Geometric Complexity Theory: a high level

overview.

K V Subrahmanyam

C. M. I.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 1 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 2 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 3 / 71
 Functions and Algorithms
Functions considered in Complexity theory:
Number Functions f : Zn 7→ Z
I The determinant of a matrix.
I Optimal value of an LP.
Decision functions f : Zn 7→ {0, 1}.
I Is the determinant of a matrix zero?
I Is the optimum of an LP greater than a certain value?
I Is there an assignment of values to the variables of a Boolean
formula so that the formula is true?
In general consider a family of functions - example detm .

An Algorithm is a program which computes a family of functions.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 4 / 71
 Measures
Every problem for which we need to design an algorithm comes with:
The input instance
I The matrix whose determinant is to be computed.
I The specific Boolean formula.
Input size
I for detm , the size, m2 , of the matrix.
I Also the number of digits needed to write an entry of the matrix
What is important?
The resources consumed by the algorithm to solve the problem.
I Time taken by the algorithm.
I How many digits are needed to store each intermediate value;
how many intermediate values need to be stored.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 5 / 71
 Computing with formulas

Let p(X1 , . . . , Xn ) be a polynomial. A formula is a particular way of

writing it using + and ∗.

formula = formula ∗ formulakformula + formula

Formula size: The number of ∗ and + operations.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 6 / 71
 Computing with formulas

Let p(X1 , . . . , Xn ) be a polynomial. A formula is a particular way of

writing it using + and ∗.

formula = formula ∗ formulakformula + formula

Formula size: The number of ∗ and + operations.

Example
a3 − b 3 = (a − b) ∗ (a2 + a ∗ b + b2 )
Van-der-Monde (λ1 , . . . , λn ) = Πi 6=j (λi − λj )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 6 / 71
 The permanent function
M = (mij ), a square n × n matrix Then
X
Permn (M) = Π1≤i ≤n mi ,σ(i )
σ∈Sn

X
Detn (M) = Π1≤i ≤n (−1)l(σ) mi ,σ(i )
σ∈Sn

Question
Does Permn have a formula of size polynomially bounded in n?

If NOT, what is the proof?

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 7 / 71
 The complexity theory of detm and Permn

The class of functions which can be computed in polynomial

time in input size - P. Intuitively, what can be considered
feasibly computable.
detm is easy to compute - The standard Gaussian elimination for
example shows feasibility. A combinatorial algorithm (Meena,
Vinay) - allows us to compute it by a polynomial number of
computers running in poly (log n)time.
The permanent is believed to be hard. Intuitively, among the
hardest among functions f for which:
I There is an expression for f which involves only positive
coeffecients.
I Each term of f is computable in polynomial time.
I Number of terms is at most exponential in input size.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 8 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 9 / 71
G a group. V a vector space.
ρ : G 7→ GL(V )
I V is a representation of G .
W ⊆ V a subspace of V is a subrepresentation.
I ρ(g ) · w ∈ W , ∀g ∈ G , ∀w ∈ W .
Irreducible representation.
I No proper non-trivial subrepresentation.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 10 / 71
G a group. V a vector space.
ρ : G 7→ GL(V )
I V is a representation of G .
W ⊆ V a subspace of V is a subrepresentation.
I ρ(g ) · w ∈ W , ∀g ∈ G , ∀w ∈ W .
Irreducible representation.
I No proper non-trivial subrepresentation.
G reductive.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 10 / 71
G a group. V a vector space.
ρ : G 7→ GL(V )
I V is a representation of G .
W ⊆ V a subspace of V is a subrepresentation.
I ρ(g ) · w ∈ W , ∀g ∈ G , ∀w ∈ W .
Irreducible representation.
I No proper non-trivial subrepresentation.
G reductive.
I Every finite dimensional representation is a direct sum of
irreducible representations.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 10 / 71
G a group. V a vector space.
ρ : G 7→ GL(V )
I V is a representation of G .
W ⊆ V a subspace of V is a subrepresentation.
I ρ(g ) · w ∈ W , ∀g ∈ G , ∀w ∈ W .
Irreducible representation.
I No proper non-trivial subrepresentation.
G reductive.
I Every finite dimensional representation is a direct sum of
irreducible representations.
I V = ⊕λ mλ Vλ (G )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 10 / 71
G a group. V a vector space.
ρ : G 7→ GL(V )
I V is a representation of G .
W ⊆ V a subspace of V is a subrepresentation.
I ρ(g ) · w ∈ W , ∀g ∈ G , ∀w ∈ W .
Irreducible representation.
I No proper non-trivial subrepresentation.
G reductive.
I Every finite dimensional representation is a direct sum of
irreducible representations.
I V = ⊕λ mλ Vλ (G )
I λ - labels of irreducible representations of G - building blocks in
the representation theory of reductive groups.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 10 / 71
 Examples of representations of GLn (C)

G acting on an Cn by A · v = A ∗ v .
I The zero vector is left fixed. A subrepresentation.
I Ireducible representation.
G acting on the vector space V of n × n matrices,
A · X = AXA−1 .
I The vector space of scalar matrices is left invariant.
I Not irreducible.
X = (xi ,j ) an n × n matrix of indeterminates. G acts on the
vector space of functions C[xi ,j ] by A · xi ,j = k=n
P
k=1 ak,i xk,j .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 11 / 71
 Representations of GLn (C),Sd
GL(n) – the general linear group over C
Sd – the symmetric group on d letters
Partition – a decreasing sequence of positive integers

Pictorially, partition (5, 2, 2) is represented as

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 12 / 71
 Representations of GLn (C),Sd
GL(n) – the general linear group over C
Sd – the symmetric group on d letters
Partition – a decreasing sequence of positive integers

Pictorially, partition (5, 2, 2) is represented as

GL(n), Sd are reductive

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 12 / 71
 Representations of GLn (C),Sd
GL(n) – the general linear group over C
Sd – the symmetric group on d letters
Partition – a decreasing sequence of positive integers

Pictorially, partition (5, 2, 2) is represented as

GL(n), Sd are reductive
Irreducible polynomial representations of GL(n) are
parameterized by partitions with at most n rows - denoted
Vλ (GLn )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 12 / 71
 Representations of GLn (C),Sd
GL(n) – the general linear group over C
Sd – the symmetric group on d letters
Partition – a decreasing sequence of positive integers

Pictorially, partition (5, 2, 2) is represented as

GL(n), Sd are reductive
Irreducible polynomial representations of GL(n) are
parameterized by partitions with at most n rows - denoted
Vλ (GLn )
Irreducible representations of Sd are parameterized by partitions
of d - denoted Wλ (Sd )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 12 / 71
 The irreducible representations of GLn
Z - an n × n variable matrix.
C[Z ] - ring of polynomial functions in the entries of Z - a
representation of GLn .
I (σ·f )(Z ) = f (Z σ)
Semi-standard tableau T -
I

1 1 3 3 4
2 3
3 5
C
T is the productofdeterminants ofminors such as
x1,1 x1,2 x1,3 x1,1 x1,3 x1,5
x2,1 x2,2 x2,3 , x2,1 x2,3 x2,5 , x1,3 , x1,3 , x1,4
x3,1 x3,2 x3,3 x3,1 x3,3 x3,5

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 13 / 71
 The irreducible representations of GLn
Z - an n × n variable matrix.
C[Z ] - ring of polynomial functions in the entries of Z - a
representation of GLn .
I (σ·f )(Z ) = f (Z σ)
Semi-standard tableau T -
I

