The Computational Complexity of Plethysm Coefficients

In two papers, B\"urgisser and Ikenmeyer (STOC 2011, STOC 2013) used an adaption of the geometric complexity theory (GCT) approach by Mulmuley and Sohoni (Siam J Comput 2001, 2008) to prove lower bounds on the border rank of the matrix multiplication tensor. A key ingredient was information about certain Kronecker coefficients. While tensors are an interesting test bed for GCT ideas, the far-away goal is the separation of algebraic complexity classes. The role of the Kronecker coefficients in that setting is taken by the so-called plethysm coefficients: These are the multiplicities in the coordinate rings of spaces of polynomials. Even though several hardness results for Kronecker coefficients are known, there are almost no results about the complexity of computing the plethysm coefficients or even deciding their positivity. In this paper we show that deciding positivity of plethysm coefficients is NP-hard, and that computing plethysm coefficients is #P-hard. In fact, both problems remain hard even if the inner parameter of the plethysm coefficient is fixed. In this way we obtain an inner versus outer contrast: If the outer parameter of the plethysm coefficient is fixed, then the plethysm coefficient can be computed in polynomial time. Moreover, we derive new lower and upper bounds and in special cases even combinatorial descriptions for plethysm coefficients, which we consider to be of independent interest. Our technique uses discrete tomography in a more refined way than the recent work on Kronecker coefficients by Ikenmeyer, Mulmuley, and Walter (Comput Compl 2017). This makes our work the first to apply techniques from discrete tomography to the study of plethysm coefficients. Quite surprisingly, that interpretation also leads to new equalities between certain plethysm coefficients and Kronecker coefficients.


Introduction
Geometric complexity theory (GCT) is an approach towards the separation of Valiant's algebraic complexity classes using algebraic geometry and representation theory. These ideas were introduced by Mulmuley and Sohoni in [MS01,MS08]. A first implementation of that framework proves lower bounds on the border rank of the matrix multiplication tensor [BI11,BI13b]. In these papers, Kronecker coefficients play a key role, as they are the multiplicities in the coordinate rings of spaces of tensors. While that line of research turns out as an interesting test case for GCT ideas, the ultimate goal remains to separate algebraic complexity classes, that is, certain sets of (families of) polynomials. Switching from tensors to polynomials, the role of Kronecker coefficients in [BI11,BI13b] is now taken by plethysm coefficients (see e.g. [BI18,Sec. 12.4(i)]). Plethysm coefficients are the main subject of "GCT7" [Mul07], and they also appear prominently in the GCT publications [KL14,Kum15,Bür16,BHI17].
The main subject of study in this paper are plethysm coefficients and, to a lesser extent, Kronecker coefficients. Both types of representation theoretic coefficients are not only important in GCT, but they are fundamental objects in algebraic combinatorics. Indeed, in 1999, Richard Stanley highlighted the quest for a combinatorial description for plethysm coefficients as Problem 9 in his survey on "outstanding open problems" in algebraic combinatorics [Sta99b]; finding a combinatorial interpretation for Kronecker coefficients is Problem 10 in the same paper 1 . Even though GCT asks to compare these (and related) coefficients (see e.g. [Ike12,Appendix]), both coefficients have been studied mostly independently. But this has started to change recently, as can be seen for example from the fact that the Fall 2016 Eastern Sectional Meeting of the American Mathematical Society held a special session on "Plethysm and Kronecker products in Representation Theory" and also from the 2020 Oberwolfach Mini-Workshop on "Kronecker, Plethysm, and Sylow Branching Coefficients and their applications to Complexity Theory". To the best of our knowledge the only papers that give inequalities between these two sorts of coefficients are [Man11, IP17, BIP19] 2 .
In GCT, both coefficients do not only serve as the multiplicities in the coordinate rings of spaces of tensors or polynomials, but they also appear as terms in nonnegative formulas that describe multiplicities of coordinate rings of important group orbits: The plethysm coefficients appear as terms in the nonnegative formulas for the multiplicities in the coordinate ring of the orbit of the product of variables ([Lan17, Sec. 9.2.3]), the power sum polynomial ( [IK19]), the unit tensor ([Ike19, Sec. 3.7]), and the permanent polynomial ([BLMW11, Eq. (5.5.2)]), whereas the Kronecker coefficients appear as terms in the nonnegative formulas for the multiplicities in the coordinate ring of the orbit of the determinant polynomial [BLMW11, Eq. (5.2.6)], the permanent polynomial ([BLMW11, Eq. (5.5.2)], both the plethysm coefficients and the Kronecker coefficients appear in this formula), the matrix multiplication tensor [BI11,Thm. 5.3], the unit tensor ([Ike19, Sec. 3.7], both the plethysm coefficients and the Kronecker coefficients appear in this formula), and the trace of the matrix power polynomial [GIP17, Thm 2.10]. In theoretical physics, the plethysm and Kronecker coefficients are closely related to the quantum marginal problem for indistinguishable and distinguishable particles, respectively, see e.g. [CHM07] and [CDKW14,Sec. 7], where the asymptotic positivity of the coefficients is studied.
While the #P-hardness of the computation of the Kronecker coefficient dates back to [BI08,BOR09], the recent paper [IMW17] proves the NP-hardness of deciding the positivity of the Kronecker coefficient. This was a setback for the GCT program, as it was originally conjectured that Kronecker positivity would be decidable in polynomial time, much in the same way as the well-known Littlewood-Richardson coefficients: Even though the computation of Littlewood-Richardson coefficients is #P-complete [Nar06], positivity of Littlewood-Richardson coefficients can be decided in polynomial time [DLM06] using linear programming (see also [MNS12]), and even with a combinatorial max-flow algorithm [BI13a]. Such a result would have made the search for obstructions (i.e., inequalities between coefficients) much easier.
In this paper we prove NP-hardness of deciding positivity of plethysm coefficients (Theorem 3.1) and #P-hardness of the computation of plethysm coefficients (Theorem 3.2). Not even the #Phardness of the computation of plethysm coefficients was known prior to our work. Indeed, our reduction is quite subtle compared to the one used to prove NP-hardness of deciding positivity of Kronecker coefficients [IMW17].

