Geometric rank of tensors and subrank of matrix multiplication
GGeometric Rank of Tensors andSubrank of Matrix Multiplication
Swastik Kopparty , Guy Moshkovitz and Jeroen Zuiddam Abstract
Motivated by problems in algebraic complexity theory (e.g., matrix multiplication) andextremal combinatorics (e.g., the cap set problem and the sunflower problem), we introducethe geometric rank as a new tool in the study of tensors and hypergraphs. We prove that thegeometric rank is an upper bound on the subrank of tensors and the independence number ofhypergraphs. We prove that the geometric rank is smaller than the slice rank of Tao, and relategeometric rank to the analytic rank of Gowers and Wolf in an asymptotic fashion. As a firstapplication, we use geometric rank to prove a tight upper bound on the (border) subrank of thematrix multiplication tensors, matching Strassen’s well-known lower bound from 1987.
Tensors play a central role in computer science and mathematics. Motivated by problems in algebraiccomplexity theory (e.g., the arithmetic complexity of matrix multiplication), extremal combinatorics(e.g., the cap set problem and the Erd˝os–Szemer´edi sunflower problem) and quantum informationtheory (the resource theory of quantum entanglement), we introduce and study a new tensorparameter called geometric rank . Like the many widely studied notions of rank for tensors (rank,subrank, border rank, border subrank, flattening rank, slice rank, analytic rank), geometric rank oftensors generalizes the classical rank of matrices. In this paper, we: • prove a number of basic properties and invariances of geometric rank, • develop several tools to reason about, and sometimes exactly compute, the geometric rank, • show intimate connections between geometric rank and the other important notions of rankfor tensors, • and as a simple application of the above, we answer an old question of Strassen by showingthat the (border) subrank of m × m matrix multiplication is at most (cid:100) m / (cid:101) (this is tight forborder subrank; previously the border subrank of the matrix multiplication tensor was knownto lie between m and (1 − o (1)) m ). Rutgers University, [email protected] Institute for Advanced Study and DIMACS (Rutgers University), [email protected] Institute for Advanced Study, [email protected] a r X i v : . [ c s . CC ] J u l ore generally, we believe that geometric rank provides an interesting new route to prove upperbounds on subrank of tensors (and hence independence numbers of hypergraphs). Such upperbounds are important in complexity theory in the context of matrix multiplication and barriers tomatrix multiplication, and combinatorics in the context of specific natural hypergraphs (as in thecap set problem and the Erd˝os–Szemeredi sunflower problem). We define the geometric rank of a tensor as the codimension of the (possibly reducible) algebraicvariety defined by the bilinear forms given by the slices of the tensor. Here we use the standardnotions of dimension and codimension of affine varieties from algebraic geometry. That is, for anytensor T = ( T i,j,k ) i,j,k ∈ F n × n × n with coefficients T i,j,k in an algebraically closed field F (e.g., thecomplex numbers C ) and with 3-slices M k = ( T i,j,k ) i,j ∈ F n × n we define the geometric rank GR( T )as GR( T ) = codim { ( x, y ) ∈ F n × F n | x T M y = · · · = x T M n y = 0 } . Viewing T as the trilinear map T : F n × F n × F n → F : ( x, y, z ) (cid:55)→ (cid:80) i,j,k T i,j,k x i y j z k , we canequivalently write the geometric rank of T asGR( T ) = codim { ( x, y ) ∈ F n × F n | ∀ z ∈ F n : T ( x, y, z ) = 0 } . The definition of geometric rank is expressed asymmetrically in x , y and z . We will see, however, thatthe codimensions of { ( x, y ) ∈ F n × F n | ∀ z : T ( x, y, z ) = 0 } , { ( x, z ) ∈ F n × F n | ∀ y : T ( x, y, z ) = 0 } and { ( y, z ) ∈ F n × F n | ∀ x : T ( x, y, z ) = 0 } coincide (Theorem 3.2).The motivation for this definition is a bit hard to explain right away. We arrived at it whilesearching for a characteristic 0 analogue of the analytic rank of Gowers and Wolf [GW11] (seeSection 8). Example 1.
We give an example of how to compute the geometric rank. Let T ∈ F × × be thetensor with 3-slices M = (cid:18) (cid:19) , M = (cid:18) (cid:19) . (This is sometimes called the W -tensor). One verifies that the algebraic variety V = { ( x, y ) ∈ F × F | x y = 0 , x y + x y = 0 } has the three irreducible components { ( x, y ) ∈ F × F | x = 0 , x = 0 } , { ( x, y ) ∈ F × F | x = 0 , y = 0 } and { ( x, y ) ∈ F × F | y = 0 , y = 0 } . Each irreduciblecomponent has dimension 2 and thus V has dimension 2. Hence GR( T ) = codim V = 4 − Before discussing our results we give an introduction to some of the existing notions of rank andtheir usefulness. Several interesting notions of rank of tensors have been studied in mathematicsand computer science, each with their own applications. As a warm-up we first discuss the familiarsituation for matrices. 2 atrices.
For any two matrices M ∈ F m × m and N ∈ F n × n we write M ≤ N if there existmatrices A, B such that M = AN B . Defining the matrix rank R( M ) of M as the smallest number r such that M can be written as a sum of r matrices that are outer products ( u i v j ) ij (i.e., rank-1matrices), we see that in terms of the relation ≤ we can write the matrix rank as the minimisationR( M ) = min { r ∈ N | M ≤ I r } , where I r is the r × r identity matrix. Matrix rank thus measures the “cost” of M in terms of identitymatrices. Let us define the subrank Q( M ) of M as the “value” of M in terms of identity matrices,Q( M ) = max { s ∈ N | I s ≤ M } . It turns out that subrank equals rank for matrices,Q( M ) = R( M ) . Namely, if R( M ) = r , then by using Gaussian elimination we can bring M in diagonal formwith exactly r nonzero elements on the diagonal, and so I r ≤ M . In fact, M ≤ N if and onlyif R( M ) ≤ R( N ). Tensors.
For any two tensors S ∈ F m × m × m and T ∈ F n × n × n we write S ≤ T if there are ma-trices A, B, C such that S = ( A, B, C ) · T where we define ( A, B, C ) · T := ( (cid:80) a,b,c A ia B jb C kc T a,b,c ) i,j,k .Thus ( A, B, C ) · T denotes taking linear combinations of the slices of T in three directions accord-ing to A , B and C . Let T ∈ F n × n × n be a tensor. The tensor rank R( T ) of T is defined asthe smallest number r such that T can be written as a sum of r tensors that are outer prod-ucts ( u i v j w k ) i,j,k . Similarly as for matrices, we can write tensor rank in terms of the relation ≤ asthe “cost” minimisationR( T ) = min { r ∈ N | M ≤ I r } where I r is the r × r × r identity tensor (i.e., the diagonal tensor with ones on the main diagonal).Strassen defined the subrank of T as the “value” of T in terms of identity tensors,Q( T ) = max { s ∈ N | I s ≤ M } . Naturally, since ≤ is transitive, we have that value is at most cost: Q( T ) ≤ R( T ). Unlike thesituation for matrices, however, there exist tensors for which this inequality is strict. One wayto see this is using the fact that a random tensor in F n × n × n has tensor rank close to n whereasits subrank is at most n . Another way to see this is using the ranks R ( i ) ( T ) := R( T ( i ) ) of thematrices T (1) = ( T i,j,k ) i, ( j,k ) ∈ F n × n n , T (2) = ( T i,j,k ) j, ( i,k ) ∈ F n × n n , and T (3) = ( T i,j,k ) k, ( i,j ) ∈ F n × n n obtained from T by grouping two of the three indices together, sinceQ( T ) ≤ R ( i ) ( T ) ≤ R( T ) . Namely, it is not hard to find tensors T for which R (1) ( T ) < R (2) ( T ). We will now discuss two upperbounds on the subrank Q( T ) that improve on the flattening ranks R ( i ) ( T ). Then we will discussconnections between subrank and problems in complexity theory and combinatorics.3 lice rank. In the context of the cap set problem, Tao [Tao16] defined the slice rank of anytensor T as the minimum number r such that T can be written as a sum of r tensors of the form( u i V jk ) i,j,k , ( u j V ik ) i,j,k or ( u k V ij ) i,j,k (i.e., an outer product of a vector and a matrix). In otherwords, SR( T ) := min { R (1) ( S ) + R (2) ( S ) + R (3) ( S ) : S + S + S = T } . Clearly slice rank is atmost any flattening rank, and Tao proved that slice rank upper bounds subrank,Q( T ) ≤ SR( T ) ≤ R ( i ) ( T ) . The lower bound connects slice rank to problems in extremal combinatorics, which we will discussfurther in Section 1.3. The slice rank of large Kronecker powers of tensors was studied in [BCC + Analytic rank.