1 1 3 3 4
2 3
3 5
C
T is the productofdeterminants ofminors such as
x1,1 x1,2 x1,3 x1,1 x1,3 x1,5
x2,1 x2,2 x2,3 , x2,1 x2,3 x2,5 , x1,3 , x1,3 , x1,4
x3,1 x3,2 x3,3 x3,1 x3,3 x3,5
Theorem
Vλ is the subrepresentation spanned by CT , T-semi standard.
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 13 / 71
 Irreducible representations of Sd
C[x1 , · · · , xd ] - polynomials in n variables - a representation of
Sd .
I (σ · f ) = f( xσ(1) , · · · , xσ(d) ).
Standard tableau T -
I

1 4 5 7 9
2 6
3 8
I fT is the product of discriminant of columns

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 14 / 71
 Irreducible representations of Sd
C[x1 , · · · , xd ] - polynomials in n variables - a representation of
Sd .
I (σ · f ) = f( xσ(1) , · · · , xσ(d) ).
Standard tableau T -
I

1 4 5 7 9
2 6
3 8
I fT is the product of discriminant of columns
I Πi <j,i ,j∈{1,2,3} (xi − xj ) Πi <j,i ,j∈{4,6,8} (xi − xj ) x(5) x(7) x(9)

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 14 / 71
 Irreducible representations of Sd
C[x1 , · · · , xd ] - polynomials in n variables - a representation of
Sd .
I (σ · f ) = f( xσ(1) , · · · , xσ(d) ).
Standard tableau T -
I

1 4 5 7 9
2 6
3 8
I fT is the product of discriminant of columns
I Πi <j,i ,j∈{1,2,3} (xi − xj ) Πi <j,i ,j∈{4,6,8} (xi − xj ) x(5) x(7) x(9)

Theorem
Wλ - the subrepresentation spanned by fT , T-standard.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 14 / 71
 Tensor Products
V a representation of G .
W a representation of G .
V ⊗ W a representation of G :
I σ(v ⊗ w ) = (σ · v ) ⊗ (σ · w )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 15 / 71
 Tensor Products
V a representation of G .
W a representation of G .
V ⊗ W a representation of G :
I σ(v ⊗ w ) = (σ · v ) ⊗ (σ · w )

Question
Find an explicit decomposition of the tensor product in terms of
irreducible representations of G .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 15 / 71
 Littlewood-Richardson coefficients
G = GLn ((C )).
γ
Find the multiplicity, cα,β , of the irreducible representation of
shape γ in Vα ⊗ Vβ .
γ
I Vα ⊗ Vβ = ⊕γ cα,β Vγ

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 16 / 71
 Littlewood-Richardson coefficients
G = GLn ((C )).
γ
Find the multiplicity, cα,β , of the irreducible representation of
shape γ in Vα ⊗ Vβ .
γ
I Vα ⊗ Vβ = ⊕γ cα,β Vγ

LR-rule:
I γ = (4, 3, 3, 2), α = (2, 2, 1), β = (3, 2, 2)

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 16 / 71
 Littlewood-Richardson coefficients
G = GLn ((C )).
γ
Find the multiplicity, cα,β , of the irreducible representation of
shape γ in Vα ⊗ Vβ .
γ
I Vα ⊗ Vβ = ⊕γ cα,β Vγ

LR-rule:
I γ = (4, 3, 3, 2), α = (2, 2, 1), β = (3, 2, 2)
I If α ≤ γ, form LR skew-tableau with content β:

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 16 / 71
 Littlewood-Richardson coefficients
G = GLn ((C )).
γ
Find the multiplicity, cα,β , of the irreducible representation of
shape γ in Vα ⊗ Vβ .
γ
I Vα ⊗ Vβ = ⊕γ cα,β Vγ

LR-rule:
I γ = (4, 3, 3, 2), α = (2, 2, 1), β = (3, 2, 2)
I If α ≤ γ, form LR skew-tableau with content β:
I

1 1
2
2 3
1 3

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 16 / 71
 Littlewood-Richardson coefficients
G = GLn ((C )).
γ
Find the multiplicity, cα,β , of the irreducible representation of
shape γ in Vα ⊗ Vβ .
γ
I Vα ⊗ Vβ = ⊕γ cα,β Vγ

LR-rule:
I γ = (4, 3, 3, 2), α = (2, 2, 1), β = (3, 2, 2)
I If α ≤ γ, form LR skew-tableau with content β:
I

1 1
2
2 3
1 3

Theorem
γ
cα,β = # LR skew-tableau of shape γ\α with content β

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 16 / 71
 Complexity theory implications of LR

γ
cα,β in #P, like permanent - it has a positive formula

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 17 / 71
 Complexity theory implications of LR

γ
cα,β in #P, like permanent - it has a positive formula
It is among the hardest such functions - computing this is #P
complete - as hard as permanent.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 17 / 71
 Complexity theory implications of LR

γ
cα,β in #P, like permanent - it has a positive formula
It is among the hardest such functions - computing this is #P
complete - as hard as permanent.
If we could compute this fast, then we would also be able to
compute the Permanent fast!

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 17 / 71
 Complexity theory implications of LR

γ
cα,β in #P, like permanent - it has a positive formula
It is among the hardest such functions - computing this is #P
complete - as hard as permanent.
If we could compute this fast, then we would also be able to
compute the Permanent fast!
Theorem, GCT III
Checking non-zeroness of LR coeff is in P

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 17 / 71
 Complexity theory implications of LR

γ
cα,β in #P, like permanent - it has a positive formula
It is among the hardest such functions - computing this is #P
complete - as hard as permanent.
If we could compute this fast, then we would also be able to
compute the Permanent fast!
Theorem, GCT III
Checking non-zeroness of LR coeff is in P

The precursor to Saturated linear programming.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 17 / 71
 Kronecker coefficients
G = Sd :
γ
Find the multiplicity, kα,β , of the irreducible representation of
shape γ in Wα ⊗ Wβ .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 18 / 71
 Kronecker coefficients
G = Sd :
γ
Find the multiplicity, kα,β , of the irreducible representation of
shape γ in Wα ⊗ Wβ .
γ
I Wα ⊗ Wβ = ⊕γ kα,β Wγ

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 18 / 71
 Kronecker coefficients
G = Sd :
γ
Find the multiplicity, kα,β , of the irreducible representation of
shape γ in Wα ⊗ Wβ .
γ
I Wα ⊗ Wβ = ⊕γ kα,β Wγ
No rule akin to LR rule known.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 18 / 71
 Kronecker coefficients
G = Sd :
γ
Find the multiplicity, kα,β , of the irreducible representation of
shape γ in Wα ⊗ Wβ .
γ
I Wα ⊗ Wβ = ⊕γ kα,β Wγ
No rule akin to LR rule known.
Question
Does the Kronecker coefficient belong to #P?

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 18 / 71
 Kronecker coefficients
G = Sd :
γ
Find the multiplicity, kα,β , of the irreducible representation of
shape γ in Wα ⊗ Wβ .
γ
I Wα ⊗ Wβ = ⊕γ kα,β Wγ
No rule akin to LR rule known.
Question
Does the Kronecker coefficient belong to #P?