Structure of the Paper
Section 2 explains the necessary background from representation theory. Section 3 states our main hardness results and puts them in contrast to what is known about efficiently computable subcases. Section 4 is a first important step in our reduction: We show that the inner parameter of the plethysm coefficient can be fixed to be three. Section 5 translates an interesting subcase of the plethysm coefficient problem into a problem in discrete tomography. In Section 6 we use this tomography description to design a sequence of reductions that finally prove our hardness result. Sections 7 and 8 contain results of independent interest: Section 7 highlights new close connections to Kronecker coefficients, while Section 8 contains more combinatorial descriptions for the plethysm coefficients. Finally, Section 9 uses classical results to provide an algorithm that places the problem of computing plethysm coefficients in the complexity class GapP.

Preliminaries from Representation Theory
The definition of plethysm and Kronecker coefficients that we use requires some elementary algebraic geometry and representation theory, as can be found for example in the standard textbooks [Kra85,FH91,Sag01]. The necessary material can also for example be found in [Lan11,Ch. 6].
A composition λ is a list of nonnegative integers λ = (λ 0 , λ 1 , . . .), which we always treat as finite by omitting trailing zeros. If the entries are nonincreasing, then we call λ a partition. A partition λ is often presented as its Young diagram, which is a top-left justified array containing λ i boxes in the i-th row. For instance, (3, 1) is identified with the Young diagram . In that presentation it makes sense to term λ 0 the width of λ and the number of non-zero entries the height of λ. We refer to the number |λ| := i λ i of boxes in λ's Young diagram as the size of λ. If λ is of height 1, then we call λ a 1-row partition, and similarly, if λ is of width 1, then λ is a 1-column partition. Finally, let the transpose λ t of λ denote the partition obtained by exchanging rows and columns. For example, (3, 1) t = (2, 1, 1) is depicted as .
Let V be a finite-dimensional complex vector space. There is a canonical action of the general linear group GL(V ) on V given by gv := g(v) for all g ∈ GL(V ), v ∈ V . For a composition λ, if diag(α 0 , . . . , α k )v = α λ 0 0 · · · α λ k k v, then we say that v is a weight vector of weight λ. Note that for all matrices m ∈ gl(V ) := C dim V ×dim V we have that Id +εm ∈ GL(V ) for all small enough ε, where Id is the dim V × dim V identity matrix. We say that V admits an action of the Liealgebra gl(V ) defined by mv := lim ε→0 ε −1 ((Id +εm)v − v) for all m ∈ gl(V ), v ∈ V . For i < j, define the raising operator E i,j ∈ gl(V ) as the matrix with a single one at position (i, j) and zeros everywhere else. A weight vector v of weight λ is called a highest weight vector if v vanishes under the action of all raising operators. In that case, λ is guaranteed to be a partition. It is wellknown that the irreducible polynomial representations of GL(V ) are characterized by their unique (up to scale) highest weight vectors, which in turn, are indexed by partitions of height at most dim V . The irreducible GL(V )-representation of type ν is called the Weyl module and denoted by S ν V . The Weyl module S ν V is constructed as follows. Given any GL(V )-representation W (for example W = V ), there is a natural linear GL(V )-action on its m-th tensor power m W given by g(w 1 ⊗ w 2 ⊗ · · · ⊗ w m ) := (gw 1 ) ⊗ (gw 2 ) ⊗ · · · ⊗ (gw m ) and linear continuation. We define S ν W to be the image of the Schur functor S ν that maps W to its m-th tensor power m W and then projects onto the ν-th isotypic component via where χ ν is the character of the irreducible representation of type ν of the symmetric group S m . For example, if ν = (m), then S ν V is the vector space of symmetric tensors of order m, which we denote by Sym m V . If ν = (1, 1, . . . , 1) is of size m, then S ν V is the vector space of alternating tensors of order m, which we denote by m V . A plethysm is the application of S µ to S ν V . The discussion above shows that this space The nonnegative integers p λ (µ, ν) are called general plethysm coefficients. The case where µ = (n) and ν = (m) is of special interest, so we write a λ (n, m) := p λ ((n), (m)), which is the multiplicity of type λ in Sym n Sym m V . The coefficient a λ (n, m) is often called the plethysm coefficient in the literature. We introduce the dual plethysm coefficient b λ (n, m) as another special case of general plethysm coefficients, defined by i.e., by restricting the Young diagrams µ and ν to a single row (in case of Sym) or a single column (in case of ). By some well-known dualities [Car90,Man98,MM15] we can relate the above decompositions with their respective analogues after exchanging Sym and .