Gowers and Wolf [GW11] defined the analytic rank of any tensor T ∈ F n × n × n p over the finite field F p for a prime p as AR( T ) := − log p bias( T ), where the bias of T is definedas bias( T ) := E exp(2 πi T ( x, y, z ) /p ) with the expectation taken over all vectors x ∈ F n p , y ∈ F n p and z ∈ F n p . The analytic rank relates to subrank and tensor rank as follows:Q( T ) ≤ AR( T )AR( I ) ≤ R( T )where AR( I ) = − log p (1 − (1 − /p ) ). The upper bound was proven in [BHH + T ) / AR( I ) can be larger than max i R ( i ) ( T ) for small p . The lower bound is essentiallyby Lovett [Lov19]. Namely, Lovett proves that AR( T ) / AR( I ) upper bounds the size of the largestprincipal subtensor of T that is diagonal. (We will discuss this further in Section 1.3.) Lovettmoreover proved that AR( T ) ≤ SR( T ) and he thus proposes analytic rank as an effective upperbound tool for any type of problem where slice rank works well asymptotically. Lovett’s resultmotivated us to study other parameters to upper bound the subrank, which led to geometric rank.Another line of work has shown upper bounds on SR( T ) in terms of AR( T ). This was firstproven by Bhowmick and Lovett [BL15], with an Ackerman-type dependence. The dependence waslater improved significantly by Janzer [Jan18]. Recently, Janzer [Jan20] and Mili´cevi´c [Mil19] provedpolynomial upper bounds of SR in terms of AR. It is not known whether these parameters can berelated by a multiplicative constant. Arithmetic complexity of matrix multiplication and barriers.
A well-known problem incomputer science concerning tensors is about the arithmetic complexity of matrix multiplication.Asymptotically how many scalar additions and multiplications are required to multiply two m × m matrices? The answer is known to be between n and Cn . ... , or in other words, the exponentof matrix multiplication ω is known to be between 2 and 2 . ... [LG14]. The complexity ofmatrix multiplication turns out to be determined by the tensor rank of the matrix multiplicationtensors (cid:104) m, m, m (cid:105) corresponding to taking the trace of the product of three m × m matrices. Explicitly, (cid:104) m, m, m (cid:105) corresponds to the trilinear map (cid:80) mi,j,k =1 x ij y jk z ki . In practice, upper bounds on therank of the matrix multiplication tensors are obtained by proving a chain of inequalities (cid:104) m, m, m (cid:105) ≤ T ≤ I r T , which is usually taken to be a Coppersmith–Winograd tensor, andan r that is small relatively to m . It was first shown by Ambainis, Filmus and Le Gall [AFLG15]that there is a barrier for this strategy to give fast algorithms. This barrier was recently extendedand simplified in several works [BCC + + n →∞ Q( T ⊗ n ) /n isstrictly smaller than the asymptotic rank lim n →∞ R( T ⊗ n ) /n , then one cannot obtain ω = 2 via T .These barriers rely on the fact that the asymptotic subrank of the matrix multiplication tensors ismaximal. Summarizing, the rank of the matrix multiplication tensors corresponds to the complexityof matrix multiplication whereas the subrank of any tensor corresponds to the a priori suitability ofthat tensor for use as an intermediate tensor. The upper bounds on the asymptotic subrank used inthe aforementioned results were obtained via slice rank or the related theory of support functionalsand quantum functionals [CVZ18]. Cap sets, sunflowers and independent sets in hypergraphs.
Several well-known problemsin extremal combinatorics can be phrased in terms of the independence number of families ofhypergraphs. One effective collection of upper bound methods proceeds via the subrank of tensors.(For other upper bound methods, see e.g. the recent work of Filmus, Golubev and Lifshitz [FGL19].) Ahypergraph is a a symmetric subset E ⊆ V × V × V . An independent set of E is any subset S ⊆ V suchthat S does not induce any edges in E , that is, E ∩ ( S × S × S ) = ∅ . The independence number α ( E )of E is the largest size of any independence set in E . For any hypergraph E ⊆ [ n ] × [ n ] × [ n ], if T ⊆ F n × n × n is any tensor supported on E ∪ { ( i, i, i ) : i ∈ [ n ] } , then α ( E ) ≤ Q( T ) . Indeed, for any independent set S of E the subtensor T | S × S × S is a diagonal tensor with nonzerodiagonal and T ≥ T | S × S × S . For example, the resolution of the cap set problem by Ellenberg andGijswijt [EG17], as simplified by Tao [Tao16], can be thought of as upper bounding the subrankof tensors corresponding to strong powers of the hypergraph consisting of the edge (1 , ,
3) andpermutations. The Erd˝os–Szemer´edi sunflower problem for three petals was resolved by Naslund andSawin [NS17] by similarly considering the strong powers of the hypergraph consisting of the edge(1 , ,
2) and permutations. In both cases slice rank was used to obtain the upper bound. Anotherresult in extremal combinatorics via analytic rank was recently obtained by Bri¨et [Bri19].
We establish a number of basic properties of geometric rank. These imply close connections betweengeometric rank and other notions of rank, and thus bring in a new set of algebraic geometric tools tohelp reason about the various notions of rank. In particular, our new upper bounds on the (border)subrank of matrix multiplication follow easily from our basic results.
Subrank and slice rank.
We prove that the geometric rank GR( T ) is at most the slice rank SR( T )of Tao [Tao16] and at least the subrank Q( T ) of Strassen [Str87] (see Theorem 4.1). Theorem 1.
For any tensor T , Q( T ) ≤ GR( T ) ≤ SR( T ) .
5e thus add GR to the collection of tools to upper bound the subrank of tensors Q and inturn the independence number of hypergraphs. We prove these inequalities by proving that GRis monotone under ≤ , additive under the direct sum of tensors, and has value 1 on the trivial I tensor. We also give a second more direct proof of this inequality (Theorem 7.1). Border subrank.
We extend our upper bound on subrank to border subrank , the (widely studied)approximative version of subrank.The main ingredient in this extension is the following fact (which itself exploits the algebraic-geometric nature of definition of GR): the set of tensors { T ∈ F n × n × n | GR( T ) ≤ m } is closed inthe Zariski topology. In other words, geometric rank is lower-semicontinuous. This implies that thegeometric rank also upper bounds the border subrank Q( T ) (see Theorem 5.1). Theorem 2.
For any tensor T , Q( T ) ≤ GR( T ) . As far as we know, GR is a new tensor parameter. We show that GR is not the same parameteras Q, Q or SR (Remark 6.4 and Remark 6.6).
Matrix multiplication.
In the study of the complexity of matrix multiplication, Strassen [Str87]proved that for the matrix multiplication tensors (cid:104) m, m, m (cid:105) ∈ F m × m × m the border subrank islower bounded by (cid:100) m (cid:101) ≤ Q( (cid:104) m, m, m (cid:105) ). We prove that this lower bound is optimal by provingthe following (see Theorem 6.1). Theorem 3.
For any positive integers e ≤ h ≤ (cid:96) , Q( (cid:104) e, h, (cid:96) (cid:105) ) = GR( (cid:104) e, h, (cid:96) (cid:105) ) = (cid:40) eh − (cid:98) ( e + h − (cid:96) ) (cid:99) if e + h ≥ (cid:96),eh otherwise . In particular, we have Q( (cid:104) m, m, m (cid:105) ) ≤ Q( (cid:104) m, m, m (cid:105) ) = GR( (cid:104) m, m, m (cid:105) ) = (cid:100) m (cid:101) for any m ∈ N . Our computation of GR here is a calcluation of the dimension of a variety. We do this bystudying the dimension of various sections of that variety, which then reduces to linear algebraicquestions about matrices (we are talking about matrix multiplication after all).Our result improves the previously best known upper bound on the subrank of matrix multiplica-tion of Christandl, Lucia, Vrana and Werner [CLVW18], which was Q( (cid:104) m, m, m (cid:105) ) ≤ m − m . In fact,our upper bound on GR( (cid:104) e, h, (cid:96) (cid:105) ) exactly matches the lower bound on Q( (cid:104) e, h, (cid:96) (cid:105) ) of Strassen [Str87],for any nonnegative integers e , h , and (cid:96) . We thus solve the problem of determining the exact valueof Q( (cid:104) e, h, (cid:96) (cid:105) ). Analytic rank.
