1996). A Probabilistic Theory of Pattern Recognition. Springer. https://doi.org/10.1007/978-1-4612-0711-5. See chapters 5, 6, 11, 26.

Stone, C. J. (1977). Consistent nonparametric regression. The Annals of Statistics, 5(4) 595 - 645. https://doi.org/10.1214/aos/1176343886

Chaudhuri, K., & Dasgupta, S. (2014). Rates of convergence for nearest neighbour classification. Advances in Neural Information Processing Systems 27, NIPS 2014. https://papers.nips.cc/paper/2014/hash/db957c626a8cd7a27231adfbf51e20eb-Abstract.html

• Many thanks @microhaus, this is very useful. I will try to go through these, though I can see it will not be straightforward given my limited maths :(
– Sos
Jul 20, 2021 at 12:53
• @Sos. I've edited the answer to parse the theorem in less loaded language. I have included the references for completeness. As a note from my own experiences with the book by Devroye et al., it is somewhat seen as a "bible" in statistical learning theory. It is therefore uncompromising in its demands on mathematical literacy if one is looking for thorough understanding of its results and proofs. IMO, that comprises fluency in real analysis, as well as results in Lebesgue integration and a smattering of measure theory [...] Jul 20, 2021 at 16:18
• @Sos. [...]. As those areas require non-trivial amounts of effort to develop literacy in, and because the results are theoretical, it may be unnecessary if you are only interested in applying k-nearest neigbours in a practical setting. Jul 20, 2021 at 16:26