Conjecture
γ
kα,β is in #P.
γ
non-zeroness of kα,β is in P.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 18 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 19 / 71
 Affine varieties
V = Cn , X = (x1 , x2, . . . , xn ) be the coordinates of V .
An affine algebraic set, Z ⊆ V , is the zero set of a collection of
polynomials in C[X ] = C[x1 , . . . , xn ].
Irreducible - if it is not the union of two proper affine algebraic
sets - Affine variety
Coordinate ring C[Z ] - C[X ]/I (Z ) where I (Z ) is the set of all
polynomial functions vanishing on Z .
Elements of C[Z ] - polynomial functions on Z .
Examples
I X axis union Y axis is the zero set of xy = 0 in C[x, y ].
I The parabola is the zero set of y 2 = x in C[x, y ].

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 20 / 71
 Projective Varieties
def
P n−1 = P(V ), the projective space of lines in V through the
origin.
The homogeneous coordinate ring of P(V ) is defined to be
C[X ].
A projective algebraic set Y ⊆ V is the set of zeros of
homogeneous functions on V .
The affine cone Ŷ ⊆ V is the union of lines in V corresponding
to points in Y .
The homogeneous coordinate ring R(Y ) = C[X ]/I (Y ).
Elements of R[Y ] - homogeneous functions on the cone over Y .
Degree d component of R[Y ], is the space of homogeneous
polynomials of degree d on Y .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 21 / 71
 G -varieties
Let V be a representation of a reductive group G .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 22 / 71
 G -varieties
Let V be a representation of a reductive group G .
I Then G takes a line through the origin in V to another line
through the origin in V .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 22 / 71
 G -varieties
Let V be a representation of a reductive group G .
I Then G takes a line through the origin in V to another line
through the origin in V .
I Get an action of G on P(V ).

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 22 / 71
 G -varieties
Let V be a representation of a reductive group G .
I Then G takes a line through the origin in V to another line
through the origin in V .
I Get an action of G on P(V ).
I Get a representation of G on C[X ]

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 22 / 71
 G -varieties
Let V be a representation of a reductive group G .
I Then G takes a line through the origin in V to another line
through the origin in V .
I Get an action of G on P(V ).
I Get a representation of G on C[X ]
I (σ · f )(X ) = f (σ −1 X ).

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 22 / 71
 G -varieties
Let V be a representation of a reductive group G .
I Then G takes a line through the origin in V to another line
through the origin in V .
I Get an action of G on P(V ).
I Get a representation of G on C[X ]
I (σ · f )(X ) = f (σ −1 X ).
Y ⊆ P(V ) is a G -variety if its ideal I (Y ) ⊆ C[X ] is a
subrepresentation of G .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 22 / 71
 G -varieties
Let V be a representation of a reductive group G .
I Then G takes a line through the origin in V to another line
through the origin in V .
I Get an action of G on P(V ).
I Get a representation of G on C[X ]
I (σ · f )(X ) = f (σ −1 X ).
Y ⊆ P(V ) is a G -variety if its ideal I (Y ) ⊆ C[X ] is a
subrepresentation of G .
C[X ]/I (Y ) - a representation of G .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 22 / 71
 G -varieties
Let V be a representation of a reductive group G .
I Then G takes a line through the origin in V to another line
through the origin in V .
I Get an action of G on P(V ).
I Get a representation of G on C[X ]
I (σ · f )(X ) = f (σ −1 X ).
Y ⊆ P(V ) is a G -variety if its ideal I (Y ) ⊆ C[X ] is a
subrepresentation of G .
C[X ]/I (Y ) - a representation of G .
For p ∈ Y , σ(p) is in Y - G moves points of Y around.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 22 / 71
 Orbit closures as special G -varieties
v ∈ P(V ), a point and Gv the orbit of v .

Gv = {gv |g ∈ G }
def
H = Gv = {g ∈ G |gv = v }
∆V [v ] = Gv ⊆ P(V ), the orbit closure of v .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 23 / 71
 Orbit closures as special G -varieties
v ∈ P(V ), a point and Gv the orbit of v .

Gv = {gv |g ∈ G }
def
H = Gv = {g ∈ G |gv = v }
∆V [v ] = Gv ⊆ P(V ), the orbit closure of v .

Example: The Grassmanian

Let Vλ the irreducible GLn module given by shape λ. Take the point
vλ. Then the orbit of vλ is closed, and is isomorphic to G /P. (Very
well studied in algebraic geometry)

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 23 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 24 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 25 / 71
 Valiant’s reduction
Theorem
If p(X1 , . . . , Xn ) has a formula of size m/2 then there is a matrix A of
size 2m × 2m, with entries being constants (from the underlying
field), or variables with p(X1 , . . . , Xn ) = detm (A).

Lets homogenize this construction:

Add an extra variable X0 .
Let p m (X0 , X1 , . . . , Xn ) be the degree m homogenization of p.
Homogenize Aij using X0 .
Then p m (X0 , . . . , Xn ) = detm (A).

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 26 / 71
 The hom reduction
Let X = (X1 , . . . , Xr ). For two forms f , g ∈ Symd (X ) we say we say
f hom g if f (X ) = g (AX ), where A ∈ gl (X ).
If permn has a formula of size m/2 -
Consider the space of all m × m matrices. For m > n we construct a
new function, permnm ∈ Symm (X ).

Let X 0 be the principal

n × n sub matrix of X . n X’

permnm (X ) = m
m−n
xmm permn (X 0 ).
X

Thus permn has been inserted into Symm (X ), of which detm (X ) is a

special element. permnm hom detm

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 27 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 28 / 71
 Group actions
Let V = Symm (X ).
Recall GL(X ) action on V : If f hom g then f = g (µ · X ), so:
(σ · f )(X ) = f (σ −1 X ).
Two notions: If µ is full rank, then f is in
Orbit: O(g ) = the GL(X ) orbit of g .
{σ · g |σ ∈ GL(X )}
If not, then µ is approximated
The projective orbit by elements of GL(X ).
closure:
In either case,
∆V (g ) = O(g )
f hom g =⇒ f ∈ ∆(g )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 29 / 71
 A faithful algebraization
If permn has a formula of size m/2 then permnm ∈ ∆(detm ).
On the other hand if permnm (X ) ∈ ∆(detm ) then for all  > 0,
there is a a σ ∈ GL(X ) such that σ · detm approximates permnm
to .
Gives a poly time approximation algorithm for permanent
Recall permn is #P-complete.

Approach
To show permn has no formula of size nc /2 it suffices to show that
c
permnn 6∈ ∆(detnc )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 30 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 31 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 32 / 71
 Group theoretic varieties, and orbit closure
membership
To determine if a form f belongs to ∆(g ) (assuming both live in the
same space), is in generalhopeless!
∆(detnc ) - the class variety associated to NC 2 is group theoretic.
c
∆(permnn ) - the extended #P variety is also group theoretic.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 33 / 71
 Group theoretic varieties, and orbit closure
membership
To determine if a form f belongs to ∆(g ) (assuming both live in the
same space), is in generalhopeless!
∆(detnc ) - the class variety associated to NC 2 is group theoretic.
c
∆(permnn ) - the extended #P variety is also group theoretic.
This is why we expect this problem to be tractable!