Fact ([Car90, Man98, MM15])
If m is odd, then: whereas in case that m is even:

Main Results: Complexity of Plethysm Coefficients
In order to establish hardness of computing general plethysm coefficients p λ (µ, ν), it is of course sufficient to focus on the interesting special cases a λ (n, m) and b λ (n, m). To that end, we study the following problems from a complexity theoretic perspective 3 .
Problem (Plethysm) Given a partition λ and two integers n and m, output a λ (n, m).
Problem (PlethysmPositivity) Given a partition λ and two integers n and m, output "accept" if a λ (n, m) > 0, and output "reject" otherwise.
Problem (DualPlethysm) Given a partition λ and two integers n and m, output b λ (n, m).
Problem (DualPlethysmPositivity) Given a partition λ and two integers n and m, output "accept" if b λ (n, m) > 0, and output "reject" otherwise.
We remark that the computation of a λ (n, 2) and b λ (n, 2) is possible in polynomial time 4 , so we need to deal with cases m ≥ 3 only. We will be particularly interested in cases where one of the parameters m or n is fixed. For each of the problems above we therefore define versions where m, respectively n, is fixed by appending (inner=m), respectively (outer=n), to the problem name. For example, Plethysm(inner=3) is the problem "given λ and n, compute a λ (n, 3)".
We remark that PlethysmPositivity is called the "zero locus of plethysm coefficients" in [KM16a], where this problem is labeled as "highly nontrivial". Our main result makes this intuitive statement precise. We prove that PlethysmPositivity is NP-hard: 3.1 Theorem Plethysm(inner=m) and DualPlethysm(inner=m) are NP-hard for any fixed m ≥ 3. In particular, PlethysmPositivity and DualPlethysmPositivity are NP-hard.
By a thorough analysis of the proof, show that Plethysm is #P-hard: 3.2 Theorem Plethysm(inner=m) and DualPlethysm(inner=m) are #P-hard for any fixed m ≥ 3. In particular, Plethysm and DualPlethysm are #P-hard.
Considering Theorem 3.1 and Theorem 3.2, we obtain an interesting complexity-theoretic contrast if we fix the outer parameter instead of the inner, as seen in the following proposition, whose proof can be extracted from [KM16a] 3.3 Proposition Plethysm(outer=n) and DualPlethysm(outer=n) are in P for any fixed n. More generally, computing p λ (µ, (m)) can be done in polynomial time if |µ| is fixed.
This striking difference of complexities between Theorem 3.1 (if m is fixed, then even deciding positivity of a λ (n, m) is NP-hard) and Proposition 3.3 (if n is fixed, then even computing the exact value of a λ (n, m) can be done in polynomial time) could be interpreted as an explanation why it is 3 In all problems the inputs can be encoded in binary or in unary, as it makes no difference for our results below. The hardness results still hold for the unary encoding, while the polynomial-time algorithms still work for the binary encoding.
4 By [Sta99a, Wey03, Proposition 2.3.8, Proposition 2.3.9], Sym n Sym 2 V decomposes multiplicity-free as λ∈A(n) S λ V for A(n) = {λ of size 2n : each λi is even}; similarly, we have n Sym 2 V = λ∈B(n) S λ V , where B(n) = {λ of size 2n : if λi ≥ i, then λi = λ t i + 1}. Testing membership λ ∈ An and λ ∈ Bn can both be performed in polynomial time in |λ|. 5 [KM16a] shows that, for any fixed partition µ, the function (λ0, . . . , λ |µ|−1 ) → p λ (µ, (m)) has a constant size arithmetic formula (with modular arithmetic) whose inputs are the λi; see the appendix of the arXiv version [KM16b] for a good exposition. considered to be much more difficult to obtain exact results about a λ (n, m) for n m compared to the case where m n, see e.g. [Wei90, TC92, KM16a, KM18, KL19] for evidence. In addition, the #P-hardness result leads to a completeness result for the complexity class GapP-that is, the class of all counting problems implementable by a nondeterministic polynomialtime Turing machine, where the output is interpreted as the number accepting computation paths minus the number of rejecting computation paths [FFK94].
3.4 Theorem Plethysm, DualPlethysm and the problem of computing general plethysm coefficients are GapP-complete.
The proof of Theorem 3.4 is provided in Section 9.
4 Fixing the Inner Parameter to m = 3 To ultimately prove Theorems 3.1 and 3.2, we preliminarily demonstrate that proving hardness of the case m = 3 is enough: 4.1 Lemma Let m be fixed. Given a partition λ encoded in unary, we can compute partitions π and π in polynomial time so that a λ (n, m) = b π (n, m + 1) and b λ (n, m) = a π (n, m + 1).
In particular, provided that deciding positivity of both a λ (n, 3) and b λ (n, 3) is already NP-hard, deciding positivity of a plethysm coefficient a λ (n, m) or b λ (n, m) is NP-hard as well for each fixed m ≥ 3. In order to prove Lemma 4.1, we leverage another fact:

Fact ([MM15])
Let ν be a partition of n, let λ be a partition of n · m of width at most n and let π be the partition after prepending a row of width n to λ. Then the multiplicity of the irreducible component S λ V in S ν m V equals the multiplicity of the irreducible component S π V in S ν m+1 V .
Proof of Lemma 4.1. We show the first claim only; the second part is analogous. Let a partition λ of n · m be given. Fact 2.2 yields that a λ (n, m) equals the multiplicity of the irreducible representation S λ t V in the decomposition of S ν m V , where ν is the Young Diagram consisting of a single row of width n if m is even and ν equals a column of length n otherwise. If λ t is of width > n, then we immediately deduce that a λ (n, m) = 0. So assume that λ t is of width ≤ n and construct a partition π t by prepending a row of width n to λ t . From Fact 4.2, we infer that the multiplicity of S π t in the decomposition of S ν m+1 equals a λ (n, m) and by applying Fact 2.2 once more, we finally obtain that a λ (n, m) = b π (n, m + 1).
We are left to deal with the case m = 3.

