2.1. Symmetric groups

For $n\geq 0$ , consider the symmetric group $S_n$ on $\underline {n}$ . For $\lambda \vdash n$ , denote by $\mathbf {S}^\lambda $ the Specht $\Bbbk [S_n]$ -module corresponding to $\lambda $ . We refer the reader to [Reference SaganSa01, Section 2.3] for an explicit definition of Specht modules; we do not need it. We only recall that $\{\mathbf {S}^\lambda \,:\,\lambda \vdash n\}$ is a complete and irredundant list of representatives of the isomorphism classes of simple $\Bbbk [S_n]$ -modules. For $\lambda ,\mu ,\nu \vdash n$ , we denote by $\mathbf {g}_{\lambda ,\mu }^\nu $ the corresponding Kronecker coefficient, that is, the multiplicity of $\mathbf {S}^\nu $ in $\mathbf {S}^\lambda \otimes _{\Bbbk }\mathbf {S}^\mu $ . The latter is a $\Bbbk [S_n]$ -module via the diagonal action of $S_n$ .

For $\lambda \vdash n$ , $\mu \vdash m$ and $\nu \vdash n+m$ , we denote by $\mathbf {c}_{\lambda ,\mu }^\nu $ the corresponding Littlewood– Richardson coefficient, that is, the multiplicity of $\mathbf {S}^\nu $ in $\mathrm {Ind}_{\Bbbk [S_n\times S_m]}^{\Bbbk [S_{n+m}]} (\mathbf {S}^\lambda \otimes _\Bbbk \mathbf {S}^\mu \vphantom {\mathbf {S}^\lambda \otimes _{\Bbbk _{\mathbf {S}_{\mu _{\mathbf {S}_{\mu }}}}}\mathbf {S}^\mu })$ , where $S_n$ is embedded into $S_{n+m}$ using the canonical embedding of $\underline {n}$ into $\underline {n+m}$ while $S_m$ is embedded into $S_{n+m}$ via the $+n$ -shift of the canonical embedding of $\underline {m}$ into $\underline {n+m}$ .

We refer to [Reference SaganSa01] for all details related to symmetric groups and their representations.

2.2. Induction and bimodules

For $k+m=n$ , consider the group $G=S_n\times S_k\times S_m$ and its subgroup H defined as the image of the embedding of $S_k\times S_m$ into $S_n\times S_k\times S_m$ constructed as follows: each $\sigma \in S_k$ is sent to $(\tilde {\sigma },\sigma ,e)$ , where $\tilde {\cdot }:S_k\to S_n$ is the natural embedding of $S_k$ into $S_n$ with respect to the first k elements; furthermore, each $\sigma \in S_m$ is sent to $(\hat {\sigma },e,\sigma )$ , where $\hat {\cdot }:S_m\to S_n$ is the natural embedding of $S_m$ into $S_n$ with respect to the last m elements.

The group G acts naturally on $G/H$ making $\Bbbk [G/H]$ into a G-module. We can view this G-module as a $\Bbbk [S_n]$ - $\Bbbk [S_k\times S_m]$ -bimodule using the canonical anti-involution $\sigma \mapsto \sigma ^{-1}$ on $S_k\times S_m$ . Since $\Bbbk [S_k\times S_m]\cong \Bbbk [S_k]\otimes _\Bbbk \Bbbk [S_m]$ , given an $S_k$ -module V and an $S_m$ -module W, we have the $S_n$ -module

(2-1) $$ \begin{align} \Bbbk[G/H]\bigotimes_{\Bbbk[S_k\times S_m]} (V\otimes_\Bbbk W). \end{align} $$

It is easy to see that mapping H to e gives rise to an isomorphism between the $\Bbbk [S_n]$ - $\Bbbk [S_k\times S_m]$ -bimodule $\Bbbk [G/H]$ and the $\Bbbk [S_n]$ - $\Bbbk [S_k\times S_m]$ -bimodule $\Bbbk [S_n]$ . Therefore, the $S_n$ -module in (2-1) is isomorphic to $\mathrm {Ind}_{\Bbbk [S_k\times S_m]}^{\Bbbk [S_n]}(V\otimes _\Bbbk W)$ .

2.3. Tensor product and bimodules

Consider the group $G=S_n\times S_n\times S_n$ and its subgroup H given by the diagonal embedding of $S_n$ , that is, $\sigma \mapsto (\sigma ,\sigma ,\sigma )$ . Then G acts naturally on $G/H$ making $\Bbbk [G/H]$ into a G-module. We can view this G-module as a $\Bbbk [S_n]$ - $\Bbbk [S_n\times S_n]$ -bimodule using the canonical anti-involution $\sigma \mapsto \sigma ^{-1}$ on $S_n\times S_n$ (the last two factors). As $\Bbbk [S_n\times S_n]\cong \Bbbk [S_n]\otimes _\Bbbk \Bbbk [S_n]$ , given two $S_n$ -modules V and W, we have the $S_n$ -module

(2-2) $$ \begin{align} \Bbbk[G/H]\bigotimes_{\Bbbk[S_n\times S_n]} V\otimes_\Bbbk W. \end{align} $$

It is easy to see that mapping H to $(e,e)$ gives rise to an isomorphism between the $\Bbbk [S_n]$ - $\Bbbk [S_n\times S_n]$ -bimodule $\Bbbk [G/H]$ and the $\Bbbk [S_n]$ - $\Bbbk [S_n\times S_n]$ -bimodule $\Bbbk [S_n\times S_n]$ , where the left action of $S_n$ is diagonal. Therefore, the $S_n$ -module in (2-2) is isomorphic to $V\otimes _\Bbbk W$ with the usual (diagonal) action of $S_n$ .