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 33 / 71
 Rich stabilizers
The stabilizer Gdetm of the The stabilizer of the
determinant in G = GL(m2 ) permanent in GL(n2 ).
I X 7→ AXB, I X 0 7→ PX 0 Q, P, Q
A, B ∈ GL(m); permutation matrices
I X 7→ X T in GL(n).
detm ∈ Symm (m2 ) I X 0 7→ D1 X 0 D2 , D1 , D2
determined by its diagonal matrices in
stabilizer. GL(n).
I X 0 7→ X 0T
permn ∈ Symn (n2 ) is
determined by its
stabilizer.
The embedding of Gdet is (almost) the natural embedding
GL(Cm ) × GL(Cm ) 7→ GL(Cm ⊗ Cm )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 34 / 71
 Consequences of permanent having poly-sized
formulas
Facts
If permnm ∈ ∆(detm ) then ∆(permnm ) ⊆ ∆(det).

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 35 / 71
 Consequences of permanent having poly-sized
formulas
Facts
If permnm ∈ ∆(detm ) then ∆(permnm ) ⊆ ∆(det).
Let R(n, m2 ) denote the coordinate ring of ∆(permnm ) and S the
coordinate ring of ∆(det). Then R d ,→ S d for all d .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 35 / 71
 Consequences of permanent having poly-sized
formulas
Facts
If permnm ∈ ∆(detm ) then ∆(permnm ) ⊆ ∆(det).
Let R(n, m2 ) denote the coordinate ring of ∆(permnm ) and S the
coordinate ring of ∆(det). Then R d ,→ S d for all d .

Recall: Both R and S are G = GL(m2 ) representations.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 35 / 71
 Consequences of permanent having poly-sized
formulas
Facts
If permnm ∈ ∆(detm ) then ∆(permnm ) ⊆ ∆(det).
Let R(n, m2 ) denote the coordinate ring of ∆(permnm ) and S the
coordinate ring of ∆(det). Then R d ,→ S d for all d .

Recall: Both R and S are G = GL(m2 ) representations.

def
So every irreducible representation Vλ = Vλ (G ) that occurs in
Rd as a subrepresentation, also occurs in Sd as a
subrepresentation.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 35 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 36 / 71
 Obstructions as witnesses
Definition
We say Vλ is an obstruction for the pair n, m2 and the pair
(perm, det) in degree d if it occurs in Rd and not in Sd .

Conjecture
a
An obstruction for n, m2 and the pair (perm, det) exists if m = 2n ,
for a small constant a > 0 as n → ∞. There exists such an
obstruction of a small degree mb , b > 0, a large constant.

The specification of an obstruction is given in the form of its label λ.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 37 / 71
 Why such obstructions?

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 38 / 71
 Why such obstructions?
Algebraic groups are completely determined by their
representations

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 38 / 71
 Why such obstructions?
Algebraic groups are completely determined by their
representations
The class varieties are essentially determined by their associated
triples.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 38 / 71
 Why such obstructions?
Algebraic groups are completely determined by their
representations
The class varieties are essentially determined by their associated
triples.
I #P by Gpermn ,→ GL(n2 ) ,→ GL(V )
I NC 2 by Gdetm ,→ GL(m2 ) ,→ GL(W )

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 38 / 71
 Why such obstructions?
Algebraic groups are completely determined by their
representations
The class varieties are essentially determined by their associated
triples.
I #P by Gpermn ,→ GL(n2 ) ,→ GL(V )
I NC 2 by Gdetm ,→ GL(m2 ) ,→ GL(W )
A witness for the non-existence of the embedding ought to be
present in the representation-theoretic datum, assuming
#P 6= NC 2 .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 38 / 71
Representation theory of homogeneous spaces,
GL(V )/H
Suppose H1 , H2 are reductive subgroups of a reductive group
G = GL(Cl ).
Assume H2 is not a conjugate of H1 .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 39 / 71
Representation theory of homogeneous spaces,
GL(V )/H
Suppose H1 , H2 are reductive subgroups of a reductive group
G = GL(Cl ).
Assume H2 is not a conjugate of H1 .
Fact
I G /H1 , G /H2 are affine algebraic varieties
I G /H1 cannot be embedded in G /H2 and vice versa.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 39 / 71
Representation theory of homogeneous spaces,
GL(V )/H
Suppose H1 , H2 are reductive subgroups of a reductive group
G = GL(Cl ).
Assume H2 is not a conjugate of H1 .
Fact
I G /H1 , G /H2 are affine algebraic varieties
I G /H1 cannot be embedded in G /H2 and vice versa.

Question Is there a representation theoretic obstruction?

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 39 / 71
Representation theory of homogeneous spaces,
GL(V )/H
Suppose H1 , H2 are reductive subgroups of a reductive group
G = GL(Cl ).
Assume H2 is not a conjugate of H1 .
Fact
I G /H1 , G /H2 are affine algebraic varieties
I G /H1 cannot be embedded in G /H2 and vice versa.

Question Is there a representation theoretic obstruction?

Peter-Weyl theorem implies,

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 39 / 71
Representation theory of homogeneous spaces,
GL(V )/H
Suppose H1 , H2 are reductive subgroups of a reductive group
G = GL(Cl ).
Assume H2 is not a conjugate of H1 .
Fact
I G /H1 , G /H2 are affine algebraic varieties
I G /H1 cannot be embedded in G /H2 and vice versa.

Question Is there a representation theoretic obstruction?

Peter-Weyl theorem implies,
I C[G ] = ⊕λ Vλ ⊗ Vλ .
I C[G /H] = ⊕λ Vλ ⊗ VλH

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 39 / 71
 The work of Larsen and Pink
Vλ (G ) is an obstruction for the pair (G /H1 , G /H2 ) is it contains an
H1 invariant and does not contain an H2 invariant.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 40 / 71
 The work of Larsen and Pink
Vλ (G ) is an obstruction for the pair (G /H1 , G /H2 ) is it contains an
H1 invariant and does not contain an H2 invariant.
Theorem:Larsen,Pink
Such an obstruction always exists.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 40 / 71
 The work of Larsen and Pink
Vλ (G ) is an obstruction for the pair (G /H1 , G /H2 ) is it contains an
H1 invariant and does not contain an H2 invariant.
Theorem:Larsen,Pink
Such an obstruction always exists.

The embeddibility problem in GCT is a generalization of this situation

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 40 / 71
 The work of Larsen and Pink
Vλ (G ) is an obstruction for the pair (G /H1 , G /H2 ) is it contains an
H1 invariant and does not contain an H2 invariant.
Theorem:Larsen,Pink
Such an obstruction always exists.

The embeddibility problem in GCT is a generalization of this situation

Conjecture, GCT II
Let V = Symm (x11 , x12 , . . . , xmm ). Let Π be the set of G = GL(m2 )
submodules of C[V ] whose duals do not contain a Gdet invariant. Let
X (Π) ⊆ P(V ) be the zero set of forms in Π. Then X (Π) = ∆(detm ).

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 40 / 71
 What is known
Theorem, GCTII
There is a dense open neighbourhood U ⊆ P(V ) of the orbit of the
determinant such that ∆(detm ) ∩ U = X (Π) ∩ U.

Assuming the conjecture and the belief that the permanent cannot be
approximated infinitely closely by circuits of poly-logarithmic depth,
Theorem
Obstructions do exist. GCT II.
We need to show obstructions exist, Unconditionally.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 41 / 71
 What is known
Theorem, GCTII
There is a dense open neighbourhood U ⊆ P(V ) of the orbit of the
determinant such that ∆(detm ) ∩ U = X (Π) ∩ U.

Assuming the conjecture and the belief that the permanent cannot be
approximated infinitely closely by circuits of poly-logarithmic depth,
Theorem
Obstructions do exist. GCT II.
We need to show obstructions exist, Unconditionally.
Important to note that existence of obstructions depends upon
the special nature of detm and so, also of the variety ∆(detm ).

