Twisted convolution and the sensitivity conjecture
What's new 2019-08-20
Earlier this month, Hao Huang (who, incidentally, was a graduate student here at UCLA) gave a remarkably short proof of a long-standing problem in theoretical computer science known as the sensitivity conjecture. See for instance this blog post of Gil Kalai for further discussion and links to many other online discussions of this result. One formulation of the theorem proved is as follows. Define the -dimensional hypercube graph to be the graph with vertex set , and with every vertex joined to the vertices , where is the standard basis of .
Theorem 1 (Lower bound on maximum degree of induced subgraphs of hypercube) Let be a set of at least vertices in . Then there is a vertex in that is adjacent (in ) to at least other vertices in .
The bound (or more precisely, ) is completely sharp, as shown by Chung, Furedi, Graham, and Seymour; we describe this example below the fold. When combined with earlier reductions of Gotsman-Linial and Nisan-Szegedy; we give these below the fold also.
Let be the adjacency matrix of (where we index the rows and columns directly by the vertices in , rather than selecting some enumeration ), thus when for some , and otherwise. The above theorem then asserts that if is a set of at least vertices, then the minor of has a row (or column) that contains at least non-zero entries.
The key step to prove this theorem is the construction of rather curious variant of the adjacency matrix :
Proposition 2 There exists a matrix which is entrywise dominated by in the sense that
Assuming this proposition, the proof of Theorem 1 can now be quickly concluded. If we view as a linear operator on the -dimensional space of functions of , then by hypothesis this space has a -dimensional subspace on which acts by multiplication by . If is a set of at least vertices in , then the space of functions on has codimension at most in , and hence intersects non-trivially. Thus the minor of also has as an eigenvalue (this can also be derived from the Cauchy interlacing inequalities), and in particular this minor has operator norm at least . By Schur’s test, this implies that one of the rows or columns of this matrix has absolute values summing to at least , giving the claim.
Remark 3 The argument actually gives a strengthening of Theorem 1: there exists a vertex of with the property that for every natural number , there are at least paths of length in the restriction of to that start from . Indeed, if we let be an eigenfunction of on , and let be a vertex in that maximises the value of , then for any we have that the component of is equal to ; on the other hand, by the triangle inequality, this component is at most times the number of length paths in starting from , giving the claim.
This argument can be viewed as an instance of a more general “interlacing method” to try to control the behaviour of a graph on all large subsets by first generating a matrix on with very good spectral properties, which are then partially inherited by the minor of by interlacing inequalities. In previous literature using this method (see e.g., this survey of Haemers, or this paper of Wilson), either the original adjacency matrix , or some non-negatively weighted version of that matrix, was used as the controlling matrix ; the novelty here is the use of signed controlling matrices. It will be interesting to see what further variants and applications of this method emerge in the near future. (Thanks to Anurag Bishoi in the comments for these references.)
The “magic” step in the above argument is constructing . In Huang’s paper, is constructed recursively in the dimension in a rather simple but mysterious fashion. Very recently, Roman Karasev gave an interpretation of this matrix in terms of the exterior algebra on . In this post I would like to give an alternate interpretation in terms of the operation of twisted convolution, which originated in the theory of the Heisenberg group in quantum mechanics.
Firstly note that the original adjacency matrix , when viewed as a linear operator on , is a convolution operator
where
is the counting measure on the standard basis , and denotes the ordinary convolution operation
As is well known, this operation is commutative and associative. Thus for instance the square of the adjacency operator is also a convolution operator
where the convolution kernel is moderately complicated:
The factor in this expansion comes from combining the two terms and , which both evaluate to .
More generally, given any bilinear form , one can define the twisted convolution
of two functions . This operation is no longer commutative (unless is symmetric). However, it remains associative; indeed, one can easily compute that
In particular, if we define the twisted convolution operator
then the square is also a twisted convolution operator
and the twisted convolution kernel can be computed as
For general bilinear forms , this twisted convolution is just as messy as is. But if we take the specific bilinear form
then for and for , and the above twisted convolution simplifies to
and now is very simple:
Thus the only eigenvalues of are and . The matrix is entrywise dominated by in the sense of (1), and in particular has trace zero; thus the and eigenvalues must occur with equal multiplicity, so in particular the eigenvalue occurs with multiplicity since the matrix has dimensions . This establishes Proposition 2.