Combinatorial Descriptions for Plethysm Coefficients via Discrete Tomography
Our results in Section 3 are based on a new combinatorial description of the plethysm coefficient in a subcase, based on new combinatorial lower and upper bounds for the plethysm coefficient. We think that the lower and upper bound as well as the combinatorial description are of independent interest. By for all (x, y, z) ∈ P and all (x , y , z ) ∈ C with x ≤ x, y ≤ y, z ≤ z we have (x , y , z ) ∈ P . Analogously, a point set P in the closed cone is called a pyramid in the closed cone if for all (x, y, z) ∈ P and all (x , y , z ) ∈ C with x ≤ x, y ≤ y, z ≤ z we have (x , y , z ) ∈ P .
Let P be a point set in the open cone or closed cone. The sum-marginal S(P ) is the sequence of numbers defined via

Definition
Let λ be a partition of 3n. We define:

Theorem
Let λ be a partition of 3n. Then: Proof. Let X 0 , X 1 , . . . , X k−1 be an ordered basis of V = C k . There is a natural bijection between the points in C <k := [0, k − 1] 3 ∩ C and the monomials in Sym 3 V . It is easy to check that a monomial X x X y X z is of weight κ if and only the corresponding point (x, y, z) has sum-marginal κ (here, κ is not necessarily a partition). In the same way, we match (x, y, z) ∈ C <k := [0, k − 1] 3 ∩ C with elementary wedge products X x ∧ X y ∧ X z . (Since we are dealing with these objects modulo scalars, the sign changes caused by the wedge product shall not bother us here.) Now let λ be a partition of 3n. We associate each point set P ⊆ C <k with sum-marginal λ with a vector v P := notice that |λ| = 3n implies that P is of size n and thus indeed v P ∈ n Sym 3 V . Analogously, for point sets P ⊆ C <k with sum-marginal λ, we define vectors From the preceding paragraphs, it follows that v P is a weight vector of weight λ, and furthermore, for P ranging over all point subsets of C <k with sum-marginal λ, {v P } P forms a basis of the weight subspace of weight λ of n Sym 3 V . Recall that b + λ (n, 3) exactly counts the number of such sets P , thus b + λ (n, 3) equals the dimension of the weight-λ subspace. Moreover, since n Sym 3 V is a GL(V )-representation, the multiplicity b λ (n, 3) of λ in the decomposition of n Sym 3 V into irreducible GL(V )-representations matches the dimension of the linear subspace spanned by all highest weight vectors of weight λ. But each highest weight vector is in particular a weight vector and thus b λ (n, 3) ≤ b + λ (n, 3). In the same way, {w P } P forms a basis of the weight-λ t subspace of n 3 V , for P ranging over all point subsets of C <k with sum-marginal λ t . Therefore, a + λ (n, 3) equals the dimension of the weight-λ t subspace and hence upper-bounds the multiplicity of the irreducible component S λ t V in n 3 V . Recall that this multiplicity equals a λ (n, 3) by Fact 2.2, so we conclude a λ (n, 3) ≤ a + λ (n, 3). For the remainder of the proof, we can treat the symmetric (closed cone) and skew-symmetric (open cone) cases identically, so we omit the proof of a − λ (n, 3) ≤ a λ (n, 3) and verify only that b − λ (n, 3) ≤ b λ (n, 3). Let P ⊆ C <k be a pyramid in the closed cone with sum-marginal λ, |λ| = 3n. It suffices to show that v P is not only a weight vector, but in fact a highest weight vector. To prove this, we demonstrate that v P is annihilated by all raising operators E i,j ; recall that, for i < j, E i,j denotes the upper triangular matrix with a single 1 in the i-th row and the j-th column and zero entries anywhere else.
We claim that each summand vanishes individually. Focus, for some (x, y, z) ∈ P , on the summand In case that x = j, this term clearly equals zero, so assume x = j. Since i < j = x, the points (i, y, z) and (x, y, z) differ. But P is a pyramid and thus contains (i, y, z), too. Consequently, the monomial X i X y X z also occurs in the big wedge product on the right-hand side and therefore cancels with X i X y X z on the left-hand side. The other summands with i in place of y or z, respectively, can be shown to vanish by an analogous argumentation.
Applying the raising operator E i,j to v P has an intuitive geometric interpretation for P : Any summand as above arises from P by replacing some coordinate j of some point in P by i. Thereby that point is moved closer to the origin as i < j. Recall however, that pyramids are point sets not containing "holes", thus P is transformed by mapping two points onto each other. Here comes the alternating property of the wedge product into play by annihilating the summand.
When the lower and upper bound coincide, we get a combinatorial description of the (dual) plethysm coefficient. The following Proposition 5.5 gives a sufficient condition for when this happens. Let λ ∈ N [0,r] be a composition. We define the coordinate sum B(λ) via Let P be a point set in the open cone or closed cone. The coordinate sum B(P ) is defined as the coordinate sum of the sum-marginal B(P ) := B(S(P )), which is the same as the sum over all coordinates: Let ξ(i) the number of cardinality-1 point sets in the closed cone with coordinate sum exactly i. This is the same as the number of partitions of i of height ≤ 3, which is OEIS sequence A001399, with explicit formula round((i + 3) 2 /12). Let ξ(i) denote number of cardinality-1 point sets in the open cone with coordinate sum i. This is the same as the number of partitions µ of i of height 3 such that µ 1 , µ 2 , and µ 3 are pairwise distinct, which is OEIS sequence A069905 6 , with explicit formula round(i 2 /12). Now let ι(n) := min{ι : ι i=0 ξ(i) ≥ n} and analogously, let ι(n) := min{ι : ι i=0 ξ(i) ≥ n}. Finally, let β(n) := n i=1 ι(i) and analogously, let β(n) := n i=1 ι(i).

Lemma
There is no point set P of cardinality n in the closed cone that has B(P ) < β(n).
Analogously, there is no point set P of cardinality n in the open cone that has B(P ) < β(n).
Proof. This is seen by induction: If P of cardinality n in the open cone has B(P ) < β(n), then remove the point (x, y, z) with highest coordinate sum to obtain the point set P . Since for each ι there are only ξ(ι) many points ( , which means that P cannot exist by induction hypothesis and hence P cannot exist. A completely analogous argument works for the closed cone.