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 41 / 71
 What is known
Theorem, GCTII
There is a dense open neighbourhood U ⊆ P(V ) of the orbit of the
determinant such that ∆(detm ) ∩ U = X (Π) ∩ U.

Assuming the conjecture and the belief that the permanent cannot be
approximated infinitely closely by circuits of poly-logarithmic depth,
Theorem
Obstructions do exist. GCT II.
We need to show obstructions exist, Unconditionally.
Important to note that existence of obstructions depends upon
the special nature of detm and so, also of the variety ∆(detm ).
The conjecture will not be true for varieties arising out of many
other NC 2 -complete forms - the class variety may not be group
theoretic.
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 41 / 71
 The Kronecker problem and NC 2 vs #P
We need to understand which GL(m2 ) modules contain the
stabilizer of the determinant form.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 42 / 71
 The Kronecker problem and NC 2 vs #P
We need to understand which GL(m2 ) modules contain the
stabilizer of the determinant form.
Recall that the stabilizer of the determinant form detm in
projective space, is GL(m) × GL(m) ,→ GL(m2 ) via
(A, B) 7→ A ⊗ B.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 42 / 71
 The Kronecker problem and NC 2 vs #P
We need to understand which GL(m2 ) modules contain the
stabilizer of the determinant form.
Recall that the stabilizer of the determinant form detm in
projective space, is GL(m) × GL(m) ,→ GL(m2 ) via
(A, B) 7→ A ⊗ B.
Now irreducible representations of GL(m) × GL(m) are of the
form Vα ⊗ Vβ , α, β, shapes, with at most m rows.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 42 / 71
 The Kronecker problem and NC 2 vs #P
We need to understand which GL(m2 ) modules contain the
stabilizer of the determinant form.
Recall that the stabilizer of the determinant form detm in
projective space, is GL(m) × GL(m) ,→ GL(m2 ) via
(A, B) 7→ A ⊗ B.
Now irreducible representations of GL(m) × GL(m) are of the
form Vα ⊗ Vβ , α, β, shapes, with at most m rows.
Given an GL(m2 ) module of shape γ, we need to understand the
γ
multiplicity kα,β - this is exactly the Kronecker problem using
Schur Weyl duality.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 42 / 71
 Implementing the flip
Flip
Non existence of algorithms reduced to existence of
representation theoretic obstructions

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 43 / 71
 Implementing the flip
Flip
Non existence of algorithms reduced to existence of
representation theoretic obstructions
n c
How to prove existence of an obstruction for permm ?

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 43 / 71
 Implementing the flip
Flip
Non existence of algorithms reduced to existence of
representation theoretic obstructions
n c
How to prove existence of an obstruction for permm ?
A probabilistic approach - choose a random label λ(n) of high
degree randomly and show that it is an obstruction with high
probability

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 43 / 71
 Implementing the flip
Flip
Non existence of algorithms reduced to existence of
representation theoretic obstructions
n c
How to prove existence of an obstruction for permm ?
A probabilistic approach - choose a random label λ(n) of high
degree randomly and show that it is an obstruction with high
probability
In the context of P vs NP this will be naturalizable

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 43 / 71
 Implementing the flip
Flip
Non existence of algorithms reduced to existence of
representation theoretic obstructions
n c
How to prove existence of an obstruction for permm ?
A probabilistic approach - choose a random label λ(n) of high
degree randomly and show that it is an obstruction with high
probability
In the context of P vs NP this will be naturalizable
The GCT VI approach: GO FOR EXPLICIT OBSTRUCTIONS!

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 43 / 71
 Implementing the flip
Flip
Non existence of algorithms reduced to existence of
representation theoretic obstructions
n c
How to prove existence of an obstruction for permm ?
A probabilistic approach - choose a random label λ(n) of high
degree randomly and show that it is an obstruction with high
probability
In the context of P vs NP this will be naturalizable
The GCT VI approach: GO FOR EXPLICIT OBSTRUCTIONS!
This will overcome the naturalizability barrier!

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 43 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 44 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 45 / 71
 The GCT VI hypothesis
Hypothesis, GCT VI
The following problems belong to P.
Verification There is a poly (m2, n, < d >, < λ >) algorithm for
deciding, given m, n, d , λ, if Vλ is an obstruction of degree d for
(n, m2 ) and the pair (permn , detm ).
a
Explicit construction of obstructions Suppose m = 2n for a
small constant a > 0. Then, for every n → ∞, a label λ(n) of
an obstruction for (n, m2 ) and the pair (permn , detm ) can be
constructed in time poly (m), thereby proving the existence of an
obstruction for every such n, m2 .
Discovery of obstructionsThere exists a poly (n, m) algorithm for
deciding, if there exists an obstruction for (n, m2 ) and the pair
(permn , detm ), and for constructing the label of one, if it exists.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 46 / 71
 Why the hypothesis?

GCT VI: Proving that the hypothesis is true, will be the hardest
step in the flip.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 47 / 71
 Why the hypothesis?

GCT VI: Proving that the hypothesis is true, will be the hardest
step in the flip.
The hypothesis is believable because of the special nature of the
perm function and, so, of ∆(permnm ).

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 47 / 71
 Why the hypothesis?

GCT VI: Proving that the hypothesis is true, will be the hardest
step in the flip.
The hypothesis is believable because of the special nature of the
perm function and, so, of ∆(permnm ).
There are many #P forms - probabilistic method would show so.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 47 / 71
 Why the hypothesis?

GCT VI: Proving that the hypothesis is true, will be the hardest
step in the flip.
The hypothesis is believable because of the special nature of the
perm function and, so, of ∆(permnm ).
There are many #P forms - probabilistic method would show so.
We could have taken any such form, say h, and argued that h is
not computable by a poly-sized formula.How?

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 47 / 71
 Why the hypothesis?

GCT VI: Proving that the hypothesis is true, will be the hardest
step in the flip.
The hypothesis is believable because of the special nature of the
perm function and, so, of ∆(permnm ).
There are many #P forms - probabilistic method would show so.
We could have taken any such form, say h, and argued that h is
not computable by a poly-sized formula.How?
By exhibiting an obstruction. SAME IDEA, take ∆(hnm ) and
show this is not embeddable in ∆(detm ).

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 47 / 71
 Why the hypothesis?

GCT VI: Proving that the hypothesis is true, will be the hardest
step in the flip.
The hypothesis is believable because of the special nature of the
perm function and, so, of ∆(permnm ).
There are many #P forms - probabilistic method would show so.
We could have taken any such form, say h, and argued that h is
not computable by a poly-sized formula.How?
By exhibiting an obstruction. SAME IDEA, take ∆(hnm ) and
show this is not embeddable in ∆(detm ).
Of course, such obstructions exist in plenty! thanks to the
special nature of determinant - , however

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 47 / 71
 The P-barrier thesis
It is unlikely that the hypothesis holds for a general h.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 48 / 71
 The P-barrier thesis
It is unlikely that the hypothesis holds for a general h.

Using currently available techniques, Gröbner basis, etc, we only

get an algorithm which is double exponential in m, or triple
exponential in n to verify, given λ, if Vλ is an obstruction for the
pair (h, detm )!
Unlikely we will be able to do better, given general lower bounds
for construction of Gröbner basis
The P-barrier thesis
For any approach towards P 6= NP to be viable and
non-naturalizeable, at least the problem of verifying an obstruction
should be in P.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 48 / 71
 Crossing the P-barrier
Theorem
Hypothesis (a) is true assuming certain mathematical positivity
hypothesis.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 49 / 71
 Crossing the P-barrier
Theorem
Hypothesis (a) is true assuming certain mathematical positivity
hypothesis.