3.1. The semigroup $IS_n$

Consider the symmetric inverse semigroup $IS_n$ of all bijections between subsets of $\underline {n}$ . The semigroup operation in $IS_n$ is a natural analogue of composition for partially defined maps. Here is an example for $n=5$ , where composition is from right to left and the elements are presented using the two-row notation for (partial) functions:

$$ \begin{align*} \left(\begin{array}{@{}ccc@{}}2&3&5\\1&5&2\end{array}\right) \circ \left(\begin{array}{@{}cccc@{}}1&2&4&5\\2&4&3&1\end{array}\right)= \left(\begin{array}{@{}cc@{}}1&4\\1&5\end{array}\right). \end{align*} $$

For $\sigma \in IS_n$ , we denote by $\mathrm {dom}(\sigma )$ and $\mathrm {cod}(\sigma )$ the domain and the codomain of $\sigma $ , respectively. Recall, see for example [Reference Ganyushkin and MazorchukGM09, Section 4.4], the following description of Green’s relations $\mathcal {L}$ , $\mathcal {R}$ , $\mathcal {H}$ and $\mathcal {D}=\mathcal {J}$ for $IS_n$ :

• • $\sigma \mathcal {L}\pi $ if and only if $\mathrm {dom}(\sigma )=\mathrm {dom}(\pi )$ ;

• • $\sigma \mathcal {R}\pi $ if and only if $\mathrm {cod}(\sigma )=\mathrm {cod}(\pi )$ ;

• • $\sigma \mathcal {H}\pi $ if and only if $\mathrm {dom}(\sigma )=\mathrm {dom}(\pi )$ and $\mathrm {cod}(\sigma )=\mathrm {cod}(\pi )$ ;

• • $\sigma \mathcal {J}\pi $ if and only if $|\mathrm {dom}(\sigma )|=|\mathrm {dom}(\pi )|$ .

The cardinality of $\mathrm {dom}(\pi )$ is called the rank of $\pi $ and denoted $\mathrm {rank}(\pi )$ .

For a subset $X\subset \underline {n}$ , we denote by $\varepsilon _X$ the identity on X. Then $\{\varepsilon _X :\, X\subset \underline {n}\}$ is the set of all idempotents in $IS_n$ . For $X,Y\subset \underline {n}$ , we have $\varepsilon _X\varepsilon _Y=\varepsilon _{X\cap Y}$ .

3.2. Cells and simple $\Bbbk [IS_n]$ -modules

Let $\mathbf {L}$ be a left cell (that is, an $\mathcal {L}$ -equivalence class) in $IS_n$ . The linearization $\Bbbk [\mathbf {L}]$ is naturally a left $\Bbbk [IS_n]$ -module with the action given as follows, for $\sigma \in IS_n$ and $\pi \in \mathbf {L}$ :

(3-1) $$ \begin{align} \sigma\cdot\pi= \begin{cases} \sigma\pi& \text{ if } \sigma\pi\in \mathbf{L},\\ 0& \text{ otherwise.} \end{cases} \end{align} $$

This is the cell $\Bbbk [IS_n]$ -module corresponding to $\mathbf {L}$ .

The set $\mathbf {L}$ contains a unique idempotent, namely $\varepsilon _X$ , where X is the common domain for all elements in $\mathbf {L}$ . Let $G(\mathbf {L})$ be the corresponding maximal subgroup in $IS_n$ , that is, the $\mathcal {H}$ -equivalence class of $\varepsilon _X$ . Set $k:=|X|$ . Then $G(\mathbf {L})=S_X\cong S_{k}$ . We have the vector space decomposition

(3-2) $$ \begin{align} \Bbbk[\mathbf{L}]\cong \bigoplus_{Y\subset \underline{n}, |Y|=k}\varepsilon_Y \Bbbk[\mathbf{L}]. \end{align} $$

The group $G(\mathbf {L})\cong S_k$ acts on $\mathbf {L}$ by multiplication on the right. This equips $\Bbbk [\mathbf {L}]$ with the natural structure of a $\Bbbk [IS_n]$ - $\Bbbk [S_k]$ -bimodule. In fact, each $\varepsilon _Y \Bbbk [\mathbf {L}]$ in (3-2) is a free right $\Bbbk [S_k]$ -module of rank one. Up to isomorphism, the bimodule $\Bbbk [\mathbf {L}]$ does not depend on the choice of $\mathbf {L}$ inside its $\mathcal {J}$ -cell (that is, for the fixed k).

For any $\lambda \vdash k$ , the $\Bbbk [IS_n]$ -module

(3-3) $$ \begin{align} \mathbf{N}^{\lambda}:=\Bbbk[\mathbf{L}]\bigotimes_{\Bbbk[S_k]} \mathbf{S}^\lambda \end{align} $$

is simple. By the above, up to isomorphism, it does not depend on the choice of $\mathbf {L}$ . Moreover, up to isomorphism, each simple $\Bbbk [IS_n]$ -module has such a form for some k and $\lambda $ as above. Putting all this together, we see that the isomorphism classes of simple $\Bbbk [IS_n]$ -modules are in bijection with partitions $\lambda \vdash k$ , where $k\leq n$ .

We refer to [Reference Ganyushkin and MazorchukGM09, Ch. 11] and [Reference SteinbergSt16] for further details.

3.3. Kronecker coefficients for $IS_n$

Being a semigroup algebra, $\Bbbk [IS_n]$ has the canonical comultiplication given by the diagonal map $\sigma \mapsto \sigma \otimes \sigma $ . Therefore, for two $\Bbbk [IS_n]$ -modules M and N, the tensor product $M\otimes _\Bbbk N$ has the natural structure of a $\Bbbk [IS_n]$ -module.