Proposition
If B(λ) = β(|λ|/3), then each point set in the closed cone with sum-marginal λ is a pyramid in the closed cone. In particular, if λ is a partition of 3n, according to Theorem 5.3, Proof. Assume for the sake of contradiction that there exists a point set P in the closed cone with sum-marginal S(P ) = λ (note that S(P ) = λ implies |P | = |λ|/3) such that P is not a pyramid in the closed cone. Then there exists (x, y, z) ∈ P and (x , y , z ) / ∈ P with x ≤ x, y ≤ y, z ≤ z. Replacing (x, y, z) in P with (x , y , z ) gives a point set P with B(P ) < B(P ). This is a contradiction to Lemma 5.4. A completely analogous argument works for the open cone.

Proof of Main Theorem 3.2
The reductions that we present are parsimonious polynomial-time reductions between counting problems. Such reductions automatically yield polynomial-time many-one reductions between the associated decision problems.

Counting and Decision Problems in Discrete Tomography
Let us start with some notations: Let P ⊆ N 3 be a finite set of points. By X i (P ) we address the number of elements in P with X-coordinate i. Similarly, Y i (P ) and Z i (P ) count the elements with Y -and Z-coordinate i, respectively. We call (X i (P )) i , (Y i (P )) i and (Z i (P )) i the X-, Yand Z-marginals of P ; these sequences are treated as finite by omitting trailing zeros. Recall the definition (5.1) of the sum-marginal S(P ) and observe that S(P ) = X(P ) + Y (P ) + Z(P ).
The intermediate 2-dimensional problems below operate on the grid in spirit, G r is two-dimensional although each entry of G r is parameterized by 3 coordinates (the so-called trilinear coordinates). This specific embedding of G r ⊆ N 3 plays a major role in the proof of Lemma 6.3.
2D-X-Ray has been identified as NP-hard (even NP-complete) before by Gardner, Gritzmann and Prangenberg [GGP99]. In fact, they prove that the counting version #2D-X-Ray is #Phard (even #P-complete). Keeping this in mind, we now continue to reduce #2D-X-Ray in a parsimonious way to computing a λ (n, 3) or b λ (n, 3), taking the detour over the other #X-Ray problems. Our reductions are illustrated in Figure 2.
Problem (Symmetric 2D-X-Ray and Skew-Symmetric 2D-X-Ray) Given λ ∈ N [0,r] , the Symmetric 2D-X-Ray problem is to decide if there exists a point set P ⊆ G r ∩ C with summarginal S(P ) = λ.
Analogously, given λ ∈ N [0,r] , the Skew-Symmetric 2D-X-Ray problem is to decide if there exists a point set P ⊆ G r ∩ C.
Finally, we introduce three-dimensional variants that differ from the above problems only in that P is taken from N 3 rather than G r : Problem (3D-X-Ray) Given µ, ν, ρ ∈ N [0,r] , the 3D-X-Ray problem is to decide if there exists a point set P ⊆ N 3 with X-, Y -and Z-marginals µ, ν and ρ, respectively.
3D-X-Ray is known to be NP-complete by [BDLG01], see Figure 2.
Problem (Symmetric 3D-X-Ray and Skew-Symmetric 3D-X-Ray) Given λ ∈ N [0,r] , the Symmetric 3D-X-Ray problem is to decide if there exists a point set P ⊆ C with sum-marginal λ.
Analogously, given λ ∈ N [0,r] , the Skew-Symmetric 3D-X-Ray problem is to decide if there exists a point set P ⊆ C with sum-marginal λ.

#P-hardness by [GGP99]
#2D-X-Ray [IMW17] Lemma 6.3 Lemma 6.2 Figure 2 The reductions that lead to #P-hardness proofs for the problems of computing Kronecker and plethysm coefficients. All depicted reductions are parsimonious and polynomialtime, so the the NP-hardness of all corresponding decision problems ultimately follows from the NP-hardness of 2D-X-Ray that was proved in [GGP99].
Our main focus in 3D will be the following promise problem version.

Reduction from #(Skew-)Symmetric Promise-3D-X-Ray to (Dual-)Plethysm
Recall that (Skew-)Symmetric Promise-3D-X-Ray instances λ need not be partitions in general. Still, we prove that any promise instance λ which is not a partition is trivially rejected: 6.1 Lemma Let λ be a composition, and assume there exists some pyramid P in the open or closed cone with sum-marginal λ. Then λ is a partition.

Reduction from #Symmetric 2D-X-Ray to #Skew-Symmetric Promise-3D-X-Ray
In this section we prove the following lemma.

Lemma
There exists a parsimonious polynomial-time reduction from #Symmetric 2D-X-Ray to #Symmetric Promise-3D-X-Ray.
Before we can prove Lemma 6.3 we make an observation (Lemma 6.4 below) of independent interest. We call P r := {(x, y, z) : x + y + z ≤ r, x > y > z} the complete pyramid in the open cone, and analogously P r := {(x, y, z) : x + y + z ≤ r, x ≥ y ≥ z} the complete pyramid in the closed cone. Note that for all r we have β(|P r |) = B(P r ). Moreover, β(n) is the piecewise linear interpolation between these values. Analogously, β(n) is the piecewise linear interpolation between β(|P r |) = B(P r ).
The proof of the following lemma is a careful adaption of a proof given by Brunetti, Del Lungo and Gérard [BDLG01] for the non-symmetric 3D-X-Ray problem. Roughly speaking, the trick is to embed the two-dimensional grid G r into the r-th layer of the complete pyramid.