Theorem, GCT III

Checking non-zeroness of LR coefficients can be done in polynomial
γ
time - that is, given partitions α, β, γ,and n, checking if cα,β is
non-zero can be done in time polynomial in n, < α >, < β >, < γ >.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 49 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 50 / 71
LR coefficients as integral points in a polytope
γ
Recall cα,β = # of LR skew-tableau of shape γ/α with content
β.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 51 / 71
LR coefficients as integral points in a polytope
γ
Recall cα,β = # of LR skew-tableau of shape γ/α with content
β.
rji (T ), 1 ≤ i ≤ n, 1 ≤ j ≤ n, denote the number of j’s in i -th
row of T . Then

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 51 / 71
LR coefficients as integral points in a polytope
γ
Recall cα,β = # of LR skew-tableau of shape γ/α with content
β.
rji (T ), 1 ≤ i ≤ n, 1 ≤ j ≤ n, denote the number of j’s in i -th
row of T . Then
Non-neg rji ≥ 0. Shape For i ≤ n, αi + j rji = γi .
P

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 51 / 71
LR coefficients as integral points in a polytope
γ
Recall cα,β = # of LR skew-tableau of shape γ/α with content
β.
rji (T ), 1 ≤ i ≤ n, 1 ≤ j ≤ n, denote the number of j’s in i -th
row of T . Then
Non-neg rji ≥ 0. Shape For i ≤ n, αi + j rji = γi .
P

Content For j ≤ n, i rji = βj .

P

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 51 / 71
LR coefficients as integral points in a polytope
γ
Recall cα,β = # of LR skew-tableau of shape γ/α with content
β.
rji (T ), 1 ≤ i ≤ n, 1 ≤ j ≤ n, denote the number of j’s in i -th
row of T . Then
Non-neg rji ≥ 0. Shape For i ≤ n, αi + j rji = γi .
P

Content For j ≤ n, i rji = βj .

P

Tableau No k < j occurs in row i + 1 of T below a j or a higher

integer in row i of T .
X X
αi +1 + rki +1 ≤ αi + rki 0
k≤j k 0 <j

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 51 / 71
LR coefficients as integral points in a polytope
γ
Recall cα,β = # of LR skew-tableau of shape γ/α with content
β.
rji (T ), 1 ≤ i ≤ n, 1 ≤ j ≤ n, denote the number of j’s in i -th
row of T . Then
Non-neg rji ≥ 0. Shape For i ≤ n, αi + j rji = γi .
P

Content For j ≤ n, i rji = βj .

P

Tableau No k < j occurs in row i + 1 of T below a j or a higher

integer in row i of T .
X X
αi +1 + rki +1 ≤ αi + rki 0
k≤j k 0 <j

LR constraint
I rji = 0, i < j.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 51 / 71
LR coefficients as integral points in a polytope
γ
Recall cα,β = # of LR skew-tableau of shape γ/α with content
β.
rji (T ), 1 ≤ i ≤ n, 1 ≤ j ≤ n, denote the number of j’s in i -th
row of T . Then
Non-neg rji ≥ 0. Shape For i ≤ n, αi + j rji = γi .
P

Content For j ≤ n, i rji = βj .

P

Tableau No k < j occurs in row i + 1 of T below a j or a higher

integer in row i of T .
X X
αi +1 + rki +1 ≤ αi + rki 0
k≤j k 0 <j

LR constraint
I rji = 0, i < j.
i0 i0
P P
i 0 <i rj ≤ i 0 <i rj−1 .
I

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 51 / 71
 Consequences of the Knutson-Tao Saturation
conjecture
Theorem, GCT III
The polytope P described by the linear inequalities above has an
integer point iff it is non-empty

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 52 / 71
 Consequences of the Knutson-Tao Saturation
conjecture
Theorem, GCT III
The polytope P described by the linear inequalities above has an
integer point iff it is non-empty

Proof.
P of the form Ar ≤ b, where b is homogeneous in α, β, γ.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 52 / 71
 Consequences of the Knutson-Tao Saturation
conjecture
Theorem, GCT III
The polytope P described by the linear inequalities above has an
integer point iff it is non-empty

Proof.
P of the form Ar ≤ b, where b is homogeneous in α, β, γ.
P has rational vertices.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 52 / 71
 Consequences of the Knutson-Tao Saturation
conjecture
Theorem, GCT III
The polytope P described by the linear inequalities above has an
integer point iff it is non-empty

Proof.
P of the form Ar ≤ b, where b is homogeneous in α, β, γ.
P has rational vertices.
If P non-empty, qP non-empty for all q integer, q > 0.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 52 / 71
 Consequences of the Knutson-Tao Saturation
conjecture
Theorem, GCT III
The polytope P described by the linear inequalities above has an
integer point iff it is non-empty

Proof.
P of the form Ar ≤ b, where b is homogeneous in α, β, γ.
P has rational vertices.
If P non-empty, qP non-empty for all q integer, q > 0.
qγ
However number of integral points in qP is precisely cqα,qβ

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 52 / 71
 Consequences of the Knutson-Tao Saturation
conjecture
Theorem, GCT III
The polytope P described by the linear inequalities above has an
integer point iff it is non-empty

Proof.
P of the form Ar ≤ b, where b is homogeneous in α, β, γ.
P has rational vertices.
If P non-empty, qP non-empty for all q integer, q > 0.
qγ
However number of integral points in qP is precisely cqα,qβ
qγ
cqα,qβ is non-zero!
qγ γ
Theorem, Knutson, Tao: cqα,qβ 6= 0 =⇒ cα,β 6= 0.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 52 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 53 / 71
Generalized Littlewood Richardson coefficients
Given weights α, β, γ of a semi-simple Lie algebra G the
generalized LR coefficient is the multiplicity of Vγ in Vα ⊗ Vβ ,
where Vµ denotes the irreducible representation of G with
highest weight µ.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 54 / 71
Generalized Littlewood Richardson coefficients
Given weights α, β, γ of a semi-simple Lie algebra G the
generalized LR coefficient is the multiplicity of Vγ in Vα ⊗ Vβ ,
where Vµ denotes the irreducible representation of G with
highest weight µ.

Berenstein-Zelevinsky associate a polytope for the coefficient

γ
Cα,β .
γ
Cα,β is exactly the number of integral points in this polytope.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 54 / 71
Generalized Littlewood Richardson coefficients
Given weights α, β, γ of a semi-simple Lie algebra G the
generalized LR coefficient is the multiplicity of Vγ in Vα ⊗ Vβ ,
where Vµ denotes the irreducible representation of G with
highest weight µ.