For a fixed n, let $k,l,m\leq n$ , $\lambda \vdash k$ , $\mu \vdash l$ and $\nu \vdash m$ . We denote by $\mathbf {k}_{\lambda ,\mu }^\nu $ the multiplicity of $\mathbf {N}^\nu $ in $\mathbf {N}^\lambda \otimes _\Bbbk \mathbf {N}^\mu $ . It is natural to call these $\mathbf {k}_{\lambda ,\mu }^\nu $ the Kronecker coefficients for $IS_n$ . Our main result in this section is the following theorem.

We note that the Kronecker coefficients for $IS_n$ appear in the theory of FI-modules, that is, modules over the category of finite sets and injections, and especially [Reference Church, Ellenberg and FarbCEF15, (17)], for details. For $\lambda =(1)$ , the computation of $\mathbf {k}^{\nu }_{\lambda ,\mu }$ can be obtained from [Reference Mishra and SrivastavaMS21, Corollary 2.7].

3.4. Tensor product of cell modules

Before we can prove Theorem 3.1, we need to consider a simpler decomposition problem, namely, that for the tensor product of cell modules. For $0\leq r\leq n$ , we denote by $\mathbf {L}_r$ the left cell in $IS_n$ containing $\varepsilon _{\underline {r}}$ .

Let k, l and m be as in Theorem 3.1. Consider the $IS_n$ -module $\Bbbk [\mathbf {L}_k]\otimes _\Bbbk \Bbbk [\mathbf {L}_l]$ . We show that this module has a filtration whose subquotients are isomorphic to cell modules and we also determine the multiplicity of $\Bbbk [\mathbf {L}_m]$ as a subquotient in that filtration. The fact that this multiplicity is well defined (that is, does not depend on the choice of a filtration with cell subquotients) is clear since nonisomorphic cell modules do not have any common nonzero simple summands.

The elements in $\mathbf {L}_k$ are given by pairs $(X,\sigma )$ , where X is a cardinality k subset of $\underline {n}$ and $\sigma $ is a bijection from $\underline {k}$ to X. Similarly, the elements in $\mathbf {L}_l$ are given by pairs $(Y,\pi )$ , where Y is a cardinality l subset of $\underline {n}$ and $\pi $ is a bijection from $\underline {l}$ to Y. For such a $\xi =(X,\sigma )$ , the rank of $\varepsilon _{\underline {m}}\xi $ equals k if and only if $X\subset \underline {m}$ . Similarly, for such an $\eta =(Y,\pi )$ , the rank of $\varepsilon _{\underline {m}}\eta $ equals l if and only if $Y\subset \underline {m}$ . Hence, $\varepsilon _{\underline {m}}$ does not kill $\xi \otimes \eta $ if and only if $X\subset \underline {m}$ and $Y\subset \underline {m}$ (in particular, $m\geq \max \{k,l\}$ ).

If we assume that $X\cup Y\subsetneq \underline {m}$ , then $\xi \otimes \eta $ is not killed by the idempotent $\varepsilon _{X\cup Y}$ whose rank is strictly smaller than m. This means that such $\xi \otimes \eta $ cannot contribute to the cell module $\Bbbk [\mathbf {L}_m]$ . Therefore, we only need to consider the case $X\cup Y=\underline {m}$ (which implies $\max \{k,l\}\leq m\leq k+l$ ). In this case, $|X\cap Y|=k+l-m=b$ .

Denote by Q the set of all $\xi \otimes \eta $ as in the previous paragraph. The group $S_m$ , identified as the maximal subgroup $G(\mathbf {L}_m)$ , for the idempotent $\varepsilon _{\underline {m}}$ , acts on $\xi \otimes \eta \in Q$ naturally by left multiplication. The group $S_k\times S_l$ acts on $\xi \otimes \eta \in Q$ on the right, where $S_k$ acts on $\sigma $ by right multiplication while $S_l$ acts on $\pi $ by right multiplication. The two actions obviously commute.

Let us determine the cardinality of Q. We can choose $X\cap Y$ in $\binom {m}{b}$ different ways and then the rest of X in $\binom {m-b}{a}=\binom {m-b}{c}$ different ways. As $\sigma $ and $\pi $ are arbitrary and $a+b+c=m$ , there are

$$ \begin{align*} \binom{a+b+c}{b}\binom{a+c}{a}(a+b)!(b+c)!= \frac{(a+b+c)!(a+b)!(b+c)!}{a!b!c!} \end{align*} $$

different choices in total.

Now we can perform the following construction, similar to those presented in Sections 2.2 and 2.3. Consider the group $G=S_{a+b+c}\times S_{a+b}\times S_{c+b}$ . We split the $a+b+c$ elements on which $S_{a+b+c}$ acts into a apricot, b blue and c crimson elements. Similarly, we split the $a+b$ elements on which $S_{a+b}$ acts into a apricot and b blue elements. Finally, we split the $b+c$ elements on which $S_{b+c}$ acts into b blue and c crimson elements. Fix some identification of apricot elements for $S_{a+b+c}$ with apricot elements for $S_{a+b}$ , and similarly for blue and crimson elements.

Let H be the subgroup of G generated by the following elements:

• • all $(\rho ,\rho ,e)$ , where $\rho $ is a permutation of apricot elements;

• • all $(\rho ,\rho ,\rho )$ , where $\rho $ is a permutation of blue elements;

• • all $(\rho ,e,\rho )$ , where $\rho $ is a permutation of crimson elements.

Clearly, H is isomorphic to $S_a\times S_b\times S_c$ . Consider the linearization $\Bbbk [G/H]$ , which is naturally a G-module. In particular, it is a $\Bbbk [S_{a+b+c}]$ - $\Bbbk [S_{a+b}\times S_{c+b}]$ -bimodule, using the canonical anti-involution $\sigma \mapsto \sigma ^{-1}$ on $S_{a+b}\times S_{c+b}$ . Now we can compare the $\Bbbk [S_{a+b+c}]$ - $\Bbbk [S_{a+b}\times S_{c+b}]$ -bimodules $\Bbbk [G/H]$ and $\Bbbk Q$ .