Lemma
Let P ⊆ C be of size |P | = n. Then P has coordinate sum B(P ) = β(n) if and only if P r−1 ⊆ P ⊆ P r for some r.
Analogously, let P ⊆ C be of size |P | = n. Then P has coordinate sum B(P ) = β(n) if and only if P r−1 ⊆ P ⊆ P r for some r.
Proof. It is easy to see by a direct calculation that all point sets P r−1 ⊆ P ⊆ P r have B(P ) = β(n), since every point in P \ P r−1 has the same coordinate sum r (i.e., P \ P r−1 ⊆ G r ).
The other direction is intuitively straightforward by the following argument: Whenever P contains a "hole", then we can fill the hole by moving outer points closer to the origin. But in that we decrease the coordinate sum of P . More formally: Assume that P has minimal coordinate sum among all sets ⊆ C of the same size n. We pick r to be maximal, such that P r−1 ⊆ P and assume, for the sake of contradiction, that P ⊆ P r , i.e., there exists some (x, y, z) ∈ P \ P r . Furthermore, there exists some (x , y , z ) ∈ P r \ P , since otherwise P ⊆ P r which contradicts the way we chose r. We construct the set P := P ∪ {(x , y , z )} \ {(x, y, z)}; clearly |P | = |P | = n. The coordinate sum of P is bounded by Figure 3 Visualizes how to arrange the grids G r in three-dimensional space; the red-colored areas indicate the intersections with C (or C, by restricting to the interior). The essential step of the reduction is to embed a set P ⊆ G r into the top-most layer of the pyramid P := P r−1 ∪ P .
which is a contradiction to the definition of β(n). This proof works completely analogously for C.
The reduction now proceeds by embedding G r in the three-dimensional space. Any solution P of the given (Skew-)Symmetric 2D-X-Ray instance is interpreted as the r-th layer of the thereunder entirely filled pyramid P = P r−1 ∪ P ; Figure 3 might give some further intuition on how to view Symmetric 2D-X-Ray in three-dimensional space. We detail the construction in the following lemma: 6.5 Lemma Fix some r. Let λ ∈ N [0,r] be such that there exists n with | λ| = 3n and B( λ) = rn. (i.e., λ is not trivially rejected as a Symmetric 2D-X-Ray instance). Let λ := S(P r−1 ) + λ. Then B(λ) = β(|P r−1 | + n) and there is a one-to-one correspondence between point sets P ⊆ C with sum-marginal λ and point sets P ⊆ G r ∩ C with sum-marginal λ.
The same statements hold if instead we consider point sets in the open cone C, P r−1 and P r , and the Skew-Symmetric problems. In that case λ := S(P r−1 ) + λ.
After applying the obvious changes, the proof also works for the open cone C.
Proof of Lemma 6.3. Given a #Symmetric 2D-X-Ray instance λ, we first assert that | λ| is divisible by 3 and that B( λ i ) = rn holds for n = | λ|/3. If this step fails, output a trivially zero #Symmetric 3D-X-Ray instance. Otherwise, construct λ := S(P r−1 ) + λ and output λ as the corresponding #Symmetric 3D-X-Ray instance. Clearly, this algorithm runs in polynomial time. γ(G r ) Figure 4 Visualizes the reduction from 2D-X-Ray to Symmetric 2D-X-Ray. The lattice depicts the subdivision of the triangular grid G r into 13 regions per coordinate, each of width r . The intervals (r , 3r ), (4r , 9r ) and (10r , 13r ]-that is, the blocks of zeros in λ-are respectively indicated by gray-, blue-and red-colored areas; for instance, a point is colored gray if at least one of its coordinates lies in (r , 3r ).
Observe that the circled region equals the image of γ under G r , which is the only uncolored region intersecting the (closed) cone C.
Lemma 6.5 entirely proves the remaining goals: Since B(λ) = β(|P r−1 | + n), the promise is kept. Moreover, the one-to-one correspondence in Lemma 6.5 ensures that the reduction is correct and parsimonious. Again, the same proof works if one replaces Symmetric by Skew-Symmetric and C by C.
Reduction from #2D-X-Ray to #Symmetric 2D-X-Ray The following Lemma 6.6 is the first known result that connects X-Ray problems with their symmetric counterparts.
Proof. Let µ , ν , ρ ∈ N [0,r ] be a given #2D-X-Ray instance. We start to construct the corresponding #Symmetric 2D-X-Ray instance by choosing r := 13r and We claim that the #Symmetric 2D-X-Ray instance λ has the same number of solutions as the #2D-X-Ray instance (µ , ν , ρ ). Indeed, consider the following map γ that maps points in G r to points in G r ∩ C: γ(x, y, z) = (x + 9r , y + 3r , z) We lift γ from points to point sets in the canonical way and abuse notation by calling the resulting map γ again. We claim that γ maps bijectively which then finishes the proof. An illustration of γ is provided in Figure 4. The well-definedness and the injectivity of γ are obvious. It remains to prove the surjectivity of γ. We prove surjectivity by constructing an explicit preimage. To achieve this, we start with some observations on the codomain of γ. So let P be a point set in G r ∩ C with sum-marginal λ. Let M := [8r , 9r ], N := [3r , 4r ] and R := [0, r ]. We first establish that for all points (x, y, z) ∈ P it holds that x, y and z are all contained in M , N or R and neither two are included in the same interval. As a first insight, since λ i = 0 for all i ∈ M ∪ N ∪ R, it follows that x, y and z are indeed contained in M ∪ N ∪ R. In order to prove that the coordinates x, y and z must stem from pairwise distinct sets M , N and R, we distinguish the following scenarios and show that each case causes a contradiction: Two coordinates in R: If, say, y, z ∈ R, then x = r − y − z ≥ 13r − r − r = 11r . But the range [11r , 13r ] intersects none of M , N and R.
We conclude that each element of P is contained in M × N × R. Figure 4 visualizes this result geometrically: The remaining 6 regions σ(M × N × R) for permutations σ ∈ S 3 are precisely the 6 non-colored triangles, one of which lies in C (the left one at the very bottom). It now follows that the the preimage of P under γ is the set {(x − 9r , y − 3r , z) : (x, y, z) ∈ P }, which is a point set in G r with (X, Y, Z)-marginal (µ , ν , ρ ) as required.
Note that without any changes the whole proof also works for the open cone C instead of C, which proves the result for Skew-Symmetric 2D-X-Ray.
Connecting Problems 9 and 10 in [Sta99b] Problem 9 in [Sta99b] is asking for a combinatorial interpretation for a λ (n, m); this is also known as the pletyhsm problem. Much less is known about this question than about the Kronecker problem, although numerous algorithms exist to compute plethysm coefficients (see the references provided in [LR11]), and [KM18] rules out certain approaches towards finding combinatorial formulas. We make progress on the plethysm problem by giving a combinatorial interpretation in an interesting subcase, see Theorem 7.1 below. We carefully mimic the proof technique in [IMW17], where a combinatorial interpretation for the Kronecker coefficient is given in a subcase. Since we reach the same combinatorial interpretation, we conclude that the plethysm coefficient equals the Kronecker coefficient in this case. The detailed statement can be found in Theorem 7.1 below.
To see the connection, consider the net of reductions as depicted in Figure 2: First, observe that all reductions ultimately start from #2D-X-Ray. Furthermore, any reduction between two tomography problems is parsimonious. In a similar sense, the reductions to computing the Kronecker or plethysm coefficients preserve the number of solutions, too, by guaranteeing that the respective coefficients precisely reflect the number of solutions to the corresponding tomography problem.
To illustrate the application of Theorem 7.1, we give some brief examples starting with 2D-X-Ray instances of size r = 1 each. Unfortunately, all such instances are either uniquely satisfiable or unsatisfiable and thus the resulting plethysm and Kronecker coefficients take values 0 and 1 only.