Finally, we establish a strong connection between geometric rank and analyticrank.We prove that for any tensor T ∈ Z n × n × n ⊆ C n × n × n with integer coefficients, the geometricrank of T equals the liminf of the analytic rank of the tensors T p ∈ F n × n × n p obtained from T bereducing all coefficients modulo p and letting p go to infinity over all primes (see Theorem 8.1). That is, the statement GR( T ) ≤ m is characterized by the vanishing of a finite number of polynomials. heorem 4. For every tensor T over Z we have lim inf p →∞ AR( T p ) = GR( T ) . This result is in fact the source of our definition of geometric rank. The analytic rank of atensor is defined as the bias of a certain polynomial on random inputs. By simple transformations,computing the analytic rank over F p reduces to computing the number of solutions of a system ofpolynomial equations over F p . Namely,AR( T p ) = n + n − log p |{ ( x, y ) ∈ F n p × F n p : T p ( x, y, · ) = 0 }| . This system of polynomial equations defines a variety, and it is natural to expect that the dimensionof the variety roughly determines the number of F p -points of the variety. This expectation is nottrue in general, but under highly controlled circumstances something like it is true. This is howwe arrived at the definition of geometric rank (which eventually turned out to have very naturalproperties on its own, without this connection to analytic rank).Actually establishing the above liminf result is quite roundabout, and requires a number of toolsfrom algebraic geometry and number theory. In particular, we do not know whether this liminf canbe replaced by a limit!We stress that analytic rank is only defined for tensors over prime fields of positive characteristic,whereas geometric rank is defined for tensors over any field. By the aforementioned result, geometricrank over the complex numbers can be thought of as an extension of analytic rank to characteristic 0.Finding an extension of analytic rank beyond finite fields is mentioned as an open problem byLovett [Lov19, Problem 1.10]. Organization of this paper
In the next section we formally define geometric rank. In Section 3, we give some alternativedefinitions of geometric rank that help us reason about it. In Section 4 and Section 5 we show therelationship between geometric rank, slice rank, subrank and border subrank. In Section 6 we usethe established properties of geometric rank to give a proof of our upper bound on the (border)subrank of matrix multiplication. In Section 7 we give a more direct proof of the inequality betweenslice rank and geometric rank. Finally, in Section 8 we establish the relationship between geometricand analytic ranks.
In this section we set up some general notation and define geometric rank. Let F be an algebraicallyclosed field. Dimension and codimension.
The notion of dimension that we use is the standard notion inalgebraic geometry, and is defined as follows. Let V ⊆ F n be a (possibly reducible) algebraic variety.The codimension codim V is defined as n − dim V . The dimension dim V is defined as the lengthof a maximal chain of irreducible subvarieties of V [Har92]. In our proofs we will use basic factsabout dimension: the dimension of a linear space coincides with the notion from linear algebra, thedimension is additive under the cartesian product, the dimension of a locally open set equals thedimension of its closure and dimension behaves well under projections ( x, y ) (cid:55)→ y .7 otation about tensors. Let F n × n × n be the set of all three-dimensional arrays T = ( T i,j,k ) i ∈ [ n ] ,j ∈ [ n ] ,k ∈ [ n ] with T i,j,k ∈ F . We refer to the elements of F n × n × n as the n × n × n tensors over F . Toany tensor T ∈ F n × n × n we associate the polynomial in F [ x , . . . , x n , y , . . . , y n , z , . . . , z n ]defined by T ( x , . . . , x n , y , . . . , y n , z , . . . , z n ) = (cid:88) i ∈ [ n ] (cid:88) j ∈ [ n ] (cid:88) k ∈ [ n ] T i,j,k x i y j z k and the trilinear map F n × F n × F n → F defined by T ( x, y, z ) = T ( x , . . . , x n , y , . . . , y n , z , . . . , z n ) . Definition 2.1.
The geometric rank of a tensor T ∈ F n × n × n , written GR( T ), is the codimensionof the set of elements ( x, y ) ∈ F n × F n such that T ( x, y, z ) = 0 for all z ∈ F n . That is,GR( T ) := codim { ( x, y ) ∈ F n × F n | ∀ z ∈ F n : T ( x, y, z ) = 0 } . For any ( x, y ) ∈ F n × F n we define the vector T ( x, y, · ) = ( T ( x, y, e k )) n k =1 , where e , . . . , e n is thestandard basis of F n . In this notation the geometric rank is given byGR( T ) = codim { ( x, y ) | T ( x, y, · ) = 0 } . For later use we also define the vectors T ( x, · , z ) = T ( x, e j , z ) j and T ( · , y, z ) = T ( e i , y, z ) i , and wedefine the matrices T ( x, · , · ) = T ( x, e j , e k ) j,k , T ( · , y, · ) = T ( e i , y, e k ) i,k and T ( · , · , z ) = T ( e i , e j , z ) i,j .We defined the geometric rank of tensors with coefficients in an algebraically closed field. Fortensors with coefficients in an arbitrary field we naturally define the geometric rank via the embeddingof the field in its algebraic closure. Computer software.
One can compute the dimension of an algebraic variety V ⊆ F n usingcomputer software like Macaulay2 [GS] or Sage [Sag17]. This allows us to easily compute thegeometric rank of small tensors. For example, for Example 1 in the introduction over the field F = C ,one verifies in Macaulay2 with the commands R = CC[x1,x2,y1,y2];dim ideal(x1*y1, x2*y1 + x1*y2) or in Sage with the commands
A.
Koiran [Koi97] studied the computational complexity of the prob-lem of deciding whether the dimension of an algebraic variety V ⊆ C n is at least a given number.When V is given by polynomial equations over the integers the problem is in PSPACE, and assumingthe Generalized Riemann Hypothesis the problem is in the Arthur–Merlin class AM. Thus the sameupper bounds apply to computing GR.In the other direction, Koiran showed that computing dimension of algebraic varieties in generalis NP-hard. We know of no hardness results for computing GR.8 igher-order tensors. Our definition of geometric rank extends naturally from the set of 3-tensors F n × n × n to the set of k -tensors F n ×···× n k for any k ≥ k -tensor T ∈ F n ×···× n k asGR( T ) := codim { ( x , . . . , x k − ) ∈ F n × · · · × F n k − | ∀ x k ∈ F n k : T ( x , . . . , x k − , x k ) = 0 } . For k = 2 geometric rank coincides with matrix rank. Our results extend naturally to k -tensorswith this definition, but for clarity our exposition will be in terms of 3-tensors. We give two alternative descriptions of geometric rank that we will use later. The first descriptionrelates geometric rank to the matrix rank of the matrices T ( x, · , · ) = ( T ( x, e j , e k )) j,k . The seconddescription shows that the geometric rank of T ( x, y, z ) is symmetric under permuting the variables x , y and z . Both theorems rely on an understanding of the dimension of fibers of a (nice) map. Theorem 3.1.
For any tensor T ∈ F n × n × n , dim { ( x, y ) | T ( x, y, · ) = 0 } = max i dim (cid:8) x | dim { y | T ( x, y, · ) = 0 } = i (cid:9) + i = max i dim (cid:8) x | corank T ( x, · , · ) = i (cid:9) + i and therefore GR( T ) = codim { ( x, y ) | T ( x, y, · ) = 0 } = min j codim (cid:8) x | rank T ( x, · , · ) = j (cid:9) + j. Proof.
Let V = { ( x, y ) | T ( x, y, · ) = 0 } . Let W = F n . Let π : V → W map ( x, y ) to x . Define thesets W i = { x | corank( T ( x, · , · )) = i } . The rank-nullity theorem for matrices gives for any fixed x that corank( T ( x, · , · )) = dim { y | T ( x, y, · ) = 0 } . The sets W i are locally closed, that is, each W i is the intersection of an open set and a closed set. Let V i = π − ( W i ). The set V i is also locallyclosed. We have that W = ∪ i W i and so V = ∪ i V i . Therefore, dim V = max i dim V i . We claim thatdim V i = dim W i + i . From this claim follows dim V = max i dim W i + i , which finishes the proof.We prove the claim that dim V i = dim W i + i . For every x ∈ W i the fiber dimension dim π − ( x )equals i . Write V i as a union of irreducible components V ij . Let W ij be the closure of π ( V ij ).We now apply Theorem 3.3 (see the end of this section) with X = V i and X = V ij . For any p = ( x, y ) ∈ X we have that π − ( π ( p )) = { ( x, y (cid:48) ) | T ( x, y (cid:48) , · ) = 0 } . The set { y (cid:48) | T ( x, y (cid:48) , · ) = 0 } is a linear subspace and thus irreducible. Therefore, π − ( π ( p )) is irreducible. Then Theorem 3.3gives that dim V ij = dim W ij + i . We have that max j dim W ij = dim W i , so taking the j maximisingdim W ij gives dim V i ≤ dim W i + i . Also max j dim V ij = dim V i , so taking the j maximising dim V ij gives dim V i ≥ dim W i + i . Theorem 3.2.