Remark 4 Twisted convolution is actually just a component of ordinary convolution, but not on the original group ; instead it relates to convolution on a Heisenberg group extension of this group. More specifically, define the Heisenberg group to be the set of pairs with group law
and inverse operation
(one can dispense with the negative signs here if desired, since we are in characteristic two). Convolution on is defined in the usual manner: one has
for any . Now if is a function on the original group , we can define the lift by the formula
and then by chasing all the definitions one soon verifies that
for any , thus relating twisted convolution to Heisenberg group convolution .
Remark 5 With the twisting by the specific bilinear form given by (2), convolution by and now anticommute rather than commute. This makes the twisted convolution algebra isomorphic to a Clifford algebra (the real or complex algebra generated by formal generators subject to the relations for ) rather than the commutative algebra more familiar to abelian Fourier analysis. This connection to Clifford algebra (also observed independently by Tom Mrowka and by Daniel Matthews) may be linked to the exterior algebra interpretation of the argument in the recent preprint of Karasev mentioned above.
Remark 6 One could replace the form (2) in this argument by any other bilinear form that obeyed the relations and for . However, this additional level of generality does not add much; any such will differ from by an antisymmetric form (so that for all , which in characteristic two implied that for all ), and such forms can always be decomposed as , where . As such, the matrices and are conjugate, with the conjugation operator being the diagonal matrix with entries at each vertex .
Remark 7 (Added later) This remark combines the two previous remarks. One can view any of the matrices in Remark 6 as components of a single canonical matrix that is still of dimensions , but takes values in the Clifford algebra from Remark 5; with this “universal algebra” perspective, one no longer needs to make any arbitrary choices of form . More precisely, let denote the vector space of functions from the hypercube to the Clifford algebra; as a real vector space, this is a dimensional space, isomorphic to the direct sum of copies of , as the Clifford algebra is itself dimensional. One can then define a canonical Clifford adjacency operator on this space by
where are the generators of . This operator can either be identified with a Clifford-valued matrix or as a real-valued matrix. In either case one still has the key algebraic relations and , ensuring that when viewed as a real matrix, half of the eigenvalues are equal to and half equal to . One can then use this matrix in place of any of the to establish Theorem 1 (noting that Schur’s test continues to work for Clifford-valued matrices because of the norm structure on ).
To relate to the real matrices , first observe that each point in the hypercube can be associated with a one-dimensional real subspace (i.e., a line) in the Clifford algebra by the formula
for any (note that this definition is well-defined even if the are out of order or contain repetitions). This can be viewed as a discrete line bundle over the hypercube. Since for any , we see that the -dimensional real linear subspace of of sections of this bundle, that is to say the space of functions such that for all , is an invariant subspace of . (Indeed, using the left-action of the Clifford algebra on , which commutes with , one can naturally identify with , with the left action of acting purely on the first factor and acting purely on the second factor.) Any trivialisation of this line bundle lets us interpret the restriction of to as a real matrix. In particular, given one of the bilinear forms from Remark 6, we can identify with by identifying any real function with the lift defined by
whenever . A somewhat tedious computation using the properties of then eventually gives the intertwining identity
and so is conjugate to .
— 1. The Chung, Furedi, Graham, and Seymour example —
The paper of by Chung, Furedi, Graham, and Seymour gives, for any , an example of a subset of of cardinality for which the maximum degree of restricted to is at most , thus showing that Theorem 1 cannot be improved (beyond the trivial improvement of upgrading to , because the maximum degree is obviously a natural number).
Define the “Möbius function” to be the function
for . This function is extremely balanced on coordinate spaces. Indeed, from the binomial theorem (which uses the convention ) we have
More generally, given any index set of cardinality , we have
Now let be a partition of into disjoint non-empty sets. For each , let be the subspace of consisting of those such that for all . Then for any , we have
and the right-hand side vanishes if and equals when . Applying the inclusion-exclusion principle, we conclude that
and thus also (assuming )
so that
Thus, if denotes the set of those with , together with those with , then has to have two more elements than its complement , and hence has cardinality .
Now observe that, if with and , then , and if then unless . Thus in this case the total number of for which is at most . Conversely, if with and , then , and for each there is at most one that will make lie in . Hence in this case the total number of for which is at most . Thus the maximum degree of the subgraph of induced by is at most . By choosing the to be a partition of into pieces, each of cardinality at most , we obtain the claim.