From the definitions, it is easy to see that for any conjugate $H'$ of H in G, we have $H'\cap S_{a+b+c} = e$ . This implies that as an $S_{a+b+c}$ -module, the module $\Bbbk [G/H]$ is free. By Lemma 3.2, we thus also have that $\Bbbk Q$ is free, as an $S_{a+b+c}$ -module. In particular, Q is just a disjoint union of copies of regular $S_m$ -orbits. Combining the definitions with decomposition (3-2), we have that each such regular $S_m$ -orbit determines in $\Bbbk [\mathbf {L}_k]\otimes _\Bbbk \Bbbk [\mathbf {L}_l]$ a subquotient isomorphic to $\Bbbk [\mathbf {L}_m]$ . This, on the one hand, implies that $\Bbbk [\mathbf {L}_k]\otimes _\Bbbk \Bbbk [\mathbf {L}_l]$ has a filtration whose subquotients are cell modules. On the other hand, it also says that the multiplicity of $\Bbbk [\mathbf {L}_m]$ in $\Bbbk [\mathbf {L}_k]\otimes _\Bbbk \Bbbk [\mathbf {L}_l]$ equals the number of $S_m$ -orbits on Q. Summing everything up, we obtain the following corollary.

3.5. Proof of Theorem 3.1

Claim (a) of Theorem 3.1 follows from the two estimates $m\geq \max \{k,l\}$ and $m\leq k+l$ obtained in the previous subsection.

In the previous subsection, we considered the bimodule $\Bbbk Q$ that describes the part of the tensor product $\Bbbk [\mathbf {L}_k]\otimes _\Bbbk \Bbbk [\mathbf {L}_l]$ that determines the $\Bbbk [\mathbf {L}_m]$ -summands of that tensor product. In particular, combining the definition of Q with (3-3), it follows that

(3-5) $$ \begin{align} [\mathbf{N}^\lambda\otimes_\Bbbk\mathbf{N}^\mu:\mathbf{N}^\nu]= [\Bbbk Q\otimes_{\Bbbk[S_k]\otimes_\Bbbk \Bbbk[S_l]} (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu):\mathbf{S}^\nu]. \end{align} $$

Therefore, due to Lemma 3.2, to prove Claim (b), we need to show that (3-4) gives the multiplicity

$$ \begin{align*} [\Bbbk[G/H]\otimes_{\Bbbk[S_k]\otimes_\Bbbk \Bbbk[S_l]} (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu):\mathbf{S}^\nu]. \end{align*} $$

Define an additive functor $\mathcal {F}$ from $\Bbbk [S_{a+b}\times S_{b+c}]$ -mod to $\Bbbk [S_{a+b+c}]$ -mod for two modules $V\in S_{a+b}$ -mod and $W\in S_{b+c}$ -mod, as follows.

• • First, restrict V to $S_a\times S_b$ and W to $S_b\times S_c$ .

• • Next, consider $V\otimes _\Bbbk W$ as an $S_a\times S_b\times S_c$ -module with respect to the obvious actions of $S_a$ and $S_c$ on the first and on the second components, respectively, and the diagonal action of $S_b$ .

• • Finally, induce the obtained $S_a\times S_b\times S_c$ -module up to $S_{a+b+c}$ .

Let us now compute $\mathcal {F}(\mathbf {S}^\lambda \otimes _\Bbbk \mathbf {S}^\mu )$ . To start with,

$$ \begin{align*} \mathrm{Res}_{\Bbbk[S_a\times S_b]}^{\Bbbk[S_{a+b}]}\mathbf{S}^\lambda\cong \bigoplus_{\alpha\vdash a} \bigoplus_{\beta\vdash b} (\mathbf{S}^\alpha\otimes_\Bbbk \mathbf{S}^\beta)^{\oplus \mathbf{c}_{\alpha,\beta}^\lambda} \end{align*} $$

and

$$ \begin{align*} \mathrm{Res}_{\Bbbk[S_b\times S_c]}^{\Bbbk[S_{b+c}]}\mathbf{S}^\mu\cong \bigoplus_{\gamma\vdash c} \bigoplus_{\delta\vdash b} (\mathbf{S}^\delta\otimes_\Bbbk \mathbf{S}^\gamma)^{\oplus \mathbf{c}_{\delta,\gamma}^\mu}. \end{align*} $$

At the next step,

$$ \begin{align*} (\mathbf{S}^\alpha\otimes_\Bbbk \mathbf{S}^\beta)\otimes_\Bbbk (\mathbf{S}^\delta\otimes_\Bbbk \mathbf{S}^\gamma)\cong \bigoplus_{\epsilon\vdash b} (\mathbf{S}^\alpha\otimes_\Bbbk \mathbf{S}^\epsilon \otimes_\Bbbk \mathbf{S}^\gamma)^{\oplus \mathbf{g}_{\beta,\delta}^\epsilon}. \end{align*} $$

Finally, at the last step, which we realize as the composition of first inducing from $S_a\times S_b\times S_c$ to $S_{a+b}\times S_c$ and then from the latter to $S_{a+b+c}$ ,

$$ \begin{align*} \mathrm{Ind}_{S_a\times S_b\times S_c}^{S_{a+b+c}} \mathbf{S}^\alpha\otimes_\Bbbk \mathbf{S}^\epsilon \otimes_\Bbbk \mathbf{S}^\gamma\cong \bigoplus_{\nu\vdash a+b+c} \bigoplus_{\kappa\vdash a+b} (\mathbf{S}^\nu)^{\oplus \mathbf{c}_{\alpha,\epsilon}^\kappa\cdot \mathbf{c}_{\kappa,\gamma}^\nu}. \end{align*} $$

Collecting all these formulae together, we see that the multiplicity of $\mathbf {S}^{\nu }$ in $\mathcal {F}(\mathbf {S}^\lambda \otimes _\Bbbk \mathbf {S}^\mu )$ is given by (3-4).