For any tensor T , GR( T ) = codim { ( x, y ) | T ( x, y, · ) = 0 } = codim { ( x, z ) | T ( x, · , z ) = 0 } = codim { ( y, z ) | T ( · , y, z ) = 0 } . Proof.
We apply Theorem 3.1 to T and to T after swapping y and z to get that the codimensions of { ( x, y ) | T ( x, y, · ) = 0 } and { ( x, z ) | T ( x, · , z ) = 0 } are equal to min j codim { x | rank T ( x, · , · ) = j } + j .This proves the first equality. The second equality is proven similarly.9 heorem 3.3 ([Har92, special case of Theorem 11.12]) . Let X ⊆ F n × F n be the affine cone overa quasi-projective variety, that is, X = { ( x, y ) ∈ F n × F n | f ( x, y ) = 0 , . . . , f k ( x, y ) = 0 , g ( x, y ) (cid:54) = 0 , . . . , g m ( x, y ) (cid:54) = 0 } where the f i and g i are homogeneous polynomials. Let π : X → F n map ( x, y ) to x . Let X ⊆ X bean irreducible component. Suppose that the fiber π − ( π ( p )) is irreducible for every p ∈ X . Then dim X = dim π ( X ) + min p ∈ X dim π − ( π ( p )) . Recall that the subrank Q( T ) of T is the largest number s such that I s ≤ T and the slice rank SR( T )is the smallest number r such that T ( x, y, z ) can be written as a sum of r trilinear maps of theform f ( x ) g ( y, z ) or f ( y ) g ( x, z ) or f ( z ) g ( x, y ). Theorem 4.1.
For any tensor T , Q( T ) ≤ GR( T ) ≤ SR( T ) . Theorem 4.1 will follow from the following basic properties of GR. We will give a more directproof of the inequality GR( T ) ≤ SR( T ) in Section 7. Recall from the introduction that for any twotensors S ∈ F m × m × m and T ∈ F n × n × n we write S ≤ T if there are matrices A, B, C such that S = ( A, B, C ) · T where we define ( A, B, C ) · T := ( (cid:80) a,b,c A ia B jb C kc T a,b,c ) i,j,k . Lemma 4.2. GR is ≤ -monotone: if S ≤ T , then GR( S ) ≤ GR( T ) .Proof. Let T ∈ F n × n × n . We claim that GR((Id , Id , C ) · T ) ≤ GR( T ) for any C ∈ F m × n ,where Id denotes an identity matrix of the appropriate size. From this claim and the symmetryof GR (Theorem 3.2), follows the inequalities GR(( A, Id , Id) · T ) ≤ T and GR((Id , B, Id) · T ) ≤ GR( T )for any matrices A ∈ F m × n and B ∈ F m × n . Chaining these three inequalities gives that for anytwo tensors S and T , if S ≤ T , then GR( S ) ≤ GR( T ).We prove the claim. Let S = (Id , Id , C ) · T . Let M k = ( T i,j,k ) ij be the 3-slices of T andlet N k = ( S i,j,k ) ij be the 3-slices of S . Since S = (Id , Id , C ) · T , the matrices N , . . . , N m are inthe linear span of the matrices M , . . . , M n . Thus V = { ( x, y ) | x T M y = · · · = x T M n y = 0 } isa subset of W = { ( x, y ) | x T N y = · · · = x T N m y = 0 } . Therefore, dim V ≤ dim W and it followsthat GR( S ) = codim W ≤ codim V = GR( T ).Let T ∈ F m × m × m and T ∈ F n × n × n be tensors with 3-slices A k and B k respectively. Thedirect sum T ⊕ T ∈ F ( m + n ) × ( m + n ) × ( m + n ) is defined as the tensor with 3-slices A k ⊕ n × n for k = 1 , . . . , m and 0 m × m ⊕ B k for k = m + 1 , . . . , m + n where 0 a × b denotes the zero matrixof size a × b . In other words, T ⊕ T is the block-diagonal tensor with blocks T and T . Lemma 4.3. GR is additive under direct sums: GR( T ⊕ T ) = GR( T ) + GR( T ) .Proof. Let A k be the 3-slices of T and let B k be the 3-slices of T . Let T = T ⊕ T be the directsum with 3-slices M k . Then V = { ( x, y ) | T ( x, y, · ) = 0 } = { ( x, y ) | x T M y = · · · = x T M m + n y = 0 }
10s the cartesian product of V = { ( x, y ) | x T A y = · · · = x T A m y = 0 } and V = { ( x, y ) | x T B y = · · · = x T B n y = 0 } . Thus dim V = dim V + dim V [Har92, page 138]. Therefore, GR( T ) = GR( T ) + GR( T ). Lemma 4.4. GR is sub-additive under element-wise sums: GR( S + T ) ≤ GR( S ) + GR( T ) .Proof. Note that S + T ≤ S ⊕ T . Thus, GR( S + T ) ≤ GR( S ⊕ T ) = GR( S ) + GR( T ), where theinequality uses Lemma 4.2, and the equality uses Lemma 4.3. Lemma 4.5. If SR( T ) = 1 , then GR( T ) = 1 .Proof. It is sufficient to consider a tensor T ∈ F × n × n with one nonzero slice. Then we have that T (0 , F n , F n ) = 0, and so GR( T ) = 1 + n − n = 1. Lemma 4.6.
For every r ∈ N we have GR( I r ) = r .Proof. We have SR( I ) = 1 and so GR( I ) = 1 (Lemma 4.5). Since I r is a direct sum of r copiesof I and geometric rank is additive under taking the direct sum ⊕ (Lemma 4.4), we find thatGR( I r ) = r GR( I ) = r . Proof of Theorem 4.1.
We prove that GR( T ) ≤ SR( T ). Let r = SR( T ). Then there are tensors T , . . . , T r so that T = (cid:80) ri =1 T i and SR( T i ) = 1. Then also GR( T i ) = 1 (Lemma 4.5). Subadditivityof GR under element-wise sums (Lemma 4.4) givesGR( T ) ≤ r (cid:88) i =1 GR( T i ) = r = SR( T ) . We prove that Q( T ) ≤ GR( T ). Let s = Q( T ). Then I s ≤ T . We know GR( I s ) = s (Lemma 4.6).By the ≤ -monotonicity of GR (Lemma 4.2), we haveQ( T ) = s = GR( I s ) ≤ GR( T ) . In this section we extend the inequality Q( T ) ≤ GR( T ) (Theorem 4.1) to the approximative versionof subrank, called border subrank. To define border subrank we first define degeneration (cid:69) , whichis the approximative version of restriction ≤ . We write S (cid:69) T , and we say S is a degeneration of T ,if for some e ∈ N we have S + εS + ε S + · · · + ε e S e = ( A ( ε ) , B ( ε ) , C ( ε )) · T for some tensors S i over F and for some matrices A ( ε ) , B ( ε ) , C ( ε ) whose coefficients are Laurentpolynomials in the formal variable ε . Equivalently, S (cid:69) T if and only if S is in the orbit closure G · T where G denotes the group GL n × GL n × GL n , G · T denotes the natural group action that we also11sed in the definition of ≤ , and the closure is taken in the Zariski topology [BCS97, Theorem 20.24].(When F = C one may equivalently take the closure in the Euclidean topology.) Recall that thesubrank of T is defined as Q( T ) = max { n ∈ N | I n ≤ T } . The border subrank of T is defined asQ( T ) = max { n ∈ N | I n (cid:69) T } . Clearly, Q( T ) ≤ Q( T ). Theorem 5.1.
For any tensor T , Q( T ) ≤ GR( T ) . To prove Theorem 5.1 we use the following theorem on upper-semicontinuity of fiber dimension.
Theorem 5.2 ([Har92, special case of Corollary 11.13]) . Let X be the zero set of bi-homogeneouspolynomials, that is, X = { ( a, b ) ∈ F m × F m | f ( a, b ) = · · · = f k ( a, b ) = 0 } where the f i ( a, b ) are polynomials that are homogeneous in both a and b . Let π : X → F m map ( a, b ) to b . Let Y = π ( X ) be its image. For any q ∈ Y , let λ ( q ) = dim( π − ( q )) . Then λ ( q ) is anupper-semicontinuous function of q , that is, the set { q ∈ Y | λ ( q ) ≥ m } is Zariski closed in Y . Lemma 5.3. GR is lower-semicontinuous: for any n i , m ∈ N the set { T ∈ F n × n × n | GR( T ) ≤ m } is Zariski closed.Proof. We define the set X = { ( T, x, y ) ∈ F n × n × n × F n × F n | T ( x, y, F n ) = 0 } . Let π : X → F n × n × n map ( T, x, y ) to T . Let Y = π ( X ) = F n × n × n be the image of π . Forany T ∈ Y let λ ( T ) := dim( π − ( T )). Then λ ( T ) is an upper-semicontinuous function of T in theZariski topology on Y by Theorem 5.2. This means that the set { T ∈ F n × n × n | λ ( T ) ≥ m } isclosed for every m ∈ N . It follows that { T ∈ F n × n × n | GR( T ) ≤ m } is closed for every m ∈ N . Remark 5.4.