Remark 8 Suppose that is a perfect square, then the lower bound here exactly matches the upper bound in Theorem 1. In particular, the minor of the matrix must have an eigenvector of eigenvalue . Such an eigenvector can be explicitly constructed as follows. Let be the vector defined by setting
for some , , , and for all other (one easily verifies that the previous types of lie in ). We claim that
for all . Expanding out the left-hand side, we wish to show that
First suppose that is of the form (3). One checks that lies in precisely when for one of the , in which case
Since , this simplifies using (3) as
giving (5) in this case. Similarly, if is of the form (4), then lies in precisely when , in which case one can argue as before to show that
and (5) again follows. Finally, if is not of either of the two forms (3), (4), one can check that is never of these forms either, and so both sides of (5) vanish.
The same analysis works for any of the other bilinear forms in Remark 6. Using the Clifford-valued operator from Remark 7, the eigenfunction is cleaner; it is defined by
when is of the form (3), and
when is of the form (4), with otherwise.
— 2. From induced subgraph bounds to the sensitivity conjecture —
On the hypercube , let denote the functions
The monomials in are then the characters of , so by Fourier expansion every function can be viewed as a polynomial in the (with each monomial containing at most one copy of ; higher powers of each are unnecessary since . In particular, one can meaningfully talk about the degree of a function . Observe also that the Möbius function from the preceding section is just the monomial .
Define the sensitivity of a Boolean function to be the largest number for which there is an such that there are at least values of with . Using an argument of Gotsman and Linial, we can now relate the sensitivity of a function to its degree:
Corollary 9 (Lower bound on sensitivity) For any boolean function , one has .
Proof: Write . By permuting the indices, we may assume that contains a non-trivial multiple of the monomial . By restricting to the subspace (which cannot increase the sensitivity), we may then assume without loss of generality that . The Fourier coefficient of is just the mean value
of times the Möbius function , so this mean value is non-zero. This means that one of the sets or has cardinality at least . Let denote the larger of these two sets. By Theorem 1, there is an such that for at least values of ; since , this implies that for at least values of , giving the claim.
The construction of Chung, Furedi, Graham, and Seymour from the previous section can be easily adapted to show that this lower bound is tight (other than the trivial improvement of replacing by ).
Now we need to digress on some bounds involving polynomials of one variable. We begin with an inequality of Bernstein concerning trigonometric polynomials:
Lemma 10 (Bernstein inequality) Let be a trigonometric polynomial of degree at most , that is to say a complex linear combination of for . Then
Observe that equality holds when or . Specialising to linear combinations of , we obtain the classical Bernstein inequality
for complex polynomials of degree at most .
Proof: If one is willing to lose a constant factor in this estimate, this bound can be easily established from modern Littlewood-Paley theory (see e.g., Exercise 52 of these lecture notes). Here we use an interlacing argument due to Boas. We first restrict to the case when has real coefficients. We may normalise . Let be a real parameter in . The trigonometric polynomial alternately takes the values and at the values . Thus the trigonometric polynomial alternates in sign at these values, and thus by the intermediate value theorem has a zero on each of the intervals . On the other hand, a trigonometric polynomial of degree at most can be expressed by de Moivre’s theorem as times a complex polynomial in of degree at most , and thus has at most zeroes. Thus we see that has exactly one zero in each . Furthermore, at this zero, the derivative of this function must be positive if is increasing on this interval, and negative if is decreasing on this interval. In summary, we have shown that if and are such that , then has the same sign as . By translating the function , we also conclude that if and are such that for some , then has the same sign as .
If , then we can find such that and is positive, and we conclude that ; thus we have the upper bound
A similar argument (with now chosen to be negative) similarly bounds . This gives the claim for real-valued trigonometric polynomials . (Indeed, this argument even gives the slightly sharper bound .)
To amplify this to complex valued polynomials, we take advantage of phase rotation invariance. If is a complex trigonometric polynomial, then by applying Bernstein’s inequality to the real part we have
But then we can multiply by any complex phase and conclude that
Taking suprema in , one obtains the claim for complex polynomials .