Let us now try to understand the $\Bbbk [S_{a+b+c}]$ - $\Bbbk [S_{a+b}\times S_{b+c}]$ -bimodule that realizes $\mathcal {F}$ . Taking into account the constructions presented in Sections 2.2 and 2.3, the combination of the first two steps is given by

$$ \begin{align*} \Bbbk[S_{a+b}]\otimes_\Bbbk \Bbbk[S_{b+c}], \end{align*} $$

where the right action is regular, the left actions of $S_a$ and $S_c$ are the obvious ones, and the left action of $S_b$ is the diagonal action. At the last step, we just tensor with the regular $\Bbbk [S_{a+b+c}]$ - $\Bbbk [S_{a+b+c}]$ -bimodule, where the right action is restricted to $S_a\times S_b\times S_c$ .

Directly from the definitions, we see that the stabilizer (in G) of the element $e\otimes (e\otimes e)$ from

(3-6) $$ \begin{align} \Bbbk[S_{a+b+c}]\otimes_{\Bbbk[S_a\times S_b\times S_c]} (\Bbbk[S_{a+b}]\otimes_\Bbbk \Bbbk[S_{b+c}]) \end{align} $$

contains H. Thus, sending the element H to the element $e\otimes (e\otimes e)$ , gives rise to a $\Bbbk [S_{a+b+c}]$ - $\Bbbk [S_{a+b}\times S_{b+c}]$ -bimodule homomorphism from $\Bbbk [G/H]$ to the bimodule in (3-6). As $e\otimes (e\otimes e)$ generates the latter, this homomorphism is surjective. By comparing the dimensions, we see that it is bijective. This completes the proof of Claim (b) in Theorem 3.1.

3.6. Advance remark

We complete this section with a remark about symmetric inverse semigroups that is inspired by Example 4.6 below (this example is also studied in more detail in Section 6).

If we interpret the elements in $IS_n$ as rook $0/1$ -matrices, then the Kronecker product of matrices defines a homomorphism from $IS_n\times IS_m$ to $IS_{nm}$ . Note that this homomorphism is not injective (as tensoring anything with the zero matrix outputs the zero matrix). Nevertheless, it gives rise to a homomorphism of unital associative algebras $\Bbbk [IS_n]\otimes _\Bbbk \Bbbk [IS_m]\to \Bbbk [IS_{nm}]$ . In particular, we have the corresponding pullback functor, which allows us to consider each $IS_{nm}$ -module as an $IS_n\times IS_m$ -module.

Cell modules over symmetric inverse semigroups (and their direct products) have a special feature: the cell module for a left cell $\mathbf {L}$ is the linearization of the obvious transitive action of our semigroup on $\mathbf {L}$ by partial transformations given by left multiplication; compare with (3-1). Classification of transitive acts of inverse semigroups is a classical part of the theory of inverse semigroups, see [Reference LawsonLa20, Theorem 5.4], and it is usually formulated in terms of closed inverse subsemigroups, which serve as stabilizers of points. Each such closed inverse subsemigroup has a distinguished idempotent and contains some subgroup H of the maximal subgroup corresponding to this idempotent. Abusing terminology, we can call this H the apex of the stabilizer. A left cell act corresponds exactly to the case when H is trivial, that is, consists only of the identity (that is, only of the corresponding idempotent).

If we now take a left cell $IS_{nm}$ -act and pull it back to $IS_n\times IS_m$ , we can consider a natural filtration of the latter act whose subquotients are transitive subquotient acts. The property that the apex of each point of the original act is trivial is clearly inherited under pullback. Consequently, the pullback of any cell $IS_{nm}$ -module has a filtration whose subquotients are cell $IS_n\times IS_m$ -modules. One can compare this with the results of Section 3.4.

4.1. The semigroup $I_n^*$

Consider the dual symmetric inverse semigroup $I_n^*$ , see [Reference Fitzgerald and LeechFL98]. The elements of $I_n^*$ are all possible bijections between the quotients of $\underline {n}$ and the operation is the natural version of composition for such bijections. Here is an example for $n=5$ , where composition is from right to left and the elements are presented using the two-row notation for functions:

$$ \begin{align*} \left(\begin{array}{@{}ccc@{}}\{1\}&\{2,3\}&\{4,5\}\\\{1,2\}&\{3,4\}&\{5\} \end{array}\right) \circ \left(\begin{array}{@{}ccc@{}}\{1,2\}&\{3\}&\{4,5\}\\ \{1\}&\{2,3,4\}&\{5\}\end{array}\right) = \left(\begin{array}{@{}cc@{}}\{1,2\}&\{3,4,5\}\\\{1,2\}&\{3,4,5\}\end{array}\right). \end{align*} $$

Each element $\xi $ in $I_n^*$ is thus uniquely given by the following data:

• • a set partition $\xi _d$ of $\underline {n}$ , called the domain of $\xi $ and denoted $\mathrm {dom}(\xi )$ ;

• • a set partition $\xi _r$ of $\underline {n}$ (with the same number of parts as $\xi _d$ ), called the codomain of $\xi $ and denoted $\mathrm {cod}(\xi )$ ;

• • a bijection $\overline {\xi }$ from $\underline {n}/\xi _d$ to $\underline {n}/\xi _r$ .

The cardinality of $\underline {n}/\xi _d$ is called the rank of $\xi $ and denoted $\mathrm {rank}(\xi )$ . We also call the same number the rank of $\xi _d$ .