A well-known example of a lower-semicontinuous function is matrix rank. Indeed,the set of matrices of rank at most m is the zero set of the determinants of all ( m + 1) × ( m + 1)submatrices. For geometric rank we do not know an explicit set of generators for the vanishingideal of { T ∈ F n × n × n | GR( T ) ≤ m } . For slice rank the set { T ∈ F n × n × n | SR( T ) ≤ m } isalso known to be Zariski closed and explicit vanishing polynomials for this variety were recentlyobtained by Bl¨aser, Ikenmeyer, Lysikov, Pandey and Schreyer [BIL + Lemma 5.5. GR is (cid:69) -monotone: if S (cid:69) T , then GR( S ) ≤ GR( T ) Proof.
For all g ∈ G we have GR( g · T ) = GR( T ) by Lemma 4.2. The set { T (cid:48) | GR( T (cid:48) ) ≤ GR( T ) } is Zariski closed by Lemma 5.3. It contains the orbit G · T and hence also its Zariski closure G · T ,that is, { T (cid:48) | T (cid:48) (cid:69) T } = G · T ⊆ { T (cid:48) | GR( T (cid:48) ) ≤ GR( T ) } . Therefore, GR( S ) ≤ GR( T ). Proof of Theorem 5.1.
Let n = Q( T ). Then I n (cid:69) T by the definition of Q, and so n ≤ GR( T )by Lemma 5.5. This proves the claim. 12 The border subrank of matrix multiplication
In the context of constructing fast matrix multiplication algorithms, Strassen [Str87, Theorem 6.6]proved that for any positive integers e ≤ h ≤ (cid:96) the border subrank of the matrix multiplicationtensor (cid:104) e, h, (cid:96) (cid:105) is lower bounded byQ( (cid:104) e, h, (cid:96) (cid:105) ) ≥ (cid:40) eh − (cid:98) ( e + h − (cid:96) ) (cid:99) if e + h ≥ (cid:96),eh otherwise . (1)Here (cid:104) e, h, (cid:96) (cid:105) is the tensor that corresponds to taking the trace of the product of an e × h matrix,an h × (cid:96) matrix and an (cid:96) × e matrix. We prove using the geometric rank that this lower bound isoptimal. Theorem 6.1.
For any positive integers e ≤ h ≤ (cid:96) Q( (cid:104) e, h, (cid:96) (cid:105) ) = GR( (cid:104) e, h, (cid:96) (cid:105) ) = (cid:40) eh − (cid:98) ( e + h − (cid:96) ) (cid:99) if e + h ≥ (cid:96),eh otherwise . In particular, we have Q( (cid:104) m, m, m (cid:105) ) = GR( (cid:104) m, m, m (cid:105) ) = (cid:100) m (cid:101) for any m ∈ N .Proof. Since Q( (cid:104) e, h, (cid:96) (cid:105) ) ≤ GR( (cid:104) e, h, (cid:96) (cid:105) ) (Theorem 5.1) and since we have the lower bound in (1),it suffices to show that GR( (cid:104) e, h, (cid:96) (cid:105) ) is at most eh − (cid:98) ( e + h − (cid:96) ) / (cid:99) if e + h ≥ (cid:96) and at most eh otherwise.Let T = (cid:104) e, h, (cid:96) (cid:105) . Let V = { ( x, y ) ∈ F eh × F h(cid:96) | T ( x, y, · ) = 0 } . Then GR( T ) = eh + h(cid:96) − dim V .From Theorem 3.1 it follows thatdim V = max i dim { x ∈ F eh | dim { y ∈ F h(cid:96) | T ( x, y, · ) = 0 } = i } + i. (2)We now think of F eh , F h(cid:96) and F (cid:96)e as the matrix spaces F e × h , F h × (cid:96) and F (cid:96) × e . Then T gives thetrilinear map T : F e × h × F h × (cid:96) × F (cid:96) × e → F : ( X, Y, Z ) (cid:55)→ Tr(
XY Z ). Therefore, T ( X, Y, · ) = 0 if andonly if XY = 0. If the rank of X as an e × h matrix equals r , thendim { Y ∈ F h × (cid:96) | T ( X, Y, · ) = 0 } = ( h − r ) (cid:96), since Y is any matrix with columns from ker( X ). We havedim { X ∈ F e × h | rank( X ) = r } = er + ( h − r ) r. Thus the relevant values of i in (2) are of the form i = ( h − r ) (cid:96) and we have thatdim V = max r dim { X ∈ F e × h | rank X = r } +( h − r ) (cid:96) = max r er +( h − r ) r +( h − r ) (cid:96) = max r f ( r )+ h(cid:96) where f ( r ) = r (∆ − r ) with ∆ := e + h − (cid:96) . Thus,GR( T ) = eh − max r f ( r ) . Over the integers, the function f attains its maximum at (cid:98) ∆2 (cid:99) (and at (cid:100) ∆2 (cid:101) ), but this may be outsidethe interval [0 , e ] that we want to maximise over (recall e ≤ h ≤ l ). Observe that if ∆ ≥ ≥ ∆ / ≥
0, meaning that f does attain its global maximum in the interval [0 , e ]. On the otherhand, if ∆ ≤ r (∆ − r ) ≤ f (0) for every r ≥
0, so the maximum of f in the interval [0 , e ]is at the endpoint r = 0. Summarizing,max ≤ r ≤ e f ( r ) = (cid:40) (cid:98) ∆ (cid:99) if ∆ ≥ , . (3)This completes the proof. Remark 6.2.
Theorem 6.1 gives the upper bound Q( (cid:104) m, m, m (cid:105) ) ≤ Q( (cid:104) m, m, m (cid:105) ) = (cid:100) m (cid:101) on thesubrank of matrix multiplication Q( (cid:104) m, m, m (cid:105) ). This improves the previously best known upperbound Q( (cid:104) m, m, m (cid:105) ) ≤ m − m from [CLVW18, Equation 29]. Remark 6.3.
Geometric rank GR is not sub-multiplicative under the tensor Kronecker product ⊗ .We give an example. The matrix multiplication tensor (cid:104) m, m, m (cid:105) can be written as the product (cid:104) m, m, m (cid:105) = (cid:104) m, , (cid:105) ⊗ (cid:104) , m, (cid:105) ⊗ (cid:104) , , m (cid:105) and GR( (cid:104) m, , (cid:105) ) = GR( (cid:104) , m, (cid:105) ) = GR( (cid:104) , , m (cid:105) ) = 1whereas GR( (cid:104) m, m, m (cid:105) ) = (cid:100) m (cid:101) by Theorem 6.1. Remark 6.4.
Geometric rank GR is not the same as subrank Q or border subrank Q. For example,for the trilinear map W ( x , x , y , y , z , z ) = x y z + x y z + x y z we find GR( W ) = 2 (seethe example in the introduction), whereas Q( W ) = Q( W ) = 1. The latter follows from the fact that (cid:101) Q( W ) = 1 . ... [Str91], where (cid:101) Q( T ) := lim n →∞ Q( T ⊗ n ) /n is the asymptotic subrank of T , sinceQ( T ) ≤ (cid:101) Q( T ) [Str87]. Remark 6.5.
Geometric rank GR is not super-multiplicative under the tensor Kronecker product ⊗ .Here is an example. Let (cid:101) SR( T ) := lim n →∞ SR( T ⊗ n ) /n and let (cid:101) GR( T ) := lim n →∞ GR( T ⊗ n ) /n ,whenever these limits are defined. From the fact that Q( T ) ≤ GR( T ) ≤ SR( T ) and the fact that (cid:101) Q( W ) = (cid:101) SR( W ) = 1 . ... [CVZ18] it follows that (cid:101) GR( W ) = 1 . .. , whereas GR( W ) = 2. Weconclude that GR is not super-multiplicative. We have seen already in Remark 6.3 that GR is notsub-multiplicative. Remark 6.6.