Positive Formulas for Restricted Plethysm Coefficients
In the course of the main proof, we discovered that a restricted class of plethysm coefficients can be explained as the number of certain combinatorial objects. As motivated before, results of that sort are of independent interest.
Formally speaking, we would like to show that the problem of computing general plethysm coefficients p λ (µ, ν) is contained in the complexity class #P. Completely resolving that hypothesis seems out of reach with today's techniques. Nevertheless, we make partial progress by proving that for certain nontrivial sets of inputs (µ, ν, λ), computing p λ (µ, ν) actually is in #P. In particular, let where λ is the sum-marginal of P r−1 , and λ satisfies r i=0 i · λ i = r| λ|/3 where λ is the sum-marginal of P r−1 , and λ satisfies r i=0 i · λ i = r| λ|/3 The following theorem follows from Proposition 5.5.
Using the same idea, we identify a more general class of instances (µ, ν, λ) for which the same result applies. Namely, for any choice of µ (and ν = or ν = t ), we can construct partitions λ in the same spirit as above. Let p(r) = |P r | denote the number of points in the complete pyramid P r , or equivalently, the total number of partitions of all integers ≤ r into at most three parts. Any Young diagram µ = (n 0 , n 1 , . . . , n ) t can be written as µ = n 0 n 1 n 2 · · · · · · n = n 0 n 1 n 2 · · · n n 0 n 1 n 2 · · · n = µ µ , where, for all 0 ≤ j ≤ , we pick r j minimal so that n j < p(r j ) and assign n j := p(r j − 1) and n j := n j − n j . Furthermore, let n := j n j = j n j + n j . As depicted, we refer to the upper rows (that is, rows of small indices) by µ and denote the remaining skew shape by µ. After decomposing µ in that manner, let us define where λ j is the sum-marginal of P r j −1 , and λ j satisfies In a very similar way, we can define Ψ : This time, let p(r) be the total number of partitions of all integers ≤ r into at most three distinct parts (that is, the number of points in the complete pyramid P r ). Given µ, determine r 0 , . . . , r as before with p(·) in place of p(·). Then: where λ j is the sum-marginal of P r j −1 , and λ j satisfies
Clearly, Φ Sym ⊂ Ψ Sym and Φ ⊂ Ψ , so Theorem 8.2 really generalizes Theorem 8.1. For the following proof of Theorem 8.2, we will omit the treatment of Ψ and focus on Ψ Sym only; both argumentations are analogous and the only exceptions have been pointed out in the preceding sections.
We catch up on some representation theoretic background. A Young tableau T is a Young diagram λ together with a filling of all boxes in λ with objects taken from some alphabet. If a total ordering on the alphabet is understood, then we say that T is semistandard whenever the entries of T are strictly increasing down each column and weakly increasing along each row. The Weyl module S λ V can explicitly be constructed in terms of Young tableaux: By arbitrarily indexing the boxes of λ, we fix a basis {v T } T for |λ| V where T ranges over all Young tableaux of shape λ filled with basis elements of V . Now S λ V is the largest subspace of |λ| V satisfying the following two exchange conditions (known as the Grassmann-Plücker relations): where T is obtained from T by exchanging two vertically adjacent boxes.
2. v T = T v T where, for some i and , T ranges over all tableaux obtained from T by exchanging the top-most boxes in column i with any boxes in column i − 1 preserving the vertical order.
It is known that the set of all semistandard tableaux of shape λ filled with basis elements of V forms a basis of S λ V .
For the remainder of this section, let (µ, ν, λ) ∈ Ψ Sym and let X 0 , X 1 , . . . , X k−1 be an ordered basis of V = C k . Then the set of all monomials X x X y X z for x, y, z ∈ [0, k − 1] forms a basis of Sym 3 V ; we sometimes identify monomials X x X y X z with their representatives (x, y, z) ∈ C <k := [0, k − 1] 3 ∩ C. Next, let ≺ be the partial ordering on C <k defined by (x, y, z) ≺ (x , y , z ) whenever x + y + z < x + y + z and arbitrarily extend ≺ to a total order. From the previous paragraph we derive that {v T } T forms a basis of S µ S ν V = S µ Sym 3 V , where T ranges over all semistandard tableaux of shape µ and alphabet C <k . By construction, the weight of a vector v T equals the sum of all entry-wise weights in T .
For a Young tableau T of shape µ, we write T to denote the subtableau of shape µ and we write T to denote the skew tableau corresponding to µ.
We proceed to rework some parts of Sections 5 and 6, where we now consider semistandard tableaux in place of pyramids-the exact same proof is recovered when restricting µ (and thereby the shape of the tableaux T ) to a single column.