Let $\xi ,\zeta \in I_n^*$ . Recall, see for example [Reference Fitzgerald and LeechFL98, Theorem 2.2], the following description of Green’s relations $\mathcal {L}$ , $\mathcal {R}$ , $\mathcal {H}$ and $\mathcal {D}=\mathcal {J}$ for $I_n^*$ :

• • $\xi \mathcal {L}\zeta $ if and only if $\mathrm {dom}(\xi )=\mathrm {dom}(\zeta )$ ;

• • $\xi \mathcal {R}\zeta $ if and only if $\mathrm {cod}(\xi )=\mathrm {cod}(\zeta )$ ;

• • $\xi \mathcal {H}\zeta $ if and only if $\mathrm {dom}(\xi )=\mathrm {dom}(\zeta )$ and $\mathrm {cod}(\xi )=\mathrm {cod}(\zeta )$ ;

• • $\xi \mathcal {J}\zeta $ if and only if $\mathrm {rank}(\xi )=\mathrm {rank}(\zeta )$ .

The idempotents of $I_n^*$ are naturally identified with equivalence relations on $\underline {n}$ : given an equivalence relation $\rho $ on $\underline {n}$ , the corresponding idempotent $\varepsilon _\rho $ is the identity map on $\underline {n}/\rho $ . The maximal subgroup of $I_n^*$ corresponding to $\rho $ is the symmetric group $S_\rho $ on the equivalence classes of $\rho $ .

The semigroup $I_k^*$ appears naturally in the context of the Schur–Weyl duality for the k th tensor power of the natural representation of $IS_n$ ; see [Reference Kudryavtseva and MazorchukKMaz09b, Theorem 1].

4.2. Cells and simple $I_n^*$ -modules

Let $\rho $ be an equivalence relation on $\underline {n}$ . Let $\mathbf {L}_\rho $ be the left cell (that is, an $\mathcal {L}$ -equivalence class) in $I_n^*$ containing $\varepsilon _\rho $ .

The linearization $\Bbbk [\mathbf {L}_\rho ]$ is naturally a left $I_n^*$ -module, where the action is given as follows for $\xi \in I_n^*$ and $\zeta \in \mathbf {L}_\rho $ :

$$ \begin{align*} \xi\cdot\zeta= \begin{cases} \xi\zeta& \text{ if } \xi\zeta\in \mathbf{L}_\rho,\\ 0& \text{ otherwise.} \end{cases} \end{align*} $$

This is the cell $I_n^*$ -module corresponding to $\mathbf {L}_\rho $ .

The maximal subgroup $S_\rho $ acts on $\mathbf {L}_\rho $ by multiplication on the right. This equips $\Bbbk [\mathbf {L}_\rho ]$ with the natural structure of a $\Bbbk [I_n^*]$ - $\Bbbk [S_\rho ]$ -bimodule.

For any $\lambda \vdash \mathrm {rank}(\varepsilon _\rho )$ , the $\Bbbk [I_n^*]$ -module

$$ \begin{align*} \mathbf{N}^{\lambda}:=\Bbbk[\mathbf{L}_\rho]\bigotimes_{\Bbbk[S_\rho]} \mathbf{S}^\lambda \end{align*} $$

is simple. Up to isomorphism, it does not depend on the choice of $\mathbf {L}_\rho $ inside its $\mathcal {J}$ -cell (that is, for the fixed rank of $\varepsilon _\rho $ ). Moreover, up to isomorphism, each simple $\Bbbk [I_n^*]$ -module has such a form for some $\rho $ and $\lambda $ . We refer to [Reference Fitzgerald and LeechFL98, Reference SteinbergSt16] for details.

4.3. Tensor product of cell modules

Let $\rho $ and $\sigma $ be two set partitions (also known as equivalence relations) of $\underline {n}$ . The tensor product

$$ \begin{align*} \Bbbk[\mathbf{L}_\rho]\otimes_\Bbbk\Bbbk[\mathbf{L}_\sigma] \end{align*} $$

is an $I_n^*$ -module via the diagonal action of $I_n^*$ . In this subsection, we show that this module has a filtration whose subquotients are isomorphic to cell modules. Also, we describe a combinatorial object that determines the corresponding multiplicities (just like in the case of $IS_n$ , the fact that these multiplicities are well defined follows directly from the semi-simplicity of the semigroup algebra of $I_n^*$ combined with the fact that two nonisomorphic cell modules do not have common isomorphic simple summands).

Set $k=\mathrm {rank}(\varepsilon _\rho )$ and $l=\mathrm {rank}(\varepsilon _\sigma )$ . Let $\tilde {\rho }$ be a set partition of $\underline {n}$ of rank k and $\tilde {\sigma }$ be a set partition of $\underline {n}$ of rank l. Consider the set partition $\tau $ of $\underline {n}$ whose parts are exactly the nonempty intersections of a part of $\tilde {\rho }$ with a part of $\tilde {\sigma }$ . We denote this $\tau $ by $\tilde {\rho }\cap \tilde {\sigma }$ .

Let $\preceq $ be the natural order on set partitions: we have $\alpha \preceq \beta $ if each part of $\alpha $ is a subset of some part of $\beta $ . Let $\xi \in \mathbf {L}_\rho $ and $\zeta \in \mathbf {L}_\sigma $ be such that $\mathrm {cod}(\xi )=\tilde {\rho }$ and $\mathrm {cod}(\zeta )=\tilde {\sigma }$ . Then $\tau $ is the maximum element in the set of all set partitions $\omega $ of $\underline {n}$ such that $\varepsilon _\omega \xi \in \mathbf {L}_\rho $ and $\varepsilon _\omega \zeta \in \mathbf {L}_\sigma $ . Note also that the action of $S_\tau $ on $\xi \otimes \zeta $ is, clearly, free.