Geometric rank GR is not the same as slice rank SR. For example, for the matrixmultiplication tensor (cid:104) m, m, m (cid:105) we find that GR( (cid:104) m, m, m (cid:105) ) = (cid:100) m (cid:101) (Theorem 6.1), whereas itwas known that SR( (cid:104) m, m, m (cid:105) ) = m [BCC + In Section 4 we proved, by chaining the basic properties of geometric rank, that geometric rank isat most slice rank, that is, GR( T ) ≤ SR( T ). What is the largest gap between GR( T ) and SR( T )?Motivated by this question, and motivated by the analogous question for analytic rank insteadof geometric rank that we discussed in the introduction we give a direct proof of the inequalityGR( T ) ≤ SR( T ).In fact, we prove a chain of inequalities GR( T ) ≤ ZR( T ) ≤ SR( T ) where ZR( T ) is defined asfollows. We will henceforth use the following piece of notation for a tensor T ∈ F n × n × n ; V ( T ) = { ( x, y ) ∈ F n × n | ∀ z ∈ F n : T ( x, y, z ) = 0 } . (4)14oreover, we use the following standard notation for the variety cut out by polynomials f , . . . , f s ; V ( f , . . . , f s ) = { x | f ( x ) = · · · = f s ( x ) = 0 } . (5)Let F [ x , y ] = F [ x , . . . , x n , y , . . . , y n ] and let F [ x , y , z ] = F [ x , . . . , x n , y , . . . , y n , z , . . . , z n ].Let F [ x , y ] { (0 , , (1 , , (1 , } ⊆ F [ x , y ] be the subset of polynomials that are bi-homogeneous of bi-degree (0 , ,
0) or (1 , F [ x , y ] { (0 , , (1 , , (1 , } contains the polynomials in F [ x , . . . , x n ] that are homogeneous of degree 1, and the polynomials in F [ y , . . . , y n ] that arehomogeneous of degree 1, and the polynomials in F [ x , y ] that are homogeneous of degree 1 in x , . . . , x n and homogeneous of degree 1 in y , . . . , y n . For any tensor T we defineZR( T ) = min (cid:8) s ∈ N | ∃ f , . . . , f s ∈ F [ x , y ] { (0 , , (1 , , (1 , } : V ( f , . . . , f s ) ⊆ V ( T ) (cid:9) . Theorem 7.1.
Let T be a tensor. Then GR( T ) ≤ ZR( T ) ≤ SR( T ) .Proof. We prove that ZR( T ) ≤ SR( T ). Let r = SR( T ). View T as a polynomial T ∈ F [ x , y , z ].Write T = (cid:80) ri =1 T i with SR( T i ) = 1 for every i . Then T i = f i g i for some f i ∈ F [ x , y ] { (0 , , (1 , , (1 , } and g i ∈ F [ x , y , z ]. We claim that V ( f , . . . , f r ) ⊆ V ( T ). Indeed, if ( x, y ) ∈ V ( f , . . . , f r ), then T i ( x, y, z ) = 0 for every i and every z , and therefore T ( x, y, z ) = 0 for every z . We conclude thatZR( T ) ≤ r = SR( T ).We prove that GR( T ) ≤ ZR( T ). Let s = ZR( T ). Then there are f , . . . , f s ∈ F [ x , y ] { (0 , , (1 , , (1 , } such that V ( f , . . . , f s ) ⊆ V ( T ). We haveGR( T ) = codim V ( T ) ≤ codim V ( f , . . . , f s ) ≤ s = ZR( T ) , where the first inequality follows from the containment V ( f , . . . , f s ) ⊆ V ( T ) which implies thatdim V ( f , . . . , f s ) ≤ dim V ( T ). For a tensor T over Z and a prime number p , we denote by T p the 3-tensor over F p obtained byreducing all coefficients of T modulo p . In this section we prove the following tight relationshipbetween AR( T p ) and GR( T ). Theorem 8.1.
For every tensor T over Z we have lim inf p →∞ AR( T p ) = GR( T ) . The starting point for the proof of Theorem 8.1 is the important observation that analytic rankcan be written in terms of the number of F p -points of the algebraic variety V ( T p ), that is, for anytensor T ∈ Z n × n × n ,AR( T p ) = n + n − log p | V ( T p )( F p ) | . For the proof of Theorem 8.1 we will need to prove three auxiliary results: that the Bertini–Noether Theorem can be extended to reducible varieties (Theorem 8.3 below), that prime fields arerich enough infinitely often to contain any finite set of algebraic numbers (Lemma 8.5 below), andthat for any variety satisfying a mild assumption, its number of rational points in a finite field isdetermined by its dimension (Lemma 8.8 below).15 .1 Bertini–Noether Theorem
In this subsection we extend the Bertini–Noether Theorem to reducible varieties. The Bertini–Noether Theorem says that, roughly, if a variety is irreducible then applying a homomorphism onthe defining equations—for example the modulo- p homomorphism—typically does not change itsinvariants (see Proposition 10.4.2 in [FJ05]). Theorem 8.2 (Bertini–Noether Theorem [FJ05]) . Let f , . . . , f m ∈ R [ x ] , where R is an integraldomain, such that V = V ( f , . . . , f m ) is (absolutely) irreducible. There exists a nonzero c ∈ R suchthat for every homomorphism φ : R → K into a field K , if φ ( c ) (cid:54) = 0 then V ( φ ( f ) , . . . , φ ( f m )) ⊆ K is (absolutely) irreducible of dimension dim V and degree deg V . The version of the Bertini–Noether Theorem that we need, which does not assume irreducibility,is as follows. First, let us note that any variety defined over a field F , where F is the field of fractionsof an integral domain R , can also be defined over R , by clearing denominators. For example, anyvariety defined over the algebraic numbers Q can also be defined over the algebraic integers Z . Theorem 8.3 (Extended Bertini–Noether Theorem) . Let f , . . . , f m ∈ R [ x ] , where R is an integraldomain whose field of fractions is algebraically closed. There exists a nonzero C ∈ R such that forevery homomorphism ψ : R → K into a field K , if ψ ( C ) (cid:54) = 0 then V ψ := V ( ψ ( f ) , . . . , ψ ( f m )) ⊆ K is of dimension dim V and degree deg V . Moreover, if the irreducible components of V ( f , . . . , f m ) are V , . . . , V k , where I( V i ) = (cid:104) f i,j (cid:105) j with f i,j ∈ R [ x ] , then the irreducible components of V ψ are V ψ , . . . , V ψk , where V ψi = V ( ψ ( f i,j ) j ) . For the proof of Theorem 8.3 we will need some notation and a standard auxiliary result, asfollows. Let R be a (commutative) ring. For a ideal I in R , the radical of I (in R ) is the ideal √ I = { f ∈ R | ∃ n ∈ N : f n ∈ I } . Moreover, for a ring homomorphism ψ : R → R (cid:48) we denote ψ ( I ) = (cid:104) ψ ( f ) | f ∈ I (cid:105) , which is an ideal in R (cid:48) . Lemma 8.4.