The analogue of Bernstein’s inequality for the unit interval is known as Markov’s inequality for polynomials:
Lemma 11 (Markov’s inequality for polynomials) Let be a polynomial of degree . Then
This bound is sharp, as is seen by inspecting the Chebyshev polynomial , defined as the unique polynomial giving the trigonometric identity
Differentiating (6) using the chain rule, we see that
the right-hand side approaches as , demonstrating that the factor here is sharp.
Proof: We again use an argument of Boas. We may normalise so that
The function is a trigonometric polynomial of degree at most , so by Bernstein’s inequality and the chain rule we have
for all . This already gives Markov’s inequality except in the edge regions (since ). By reflection symmetry, it then suffices to verify Markov’s inequality in the region .
From (6), the Chebyshev polynomial attains the values alternately at the different points . Thus, if , the polynomial changes sign at least times on , and thus must have all zeroes inside this interval by the intermediate value theorem; furthermore, of these zeroes will lie to the left of . By Rolle’s theorem, the derivative then has all zeroes in the interval , and at least of these will lie to the left of . In particular, the derivative can have at most one zero to once to the right of .
Since , is positive at , and hence positive as since there are no zeroes outside of . Thus the leading coefficient of is positive, which implies the same for its derivative . Thus is positive when .
From (9) one has , hence by (7) we see that is also positive at . Thus cannot become negative for , as this would create at least two zeroes to the right of . We conclude that in this region we have
From (7) we have , and the claim follows.
Remark 12 The following slightly shorter argument gives the slightly weaker bound . We again use the normalisation (8). By two applications of Bernstein’s inequality, the function has first derivative bounded in magnitude by , and second derivative bounded in magnitude by . As this function also has vanishing first derivative at , we conclude the bounds
and thus by the chain rule
For , one easily checks that the right-hand side is at most , giving the claim.
This implies a result of Ehlich-Zeller and of Rivlin-Cheney:
Corollary 13 (Discretised Markov inequality) Let be a polynomial of degree . If
Proof: We use an argument of Nisan and Szegedy. Assume for sake of contradiction that , so in particular . From the fundamental theorem of calculus and the triangle inequality one has
By a rescaled and translated version of Markov’s inequality we have
which when inserted into the preceding inequality gives after some rearranging
and then after a second application of (11) gives
Comparing with (10), we conclude that
and the claim follows after some rearranging.
Nisan and Szegedy observed that this one-dimensional degree bound can be lifted to the hypercube by a symmetrisation argument:
Corollary 14 (Multidimensional Markov inequality bound) Let be such that and for . Then
Proof: By averaging over all permutations of the indices (which can decrease the degree of , but not increase it), we may assume that is a symmetric function of the inputs . Using the Newton identities, we can then write
for some real polynomial of degree at most , where
is the Hamming length of . By hypothesis, , , and , hence by the mean-value theorem . Applying Corollary 13 with , we obtain the claim.
Define the block sensitivity of a Boolean function to be the largest number for which there is an such that there are at least disjoint subsets of with for . We have
More precisely, the sensitivity conjecture of Nisan and Szegedy asserted the bound ; Huang’s result thus gives explicit values for the exponents. It is still open whether the exponent in this theorem can be improved; it is known that one cannot improve it to below , by analysing a variant of the Chung-Furedi-Graham-Seymour example (see these notes of Regan for details). Proof: The lower bound for is immediate from the definitions, since the sensitivity arises by restricting the in the definition of block sensitivity to singleton sets. To prove the upper bound, it suffices from Proposition 9 to establish the bound
Let . By hypothesis, there are and disjoint subsets of such that for . We may normalise , and . If we then define the pullback boolean function by the formula
then it is easy to see that , , and for . The claim now follows from Corollary 14.
Remark 16 The following slightly shorter variant of the argument lets one remove the factor . Let be as above. We again may normalise and . For any , let be iid Bernoulli random variable that equal with probability and with probability . The quantity
is a trigonometric polynomial of degree at most that is bounded in magnitude by , so by two applications of Bernstein’s inequality
On the other hand, for small , the random variable is equal to zero with probability and equal to each with probability , hence
and hence . Combining these estimates we obtain and hence .
Remark 17 The sensitivity of a Boolean function can be split as , where is largest number for which there is an such that there are at least values of with . It is not difficult to use the case of Remark 3 to improve Corollary 9 slightly to . Combining this with the previous remark, we can thus improve the upper bound in Theorem 15 slightly to