All this means exactly that the element $\xi \otimes \zeta $ contributes to a cell subquotient $\Bbbk [\mathbf {L}_\tau ]$ of $\Bbbk [\mathbf {L}_\rho ]\otimes _\Bbbk \Bbbk [\mathbf {L}_\sigma ]$ .

To provide a combinatorial object for counting the multiplicities, we need to invert our problem. Given some $\tau $ , we need to find the number of all possible $\tilde {\rho }$ and $\tilde {\sigma }$ as above that are connected to $\tau $ also in the above sense.

Set $m:=\mathrm {rank}(\varepsilon _\tau )$ . Let $Q_{k,l}^{m}$ be the set of all $k\times l$ matrices with $0/1$ -entries satisfying the following conditions:

• • exactly m entries are equal to $1$ ;

• • all rows are nonzero;

• • all columns are nonzero.

The group $G:=S_k\times S_l$ acts on $Q_{k,l}^m$ by (independent) permutations of rows and columns. Consider the corresponding orbit set $Q_{k,l}^m/G$ , and set $\mathbf {p}_{k,l}^m:=|Q_{k,l}^m|$ and $\mathbf {q}_{k,l}^m:=|Q_{k,l}^m/G|$ . Note that for $\mathbf {p}_{k,l}^m$ or $\mathbf {q}_{k,l}^m$ to be nonzero, we must have the inequalities $\max \{k,l\}\leq m\leq kl$ . The cardinalities of stabilizers in G of elements of a fixed orbit are equal (as these stabilizers are conjugate) and, for an orbit $\mathcal {O}\in Q_{k,l}^m/G$ , we denote this cardinality by $|\text {stab}_{\mathcal {O}}|$ .

Proposition 4.2. The $I_n^*$ -module $\Bbbk [\mathbf {L}_\rho ]\otimes _\Bbbk \Bbbk [\mathbf {L}_\sigma ]$ has a filtration whose subquotients are isomorphic to cell modules. Moreover, the multiplicity of $\Bbbk [\mathbf {L}_\tau ]$ in $\Bbbk [\mathbf {L}_\rho ]\otimes _\Bbbk \Bbbk [\mathbf {L}_\sigma ]$ equals

(4-1) $$ \begin{align} \mathbf{p}_{k,l}^m=\sum_{i=0}^k\sum_{j=0}^l (-1)^{k+l-i-j}\binom{k}{i}\binom{l}{j}\binom{ij}{m}. \end{align} $$

Proof. Similarly to Section 3.4, the existence of the filtration follows from the fact that the action of $S_\tau $ on the elements of the form $\xi \otimes \zeta $ described above is free.

To compute the multiplicity, for a fixed $\tau $ , we need to count the number of the elements of the form $\xi \otimes \zeta $ as above, with $\mathrm {cod}(\xi )=\tilde {\rho }$ and $\mathrm {cod}(\zeta )=\tilde {\sigma }$ , for which $\tilde {\rho }\cap \tilde {\sigma }=\tau $ .

Given $M\in Q_{k,l}^m$ , there are $m!$ bijections between the parts of $\tau $ and the ‘ $1$ ’-entries in M. If we fix such a bijection $\psi $ , we can read off the parts of $\tilde {\rho }$ by uniting the parts of $\tau $ along the rows of M. Also, we can read off the parts of $\tilde {\sigma }$ by uniting the parts of $\tau $ along the columns of M. This determines the codomains of $\xi $ and $\zeta $ , respectively (note that there is no fixed order on the parts of these codomains, which explains the equivalence with respect to the action of G). As $\rho $ and $\sigma $ are fixed, to determine $\xi $ and $\zeta $ , it remains to choose $\overline {\xi }$ , in $k!$ different ways, and $\overline {\zeta }$ , in $l!$ different ways.

Now let $\mathcal {Y}(M,\psi )$ be the set of all $\xi \otimes \zeta $ described above. Let $\mathcal {Y}(M)$ be the union of $\mathcal {Y}(M,\psi )$ over all bijections $\psi $ between the parts of $\tau $ and the ‘ $1$ ’-entries of M. If M and N belong to the same G-orbit $\mathcal {O}$ , then $\mathcal {Y}(M)=\mathcal {Y}(N)$ ; we denote this set by $\mathcal {Y}_{\mathcal {O}}$ . The cardinality of $\mathcal {Y}_{\mathcal {O}}$ is ${m!\cdot k!\cdot l! }/{|\text {stab}_{\mathcal {O}}|}$ . Let $\mathcal {Y}$ be union of $\mathcal {Y}_{\mathcal {O}}$ over the orbits $\mathcal {O}$ in $Q^{m}_{k,l}/G$ . Note that the set $\mathcal {Y}$ consists of $\xi \otimes \zeta $ with $\mathrm {cod}(\xi )\cap \mathrm {cod}(\zeta )=\tau $ with $\xi \in \mathbf {L}_\rho $ and $\zeta \in \mathbf {L}_\sigma $ , and it has the cardinality:

$$ \begin{align*} m!\cdot\sum_{\mathcal{O}\in Q_{k,l}^m/G} \frac{k!\cdot l!}{|\mathrm{stab}_{\mathcal{O}}|}. \end{align*} $$

Since the action of $S_\tau $ on $\mathcal {Y}$ is free, dividing by $|S_\tau |=m!$ gives

$$ \begin{align*} \sum_{\mathcal{O}\in Q_{k,l}^m/G} \frac{k!\cdot l!}{|\mathrm{stab}_{\mathcal{O}}|} =\mathbf{p}_{k,l}^m. \end{align*} $$

The right-hand side of (4-1) is obtained using the inclusion-exclusion principle with respect to the requirement that all rows and all columns of elements in $Q_{k,l}^m$ should be nonzero.