Berenstein-Zelevinsky associate a polytope for the coefficient

γ
Cα,β .
γ
Cα,β is exactly the number of integral points in this polytope.
No saturation type theorem known for these coefficients.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 54 / 71
 A mathematical positivity hypothesis
nγ γ
Let Cnα,nβ be the stretching function associated with Cα,β . Then
the stretching function is a quasi-polynomial with period at most
two i.e.there exists two polynomials C1 , C2 such that

nγ C1 (n), n odd
Cnα,nβ =
C2 (n), n even

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 55 / 71
 A mathematical positivity hypothesis
nγ γ
Let Cnα,nβ be the stretching function associated with Cα,β . Then
the stretching function is a quasi-polynomial with period at most
two i.e.there exists two polynomials C1 , C2 such that

nγ C1 (n), n odd
Cnα,nβ =
C2 (n), n even

Positivity hypothesis, De Loera, McAllister

The quasi-polynomial is positive - i.e. the coefficients of C1 , C2 are
positive.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 55 / 71
 Polynomiality under the hypothesis
Theorem GCT V
Assume G is simple of type B,C,D. Under the positivity hypothesis
the following are equivalent.
γ
1 Cα,β ≥1
nγ
2 There exists an odd integer n such that Cnα,nβ ≥ 1.
3 The BZ polytope contains a rational point whose denominators
are all odd.
4 The affine span of the BZ polytope contains a rational point
whose denominators are all odd.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 56 / 71
 Polynomiality under the hypothesis
Theorem GCT V
Assume G is simple of type B,C,D. Under the positivity hypothesis
the following are equivalent.
γ
1 Cα,β ≥1
nγ
2 There exists an odd integer n such that Cnα,nβ ≥ 1.
3 The BZ polytope contains a rational point whose denominators
are all odd.
4 The affine span of the BZ polytope contains a rational point
whose denominators are all odd.
Proof.
(2) =⇒ (1): If for some n odd, nP has an integral point then

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 56 / 71
 Polynomiality under the hypothesis
Theorem GCT V
Assume G is simple of type B,C,D. Under the positivity hypothesis
the following are equivalent.
γ
1 Cα,β ≥1
nγ
2 There exists an odd integer n such that Cnα,nβ ≥ 1.
3 The BZ polytope contains a rational point whose denominators
are all odd.
4 The affine span of the BZ polytope contains a rational point
whose denominators are all odd.
Proof.
(2) =⇒ (1): If for some n odd, nP has an integral point then C1 ()
should be a non-zero polynomial. Its coefficients are all positive. So
C1 (1) is also non-zero!
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 56 / 71
 Saturated quasi-polynomials
A function f (n) is called a quasi-polynomial if there exist l
polynomials fj (n), 1 ≤ j ≤ l , such that f (n)=fj (n) if n = j mod
l . Here l is supposed to be the smallest such integer, and is
called the period of f (n).
The smallest j, 1 ≤ j ≤ l , such that fj is not identically zero, is
called the index of the quasi-polynomial.
We say that the quasi-polynomial f (n) which is not identically
zero is saturated , if f (index(f )) 6= 0.
Positivity =⇒ Saturation. If each fj is guaranteed to have
positive coefficients, f is saturated.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 57 / 71
 Saturated Integer programming
Integer Programming: Does the polytope {x|Ax ≤ b} contain
an integer point?

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 58 / 71
 Saturated Integer programming
Integer Programming: Does the polytope {x|Ax ≤ b} contain
an integer point?
In general very hard - NP-complete.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 58 / 71
 Saturated Integer programming
Integer Programming: Does the polytope {x|Ax ≤ b} contain
an integer point?
In general very hard - NP-complete.
Associate to a polytope P a function;
fP (n) = #integral points in the polytope nP

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 58 / 71
 Saturated Integer programming
Integer Programming: Does the polytope {x|Ax ≤ b} contain
an integer point?
In general very hard - NP-complete.
Associate to a polytope P a function;
fP (n) = #integral points in the polytope nP

Theorem Stanley
fP () is a quasi-polynomial.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 58 / 71
 Saturated Integer programming
Integer Programming: Does the polytope {x|Ax ≤ b} contain
an integer point?
In general very hard - NP-complete.
Associate to a polytope P a function;
fP (n) = #integral points in the polytope nP

Theorem Stanley
fP () is a quasi-polynomial.

Theorem, GCT VI
The index of the quasi-polynomial fP associated to a polytope
(specified by a separation oracle), can be determined in oracle
polynomial time.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 58 / 71
 Saturated Integer programming
Integer Programming: Does the polytope {x|Ax ≤ b} contain
an integer point?
In general very hard - NP-complete.
Associate to a polytope P a function;
fP (n) = #integral points in the polytope nP

Theorem Stanley
fP () is a quasi-polynomial.

Theorem, GCT VI
The index of the quasi-polynomial fP associated to a polytope
(specified by a separation oracle), can be determined in oracle
polynomial time.
Saturated integer programming has a polynomial time algorithm.
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 58 / 71
 Questions associated with quasi-polynomials

Given a quasi-polynomial g (n) we may ask:

Is g convex? Is there a polytope P whose Ehrhart polynomial
fP (n) coincides with g (n).
Is g positive?
Is g saturated?
def P
Does G (t) = n g (n)t n , have a reduced, positive form?

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 59 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 60 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 61 / 71
 Kirillov’s conjecture
γ
Associate with the Kronecker constant kα,β the following
stretching function.
γ nγ
k̃α,β (n) = knα,β

Conjecture,Kirillov
γ nγ
t n is a rational function.
P
The generating function, Kα,β (t) = knα,β
n≥0

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 62 / 71
 Kirillov’s conjecture verified, GCT VI
Theorem
γ
(a) Rationality The generating function Kα,β (t) is rational.
γ
(b) Quasi-polynomiality The stretching function k̃α,β (n) is a
quasi-polynomial function of n.
γ
(c) There exist graded, normal C-algebras S = S(kα,β ) = ⊕n Sn , and
γ
T = T (kα,β ) = ⊕n Tn such that:
1 The schemes spec(S) and spec(T ) are normal and have rational
singularities.
2 T = S GLn (C) , the subring of GLn (C)-invariants in S.
γ
3 The quasi-polynomial k̃α,β (n) is the Hilbert function of T . In
other words, it is the Hilbert function of the homogeneous
coordinate ring of the projective scheme Proj(T ).
γ
(d) Positivity The rational function Kα,β (t) can be expressed in a
positive form:
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 63 / 71
 Kirillov .....

γ h0 + h1 t + · · · + hd t d
Kα,β (t) =
Q a(j) )d(j)
,
j (1 − t
P
where a(j)’s and d (j)’s are positive integers, j d (j) = d + 1, where
γ
d is the degree of the quasi-polynomial k̃α,β (n), h0 = 1, and hi ’s are
nonnegative integers.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 64 / 71
 Kirillov .....

γ h0 + h1 t + · · · + hd t d
Kα,β (t) =
Q a(j) )d(j)
,
j (1 − t
P
where a(j)’s and d (j)’s are positive integers, j d (j) = d + 1, where
γ
d is the degree of the quasi-polynomial k̃α,β (n), h0 = 1, and hi ’s are
nonnegative integers.
The rational function we get from the theorem is not unique.
Depends upon a homogeneous system of parameters we choose.
We have

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 64 / 71
 Kirillov .....

γ h0 + h1 t + · · · + hd t d
Kα,β (t) =
Q a(j) )d(j)
,
j (1 − t
P
where a(j)’s and d (j)’s are positive integers, j d (j) = d + 1, where
γ
d is the degree of the quasi-polynomial k̃α,β (n), h0 = 1, and hi ’s are
nonnegative integers.
The rational function we get from the theorem is not unique.
Depends upon a homogeneous system of parameters we choose.
We have
Hypothesis, PH3
γ
For the lex least system of parameters, Kα,β (t) has a reduced positive
form, with max(ai ) bounded by a polynomial in the height of the
shapes α, β and γ.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 64 / 71
 Kirillov .....