Lemma
Let T be a semistandard Young tableau of shape µ and weight λ. Then: 1. T is of weight λ := j λ j . In fact, there exists only one such tableau T , which is obtained by filling all boxes in the i-th row of µ with the i-th smallest monomial according to ≺.
2. T is of weight λ := j λ j . Moreover, the j-th column of T exclusively contains monomials X x X y X z where x + y + z = r j .
The proof of Lemma 8.3 roughly follows Lemma 6.5, and to this end we first lift the definition of coordinate sums to tableaux: Let where the sum is over all entries (x, y, z) of T . Again, it is easy to relate the coordinate sum of T with its weight: Proof of Lemma 8.3. Assume that T is a semistandard Young tableau of shape µ and weight λ. We determine B(T ) exactly: Recall that λ j equals the sum-marginal of the complete pyramid P r j . In viewing the j-th column in T as a point set in the closed cone of size n j = |P r j | + n j , Lemma 6.4 yields that the contribution of the j-th column to the coordinate sum B(T ) is at least r j i=0 i · λ j i + n j · r j . In conjunction with (8.4), that bound is tight. By applying Lemma 6.4 again, we learn that for all j, the j-th column of T corresponds to a pyramid P that satisfies P r j −1 ⊆ P ⊂ P r j . But then the only way to label the boxes in the j-th column of T is as stated: The top-most n j boxes are filled with points in P r j −1 and all n j boxes below are filled with points in G r j ∩ C.
8.5 Lemma Let T be a semistandard Young tableau of shape µ and weight λ. Then v T is a highest weight vector in S µ Sym 3 V .
Proof. We assert that v T vanishes under all raising operators E a,b . In fact, it suffices to show that v T vanishes under E a,b after canonically embedding v T ∈ S µ Sym 3 V ⊂ j=0 n j Sym 3 V (that is, the space obtained by omitting condition 2 in the above characterization of Weyl modules). Let ϕ : j=0 n j Sym 3 V → S µ Sym 3 V be the canonical projection. We where each column T j of T is interpreted as a point set in the closed cone and v P for P such a point set is chosen as in Theorem 5.3. Lemma 8.3 in particular implies that each column T j is a pyramid in the closed cone, which entails that E a,b v T j = 0 by Theorem 5.3. Hence: the second equality holds since ϕ is GL(V )-equivariant.
Recall that p λ (µ, ν) equals the dimension of the weight-λ highest weight subspace of S µ S ν V . Since each elementary weight-λ vector is a highest weight vector in this case, p λ (µ, ν) equals the dimension of the weight subspace of weight λ-or equivalently, the number of semistandard tableaux of shape µ and weight λ. Formally, we obtain the following statement which implies Theorem 8.2 as an immediate consequence.

GapP-Completeness
This section is devoted to proving Theorem 3.4, which claims that computing general plethysm coefficients p λ (µ, ν), and also the special cases a λ (n, m) and b λ (n, m), is GapP-complete. The hardness part follows immediately from Theorem 3.2 7 . In order to show containment in GapP, we will derive an explicit formula involving signs for the general plethysm coefficients p λ (µ, ν). We heavily rely on tools from the theory of symmetric functions, however, we shall refrain from describing the necessary background in detail and instead limit ourselves to the most important definitions; see for instance [Sta99a] for a thorough introduction. The following argument is mostly standard and similar to e.g. [DIP19].
Let Λ denote the ring of symmetric polynomials (that is, polynomials which are invariant under any permutation of variables) in finitely many variables X 0 , X 1 , . . .. The characters of the irreducible GL n -representation S ν C n are elements s ν (X 0 , . . . , X n−1 ) ∈ Λ called Schur functions (where X 0 , . . . , X n−1 correspond to the eigenvalues of the conjugacy class of GL n ). Recall the notion of semistandard Young tableaux as introduced in Section 8 and let T 0 , . . . , T l−1 be an arbitrary enumeration of all semistandard tableaux of shape ν filled with entries 0, . . . , n − 1. The weight wt(T ) ∈ N [0,n−1] of a tableau T is defined as the composition such that wt(T ) i is the number of entries i in T . We have that s ν (X 0 , . . . , X n−1 ) = l−1 i=0 X T i , (9.1) for X T := t∈T X t , where the product is over (the multiset of) the entries t of T . Following this description, it follows that the character of the composed representation S µ S ν C n is given by s µ [s ν ](X 0 , . . . , X n−1 ) := s µ (X T 0 , . . . , X T l−1 ); that composition s µ [s ν ] of symmetric functions is called a symmetric function plethysm. The set of Schur function forms a basis of Λ. Moreover, since the Schur functions are the characters of the irreducible GL n -representations, there exists an inner product ·, · on Λ so that the Schur functions form an orthonormal basis: s λ , s µ = δ λµ . It follows that for any GL nrepresentation with character χ, the multiplicity of S λ C n in its decomposition into irreducibles equals s λ , χ . In particular, p λ (µ, ν) = s λ , s µ [s ν ] . It remains to develop a GapP-formula for s λ , s µ [s ν ] .
From that identity it is obvious that q κ (µ, ν) is GapP-computable (even #P-computable): The coefficients K µπ are clearly #P-computable and after guessing π and ρ, we can check the condition κ = i ρ i wt(T i ) in polynomial time.