Let I be an ideal in a ring R , and let ψ : R → R (cid:48) be a ring homomorphism. Then (cid:113) ψ ( √ I ) = (cid:112) ψ ( I ) .Proof. If p ∈ (cid:112) ψ ( I ) then there is an integer n such that p n ∈ ψ ( I ) ⊆ ψ ( √ I ), hence p ∈ (cid:113) ψ ( √ I ).Let p ∈ (cid:113) ψ ( √ I ), meaning there is an integer n such that p n ∈ ψ ( √ I ). Thus, p n = (cid:80) mi =1 g i ψ ( f i )for some m ∈ N , g i ∈ R (cid:48) and f i ∈ √ I . Note that for every i there is an integer k i such that f k i i ∈ I .Let k = max ≤ i ≤ m k i . Then( p n ) km = (cid:88) d ,...,d m d + ··· + d m = km m (cid:89) i =1 ( g i ψ ( f i )) d i . Observe that every summand has a multiplicand ( g i ψ ( f i )) d i with d i ≥ k ≥ k i , which lies in ψ ( I )since ψ ( f i ) d i = ψ ( f d i i ) and f d i i = f d i − k i i f k i i ∈ I . We deduce that p nkm ∈ ψ ( I ), being a sum ofmembers of the ideal ψ ( I ). Hence p ∈ (cid:112) ψ ( I ), completing the proof. φ ( f i ) ∈ K [ x ] is obtained by applying φ on each of the coefficients of f i . That deg V remains unchanged follows along similar lines to the proof for dim V (see Corollary 9.2.2 in [FJ05]). roof of Theorem 8.3. We begin with some notation. Let F be the the field of fractions of R , whichis algebraically closed by assumption. For an ideal J in F [ x ] we use the following notation: • J R := J ∩ R [ x ] is an ideal in R [ x ], • √ J R is the radical ideal of J R in R [ x ], • ψ ( J ) := ψ ( J R ) is an ideal in K [ x ].Note that J R is indeed an ideal in R [ x ], since R [ x ] is a subring of F [ x ]. Moreover, observe that √ J R = ( √ J ) R ; indeed, f ∈ ( √ J ) R iff f n ∈ J and f ∈ R [ x ] iff f ∈ √ J R .Consider the ideals I = (cid:104) f , . . . , f m (cid:105) and I i = (cid:104) f i,j (cid:105) j in F [ x ]. We will prove the following equalityof ideals in K [ x ]; (cid:112) ψ ( I ) = (cid:113)(cid:89) ψ ( I i ) . (6)We have V ( I ) = (cid:83) i V ( I i ) = V ( (cid:81) i I i ). By Hilbert’s Nullstellensatz, √ I = (cid:112)(cid:81) i I i . We thus havethe following equality of ideals in R [ x ]; √ I R = ( √ I ) R = (cid:16)(cid:113)(cid:89) I i (cid:17) R = (cid:113)(cid:89) I Ri . (7)We deduce (6) as follows; (cid:112) ψ ( I ) = (cid:113) ψ ( I R ) = (cid:113) ψ ( √ I R ) = (cid:114) ψ (cid:0)(cid:113)(cid:89) I Ri (cid:1) = (cid:113) ψ (cid:0) (cid:89) I Ri (cid:1) = (cid:113)(cid:89) ψ ( I Ri ) = (cid:113)(cid:89) ψ ( I i ) , where the second equality follows from Lemma 8.4, the third follows from (7), the fourth again fromLemma 8.4, and the fifth using the fact that ψ is a homomorphism. Now, this implies that in K [ x ], V ψ := V ( ψ ( I )) = V ( (cid:112) ψ ( I )) = V (cid:16)(cid:113)(cid:89) ψ ( I i ) (cid:17) = V (cid:0) (cid:89) ψ ( I i ) (cid:1) = (cid:91) V ( ψ ( I i )) = (cid:91) V ψi , where (6) is used in the third equality.Recall that each V i is an irreducible variety defined over the field F and thus, by clearingdenominators, over the ring R . For each i , applying Theorem 8.2 on any generating set of I( V i ) in R [ x ] and on ψ implies that there is a nonzero c i ∈ R such that if ψ ( c i ) (cid:54) = 0 then V ψi is irreducible,of dimension dim V ψi = dim V i and degree deg V ψi = deg V i . Now, let C = (cid:81) i c i , and note that C ∈ R is nonzero. If ψ ( C ) (cid:54) = 0 then ψ ( c i ) (cid:54) = 0 for all i , which implies that the union proved above, V ψ = (cid:83) i V ψi , is a union of irreducible varieties, and moreover,dim V ψ = max i dim V ψi = max i dim V i = dim V anddeg V ψ = (cid:88) i deg V ψi = (cid:88) i deg V i = deg V. This completes the proof. 17 .2 Modular roots
In this subsection we prove that, intuitively, every finite set of algebraic integers is contained in F p ,for infinitely many primes p . Henceforth, we say that there is a positive density of primes satisfyinga property P ⊆ P (here P is the set of prime numbers) if lim n →∞ |P ∩ [ n ] | / | P ∩ [ n ] | > Lemma 8.5.
For every finite set of algebraic integers S there is a positive density of primes p forwhich there is a homomorphism from Z [ S ] to F p . We will use (a special case of) the Primitive Element Theorem (see, e.g., Section 6.10 in [vdW91]).
Theorem 8.6 (Primitive Element Theorem in Characteristic 0 [vdW91]) . Let K be a finite extensionof a field F of characteristic . Then K = F ( α ) for some α ∈ K . For example, Q ( √ , √
3) = Q ( √ √ Theorem 8.7 ([BB96]) . For every polynomial P ∈ Z [ x ] there is a positive density of prime numbers p such that P has a root modulo p .Proof of Lemma 8.5. Consider Q ( S ), the field extension of the rationals Q obtained by adjoiningall the elements of S . By the Primitive Element Theorem (Theorem 8.6) there exists α ∈ Q ( S ) suchthat Q ( S ) = Q ( α ) = Q [ α ]. Thus, for every α i ∈ S there is a (univariate) polynomial f i ∈ Q [ x ] suchthat α i = f i ( α ). We denote by P be the minimal polynomial of α over Q ; by clearing denominators,we assume without loss of generality that P ∈ Z [ x ].Let p be a prime number such that P has a root a p modulo p and, moreover, p is larger thanthe absolute value of the coefficient denominators of every f i . By Theorem 8.7, applied on P , thereis a positive density of primes satisfying both conditions. Note that f i (mod p ) is a well-definedpolynomial in F p [ x ] by our second condition on p . Consider the function φ p that maps each α i = f i ( α ) ∈ S to f i ( a p ) (mod p ). Since every member of Z [ S ] is a multivariate polynomial in thevariables α i with integer coefficients, we deduce from our first condition on p that the function φ p extends to a homomorphism φ p : Z [ S ] → F p . This completes the proof. We will also need the following asymptotically-tight estimate on the number of rational points in afinite field.
Lemma 8.8.
For every variety V defined over a finite field F , if V has an irreducible componentof dimension dim V that is also defined over F then | V ( F ) | = Θ deg V, n ( | F | dim V ) . The proof of Lemma 8.8 will follow by combining the Lang-Weil Theorem [LW54] with aSchwartz-Zippel-type upper bound (see Lemma 14 in [BT12] or Claim 7.2 in [DKL14]).
Theorem 8.9 (Lang–Weil Bound [LW54]) . For every (absolutely) irreducible variety V defined overa finite field F , | V ( F ) | = | F | dim V (1 + O deg V, n ( | F | − / )) . emma 8.10 (Generalized Schwartz–Zippel lemma [BT12, DKL14]) . For every variety V definedover a finite field F , | V ( F ) | ≤ deg V · | F | dim V .Proof of Lemma 8.8. For the upper bound, apply Lemma 8.10 on V . For the lower bound, let U be anirreducible component of V of dimension dim V that is defined over F , as guaranteed by the statement,and apply Theorem 8.9 on U to obtain | V ( F ) | ≥ | U ( F ) | = Ω deg U, n ( | F | dim U ) = Ω deg V, n ( | F | dim V ).We are now ready to prove the main result of this section. Proof of Theorem 8.1.
Put d = dim V ( T ) and r = deg V ( T ). We will use the notation in (4) and (5).We will show that V ( T ) ⊆ Q N and V ( T p ) ⊆ F pN (here N = n + n ) are related in the followingsense; | V ( T p )( F p ) | = Θ r, N ( p d ) (8)where the lower bound on | V ( T p )( F p ) | holds for infinitely many prime numbers p . This would implyAR( T p ) = − log p (cid:16) | V ( T p )( F p ) || F p | N (cid:17) = N − log p | V ( T p )( F p ) | = GR( T ) − Θ r, N (cid:16) p (cid:17) where the upper bound on AR( T p ) holds for infinitely many prime numbers p . Thus, proving (8)would imply that lim inf p →∞ AR( T p ) = GR( T ), completing the proof.We begin with the upper bound in (8). For every prime p , Theorem 8.3, applied on the ring ofalgebraic integers R = Z , the field K = F p , and on any homomorphism ψ : Z → F p extending the mod- p : Z → F p homomorphism, implies that there is 0 (cid:54) = C ∈ Z such that if ψ ( C ) (cid:54) = 0 then dim V ( T p ) = d and deg V ( T p ) = r . For such p , Lemma 8.10 implies | V ( T p )( F p ) | ≤ deg V ( T p ) · | F p | dim V ( T p ) = rp d .We claim that the condition ψ ( C ) (cid:54) = 0 is satisfied for all but finitely many primes p . Indeed, if ψ ( C ) = 0 then the minimal polynomial P ∈ Z [ x ] of C ∈ Z satisfies P (0) = P ( ψ ( C )) = ψ ( P ( C )) = ψ (0) ≡ p ) , that is, p divides P (0); thus, since P (0) (cid:54) = 0 by the irreducibility P , we have p ≤ | P (0) | , as claimed.The upper in (8) follows.It remains to prove the lower bound in (8). Let U be an irreducible component of V ( T ) ofdimension d . Note that U is defined over some finite extension Z [ S ] of the integers, where S is afinite set of algebraic integers. Lemma 8.5, applied on S , implies that for a positive density of primenumbers p there is a homomorphism φ p : Z [ S ] → F p . Thus, if I( U ) = (cid:104) f j (cid:105) j with f j ∈ Z [ S ][ x ] then U φ p := V ( φ p ( f j ) j ) is defined over F p (rather than F p ). Let p be any such prime. Theorem 8.3,applied on R = Z , K = F p and on any extension ψ p of φ p to a homomorphism from Z to F p ,implies that there is 0 (cid:54) = C ∈ Z such that if ψ p ( C ) (cid:54) = 0 then dim V ( T p ) = d , deg V ( T p ) = r , andthat U ψ p = U φ p is an irreducible component of V ( T p ) of dimension d = dim V ( T p ). Recall that ψ p ( C ) (cid:54) = 0 is satisfied for all but finitely many primes p . Lemma 8.8 therefore implies, together withall of the above, that for a positive density of primes p we have | V ( T p )( F p ) | = Θ deg V ( T p ) , N ( p dim V ( T p ) ) = Θ r, N ( p d ) . This proves (8), and thus we are done. 19 cknowledgements
We would like to thank Avi Wigderson for helpful conversations.Swastik Kopparty: Research supported in part by NSF grants CCF-1253886, CCF-1540634,CCF-1814409 and CCF-1412958, and BSF grant 2014359. Some of this research was done whilevisiting the Institute for Advanced Study.Guy Moshkovitz: This work was conducted at the Institute for Advanced Study, enabled throughsupport from the National Science Foundation under grant number CCF-1412958, and at DIMACS,enabled through support from the National Science Foundation under grant number CCF-1445755.Jeroen Zuiddam: This material is based upon work directly supported by the National ScienceFoundation Grant No. DMS-1638352 and indirectly supported by the National Science FoundationGrant No. CCF-1900460. Any opinions, findings and conclusions or recommendations expressed inthis material are those of the author and do not necessarily reflect the views of the National ScienceFoundation.
