Rust Guide: Evaluating multilinear extensions via Lagrange interpolation

This is an unofficial & unaudited Rust implementation of algorithms in Dr. Thaler’s Proofs, Arguments, & ZK book.

This week we learned why multilinear extensions are useful for interactive proofs.

Source: Lemma 3.6 in

In short, and in egregious laymen terms, multilinear extensions give us the ability to both (1) distance amplify a polynomial f , into f~, which can be evaluated by the Verifier and (2) still keep the new f~ as a low degree polynomial.

(1) ensures that prover can’t convincingly lie about knowing f , by just providing random evaluations on some other bad f’

(2) ensures efficient runtimes for both Prover and Verifier operations


Let’s implement the Lagrange interpolation of multilinear polynomials 3 different ways.

To follow the notation in the following tutorial, please reference starting lemma 3.6 in Dr. Thaler’s book linked here.

The full implementation with tests & benchmarking can be found here:

Necessary caveat: I’m neither a mathematician nor a Rust expert. Suggestions for improvement are always welcome! 🥰

Method 1: slow

In a rote approach, we stream over the evaluation table f(w) and sum up the resulting computation on each evaluation element.

Within the computation for each evaluation, we first retrieve the correct notation of w in {0,1}^v. For example, the following gives us 8->[1,0,0,0]

Then we compute its χ by iterating over the vector w and r (verifier input), calculating each step of χ:(xiwi +(1−xi)(1−wi)) at all vector positionsi.

This implementation is O(v* 2^v). When benchmarked using a function f mapping {0, 1}^2 to the field F5, we get a median runtime of 1,402 ns per iteration.

Method 2: stream (lemma 3.7)

Lemma 3.7 presents a more space efficient way of streaming over f(w), which can be achieved by making slow_mle recursive.

Source: Proof of Lemma 3.7 in

This is implemented as:

And to invoke this, we simply do:

This implementation is also O(v* 2^v) runtime but with a space improvement at O(logn).

When benchmarked using a function f mapping {0, 1}^2 to the field F5, we get a median runtime of 1,312 ns per iteration.

Method 3: dynamic (lemma 3.8)

Lemma 3.8 memoizes the computations χ and provides it as a look up table later used in the (figure 3.1) summation step.

The intuition is that, for example, the χ for w=[0,0,0] & w=[0,0,1]both involve calculating χ for w=[0,0].

In fact, the χ for w=[0,0,0] is just χ for w=[00] times (1-r_i).

A modified visual of figure 3.3 in

So we can memoize and avoid recomputing intermediate χ values as we build the lookup table.

Here, we use recursion to achieve memoization.

Finally, we invoke the function to build the χ look up table and compute the inner dot product of f~ (figure 3.1).

This implementation only has a linear runtime of O(2^v), albeit with a space utilization of O(n).

Benchmarked against methods 1 and 2, method 3 yields a huge runtime improvement (for v=2 on a field F5).

For v=2, Fp = 5

To quote the book, the main takeaway from all of this is really that:

Multivariate polynomials with ultra-low degree in each variable turn out to be especially useful when designing interactive proofs with small communication and fast verification. [Thaler, pg. 28]


This is a completely unrelated photo that I just like.



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store