γ h0 + h1 t + · · · + hd t d
Kα,β (t) =
Q a(j) )d(j)
,
j (1 − t
P
where a(j)’s and d (j)’s are positive integers, j d (j) = d + 1, where
γ
d is the degree of the quasi-polynomial k̃α,β (n), h0 = 1, and hi ’s are
nonnegative integers.
The rational function we get from the theorem is not unique.
Depends upon a homogeneous system of parameters we choose.
We have
Hypothesis, PH3
γ
For the lex least system of parameters, Kα,β (t) has a reduced positive
form, with max(ai ) bounded by a polynomial in the height of the
shapes α, β and γ.
γ
If so, then we will be able to conclude that k̃α,β is saturated.
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 64 / 71
 γ
Proof of quasipolynomiality of Kα,β
.
Theorem
Let R = ⊕d Rd be a graded ring with an action of a connected
reductive group H. Let Vπ be a fixed irreducible H module with label
π. Let sdπ denote the multiplicity of the Vπ in Rd .Take its stretching
function s̃dπ (n) to be the multiplicity of Vnπ in Rnd . Take the
generating function Sdπ (t) = n≥0 s̃dπ (n)t n . Then S satisfies the
P
claims of the theorem.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 65 / 71
 Proof
Consider the action of the cyclic group of order d with generator
k
ζ; it acts
P on R sending x ∈ Rk to ζ x. The invariant ring
B = n Rnd , for this action is normal and has rational
singularities.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 66 / 71
 Proof
Consider the action of the cyclic group of order d with generator
k
ζ; it acts
P on R sending x ∈ Rk to ζ x. The invariant ring
B = n Rnd , for this action is normal and has rational
singularities.
Let Vπ∗P= Vπ∗ . Consider the homogeneous coordinate ring
Cπ∗ = n Vnπ∗ , of the orbit of vπ∗ . This too is normal with
rational singularities.(It is a G /P)

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 66 / 71
 Proof
Consider the action of the cyclic group of order d with generator
k
ζ; it acts
P on R sending x ∈ Rk to ζ x. The invariant ring
B = n Rnd , for this action is normal and has rational
singularities.
Let Vπ∗P= Vπ∗ . Consider the homogeneous coordinate ring
Cπ∗ = n Vnπ∗ , of the orbit of vπ∗ . This too is normal with
rational singularities.(It is a G /P)
B ⊗ Cπ∗ is normal and spec(B ⊗ Cπ∗ ) has rational singularities.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 66 / 71
 Proof
Consider the action of the cyclic group of order d with generator
k
ζ; it acts
P on R sending x ∈ Rk to ζ x. The invariant ring
B = n Rnd , for this action is normal and has rational
singularities.
Let Vπ∗P= Vπ∗ . Consider the homogeneous coordinate ring
Cπ∗ = n Vnπ∗ , of the orbit of vπ∗ . This too is normal with
rational singularities.(It is a G /P)
B ⊗ Cπ∗ is normal and spec(B ⊗ Cπ∗ ) has rational singularities.
C∗ acts on this; x · (b ⊗ c) = (xb ⊗ x −1 c). The invariant ring
has rational singularities and S = ⊕Sn = ⊕n Rnd ⊗ Vnπ∗ . This is
an H module. The H invariants form a ring T . The multiplicity
of the trivial H module in Sn is precisely s̃dπ (n), and is the
dimension of Tn .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 66 / 71
 Proof
Consider the action of the cyclic group of order d with generator
k
ζ; it acts
P on R sending x ∈ Rk to ζ x. The invariant ring
B = n Rnd , for this action is normal and has rational
singularities.
Let Vπ∗P= Vπ∗ . Consider the homogeneous coordinate ring
Cπ∗ = n Vnπ∗ , of the orbit of vπ∗ . This too is normal with
rational singularities.(It is a G /P)
B ⊗ Cπ∗ is normal and spec(B ⊗ Cπ∗ ) has rational singularities.
C∗ acts on this; x · (b ⊗ c) = (xb ⊗ x −1 c). The invariant ring
has rational singularities and S = ⊕Sn = ⊕n Rnd ⊗ Vnπ∗ . This is
an H module. The H invariants form a ring T . The multiplicity
of the trivial H module in Sn is precisely s̃dπ (n), and is the
dimension of Tn .
Our desired generating function is the Hilbert series of this ring
T.
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 66 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 67 / 71
 Implications of the theorem - positivity
hypothesis
Hypothesis; PH1
The Kronecker stretching function is convex.
γ
For every (α, β, γ) there exists a polytope P=Pα,β ⊆ Rm such that:
(1) The Ehrhart quasi-polynomial of P coincides with the stretching
γ
quasi-polynomial k̃α,β (n).
(2) The dimension m of the ambient space, and hence the dimension
of P as well, are polynomial in the bit-lengths of α, β and γ.
(3) Whether a point x ∈ Rm lies in P can be decided in
poly (< α >, < β >, < γ >) time. That is, the membership problem
belongs to the complexity class P. If x does not lie in P, then this
algorithm outputs a hyper plane separating x from P.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 68 / 71
 Roadmap
1 Basics in Complexity theory, Algebraic Geometry
Complexity Theory
Representation theory
Algebraic geometry
2 Algebraizing the formula complexity question
Reduction
Geometry and class varieties
3 From lower bounds to obstructions
Geometry of class varieties is tractable
The first flip
4 Saturated Integer programming
Overcoming the Razbarov-Rudich barrier
Non-zeroness of LR-coeffs in poly time
Saturated functions
5 Implementing the flip for #P vs NC 2
Saturation of Kronecker coefficients
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 69 / 71
 γ
Polynomiality of kα,β under PH1,PH3

Theorem
γ
Under PH1, PH3, determining if kα,β is non-zero can be done in
polynomial time.

γ
Experimental evidence for PH3, and positivity of k̃α,β .

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 70 / 71
 γ
Polynomiality of kα,β under PH1,PH3

Theorem
γ
Under PH1, PH3, determining if kα,β is non-zero can be done in
polynomial time.

γ
Experimental evidence for PH3, and positivity of k̃α,β .

Current proofs of PH1 for type B, C , D are based on quantum

groups
The homogeneous coordinate rings of the canonical models
associated with the Littlewood-Richardson coefficients have
quantizations endowed with canonical bases Kashiwara and
Lusztig.
Positivity of the basis is also based on the Riemann hypothesis
over finite fields, and work of Beilinson, Bernstein, Deligne.
K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.
5/09/07 @CMI 70 / 71
 A plan to prove PH1, PH2, PH3

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 71 / 71
 A plan to prove PH1, PH2, PH3

1 Construct quantizations of the homogeneous coordinate rings of

the canonical models associated with these structural constants,

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 71 / 71
 A plan to prove PH1, PH2, PH3

1 Construct quantizations of the homogeneous coordinate rings of

the canonical models associated with these structural constants,
2 Show that they have positive bases as per PH0.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 71 / 71
 A plan to prove PH1, PH2, PH3

1 Construct quantizations of the homogeneous coordinate rings of

the canonical models associated with these structural constants,
2 Show that they have positive bases as per PH0.
3 Prove PH1 and SH (and, possibly, the stronger PH2, and 3 as
well) by a detailed analysis and study of these positive bases.

K V Subrahmanyam ( C. M. I. ) Geometric Complexity Theory: a high level overview.

5/09/07 @CMI 71 / 71