References [AFLG15] Andris Ambainis, Yuval Filmus, and Fran¸cois Le Gall. Fast matrix multiplication:limitations of the Coppersmith-Winograd method (extended abstract). In
Proceedingsof the 47th Annual ACM Symposium on Theory of Computing (STOC 2015) , pages585–593, 2015. arXiv:1411 . , doi:10 . . .[Alm19] Josh Alman. Limits on the universal method for matrix multiplication. In Proceedingsof the 34th Computational Complexity Conference (CCC 2019) , pages 12:1–12:24, 2019. arXiv:1812 . , doi:10 . . CCC . . .[AW18] Josh Alman and Virginia Vassilevska Williams. Limits on all known (and someunknown) approaches to matrix multiplication. In Proceedings of the 59th AnnualIEEE Symposium on Foundations of Computer Science (FOCS 2018) , pages 580–591,2018. arXiv:1810 . , doi:10 . . . .[BB96] Daniel Berend and Yuri Bilu. Polynomials with roots modulo every integer. Proc.Amer. Math. Soc. , 124(6):1663–1671, 1996. doi:10 . .[BCC + Discrete Anal. , 2017. arXiv:1605 . , doi:10 . . .[BCC + arXiv ,2017. arXiv:1712 . .[BCS97] Peter B¨urgisser, Michael Clausen, and M. Amin Shokrollahi. Algebraic complexitytheory , volume 315 of
Grundlehren Math. Wiss.
Springer-Verlag, Berlin, 1997. doi:10 . . 20BHH +
18] Abhishek Bhrushundi, Prahladh Harsha, Pooya Hatami, Swastik Kopparty, and MrinalKumar. On multilinear forms: Bias, correlation, and tensor rank. arXiv , 2018. arXiv:1804 . .[BIL +
19] Markus Bl¨aser, Christian Ikenmeyer, Vladimir Lysikov, Anurag Pandey, andFrank-Olaf Schreyer. Variety membership testing, algebraic natural proofs, andgeometric complexity theory, 2019. arXiv:1911 . .[BL15] Abhishek Bhowmick and Shachar Lovett. Bias vs structure of polynomials in largefields, and applications in effective algebraic geometry and coding theory, 2015. arXiv:1506 . .[Bri19] Jop Bri¨et. Subspaces of tensors with high analytic rank. arXiv , 2019. arXiv:1908 . .[BT12] Boris Bukh and Jacob Tsimerman. Sum–product estimates for rational functions. Proc.Lond. Math. Soc. (3) , 104(1):1–26, 2012. arXiv:1002 . , doi:10 . .[CLVW18] Matthias Christandl, Angelo Lucia, P´eter Vrana, and Albert H. Werner. Tensornetwork representations from the geometry of entangled states, 2018. arXiv:1809 . .[CVZ18] Matthias Christandl, P´eter Vrana, and Jeroen Zuiddam. Universal points in theasymptotic spectrum of tensors (extended abstract). In Proceedings of the 50th AnnualACM SIGACT Symposium on the Theory of Computing (STOC 2018) , 2018. arXiv:1709 . , doi:10 . . .[CVZ19] Matthias Christandl, P´eter Vrana, and Jeroen Zuiddam. Barriers for Fast MatrixMultiplication from Irreversibility. In Proceedings of the 34th ComputationalComplexity Conference (CCC 2019) , volume 137, pages 26:1–26:17, 2019. arXiv:1812 . , doi:10 . . CCC . . .[DKL14] Zeev Dvir, J´anos Koll´ar, and Shachar Lovett. Variety evasive sets. Comput.Complexity , 23(4):509–529, 2014. doi:10 . .[EG17] Jordan S. Ellenberg and Dion Gijswijt. On large subsets of F nq with no three-termarithmetic progression. Ann. of Math. (2) , 185(1):339–343, 2017. doi:10 . . . . . .[FGL19] Yuval Filmus, Konstantin Golubev, and Noam Lifshitz. High dimensional Hoffmanbound and applications in extremal combinatorics, 2019. arXiv:1911 . .[FJ05] Michael D. Fried and Moshe Jarden. Field arithmetic , volume 11 of
Ergebnisse derMathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys inMathematics . Springer-Verlag, Berlin, second edition, 2005.[GS] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for researchin algebraic geometry. Available at . math . uiuc . edu/Macaulay2/ .21GW11] W. T. Gowers and J. Wolf. Linear forms and higher-degree uniformity for functionson F np . Geom. Funct. Anal. , 21(1):36–69, 2011. doi:10 . .[Har92] Joe Harris. Algebraic geometry: a first course , volume 133. Springer Science &Business Media, 1992.[Jan18] Oliver Janzer. Low analytic rank implies low partition rank for tensors, 2018. arXiv:1809 . .[Jan20] Oliver Janzer. Polynomial bound for the partition rank vs the analytic rank of tensors. Discrete Anal. , 2020. arXiv:1902 . .[Koi97] Pascal Koiran. Randomized and deterministic algorithms for the dimension ofalgebraic varieties. In Proceedings 38th Annual Symposium on Foundations ofComputer Science , pages 36–45, Oct 1997. doi:10 . . . .[LG14] Fran¸cois Le Gall. Powers of tensors and fast matrix multiplication. In Proceedings ofthe 39th International Symposium on Symbolic and Algebraic Computation (ISSAC2014) , pages 296–303, 2014. arXiv:1401 . , doi:10 . . .[Lov19] Shachar Lovett. The analytic rank of tensors and its applications. Discrete Anal. , 2019. arXiv:1806 . , doi:10 . .[LW54] Serge Lang and Andr´e Weil. Number of points of varieties in finite fields. Amer. J.Math. , 76:819–827, 1954. doi:10 . .[Mil19] Luka Mili´cevi´c. Polynomial bound for partition rank in terms of analytic rank. Geom. Funct. Anal. , 29(5):1503–1530, 2019. arXiv:1902 . , doi:10 . .[NS17] Eric Naslund and Will Sawin. Upper bounds for sunflower-free sets. Forum Math.Sigma , 5:e15, 2017. arXiv:1606 . , doi:10 . . . .[Sag17] Sage. SageMath, the Sage Mathematics Software System (Version 8.1) , 2017. .[Str87] Volker Strassen. Relative bilinear complexity and matrix multiplication.
J. ReineAngew. Math. , 375/376:406–443, 1987. doi:10 . . . . .[Str88] Volker Strassen. The asymptotic spectrum of tensors. J. Reine Angew. Math. ,384:102–152, 1988. doi:10 . . . . .[Str91] Volker Strassen. Degeneration and complexity of bilinear maps: some asymptoticspectra. J. Reine Angew. Math. , 413:127–180, 1991. doi:10 . . . . .[Tao16] Terence Tao. A symmetric formulation of the Croot-Lev-Pach-Ellenberg-Gijswijtcapset bound. https://terrytao . wordpress . com , 2016. URL: https://terrytao . wordpress . com/2016/05/18/a-symmetric-formulation-of-the-croot-lev-pach-ellenberg-gijswijt-capset-bound .[vdW91] B. L. van der Waerden. Algebra. Vol. I . Springer-Verlag, New York, 1991. doi:10 .1007/978-1-4612-4420-2