It would be interesting to have a combinatorial formula for $\mathbf {q}_{k,l}^m$ . Although it is easy to determine $\mathbf {q}_{k,l}^m$ in some special cases (for instance, $\mathbf {q}_{k,l}^{kl}=1$ and $\mathbf {q}_{k,l}^{\max \{k,l\}}$ is the number of partitions of $\max \{k,l\}$ with exactly $\min \{k,l\}$ parts), we do not know the answer in general.

In general it is easy to see that if $\mathbf {q}^m_{k,l}=1,$ then the multiplicity of $\Bbbk [\mathbf {L}_\tau ]$ in $\Bbbk [\mathbf {L}_\rho ]\otimes \Bbbk [\mathbf {L}_\sigma ]$ is the cardinality of the set $Q^m_{k,l}$ .

4.4. Tensor product of simple modules

The results of Section 4.3 suggest what should be done to understand tensor products of simple $I_n^*$ -modules. This seems to be a rather difficult problem in the general case. Below we outline how to apply Section 4.3 to this problem in more detail.

For $0\leq k,l,m\leq n$ , consider the corresponding set $Q_{k,l}^m$ . Fix:

• • a set partition $\rho $ of $\underline {n}$ with exactly k parts;

• • a set partition $\sigma $ of $\underline {n}$ with exactly l parts;

• • a set partition $\tau $ of $\underline {n}$ with exactly m parts.

Recall the set $\mathcal {Y}$ from the proof of Proposition 4.2. Note that for $\xi \otimes \zeta \in \mathcal {Y}$ , we have $\varepsilon _{\tau } (\xi \otimes \zeta )=\xi \otimes \zeta $ and, in fact, $\varepsilon _\tau $ is the maximum (with respect to $\preceq $ ) element in the set of all idempotents in $I_n^*$ that do not annihilate $\xi \otimes \zeta $ . Also, note that $\Bbbk \,\mathcal {Y}$ has the natural structure of a $\Bbbk [S_m]$ - $\Bbbk [S_k\times S_l]$ -bimodule, where the group $S_m$ acts on the left by permuting the parts of $\tau $ , the group $S_k$ acts on the right by permuting the parts of $\rho $ and the group $S_l$ acts on the right by permuting the parts of $\sigma $ .

Similarly to (3-5), for $\lambda \vdash k$ , $\mu \vdash l$ and $\nu \vdash m$ ,

$$ \begin{align*} [\mathbf{N}^\lambda\otimes_\Bbbk\mathbf{N}^\mu:\mathbf{N}^\nu]= [\Bbbk \,\mathcal{Y}\otimes_{\Bbbk[S_k\times S_l]} (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu):\mathbf{S}^\nu]. \end{align*} $$

By construction, the bimodule $\Bbbk \,\mathcal {Y}$ splits into a direct sum of subbimodules given by varying M from the previous paragraph inside a fixed $S_k\times S_l$ -orbit in $Q_{k,l}^m$ . For a fixed such orbit $\mathcal {O}$ , we denote by $\Bbbk \mathcal {Y}_{\mathcal {O}}$ the corresponding direct summand of $\Bbbk \,\mathcal {Y}$ . Therefore, the real problem is to determine the multiplicities

$$ \begin{align*} [\Bbbk \,\mathcal{Y}_{\mathcal{O}}\otimes_{\Bbbk[S_k\times S_l]} (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu):\mathbf{S}^\nu] \end{align*} $$

for each $\mathcal {O}\in Q_{k,l}^m/_{S_k\times S_l}$ . Below we give a more group-theoretic reformulation of this problem and discuss various examples of different orbits. An explicit solution in all cases seems to be quite difficult.

In the above construction, assume that $\mathcal {O}=(S_k\times S_l)\cdot M$ and let H be the stabilizer of M in $S_k\times S_l$ . Consider the set A of all possible bijections between the elements of $\underline {m}$ and the ‘ $1$ ’-entries in M. Then each element of H gives rise to a permutation of the ‘ $1$ ’-entries in M, defining a homomorphism from H to $S_m$ . Moreover, a permutation in $S_m$ corresponding to a nontrivial element of H is, obviously, nontrivial. Therefore, this homomorphism is injective and thus we can identify H as a subgroup of $S_m$ .

Proposition 4.4. We have

$$ \begin{align*} [\Bbbk \,\mathcal{Y}_{\mathcal{O}}\otimes_{\Bbbk[S_k\times S_l]} (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu): \mathbf{S}^\nu]= [\mathrm{Ind}^{S_m}_H\mathrm{Res}^{S_k\times S_l}_H (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu) :\mathbf{S}^\nu]. \end{align*} $$

Proof. For $G=S_k\times S_l$ , the $\Bbbk [S_m]$ - $\Bbbk [G]$ -bimodule corresponding to the functor $\mathrm {Ind}^{S_m}_H\mathrm {Res}^G_H$ is given by

$$ \begin{align*} B:=\Bbbk[S_m]\otimes_{\Bbbk[H]}\Bbbk[G]. \end{align*} $$

Note that the dimension of this bimodule is ${|S_m||G|}/{|H|}$ and that this coincides with the dimension of $\Bbbk \,\mathcal {Y}_{\mathcal {O}}$ . From our definition of H, it follows that the right action of H on $\mathcal {Y}_{\mathcal {O}}$ (that is, the action in terms of G) coincides with the left action of H on $\mathcal {Y}_{\mathcal {O}}$ (that is, the action in terms of $S_m$ ).

Therefore, sending $e\otimes e\in B$ to the element in $\mathcal {Y}_{\mathcal {O}}$ corresponding to $(M,\psi )$ gives rise to a homomorphism of $\Bbbk [S_m]$ - $\Bbbk [G]$ -bimodules from B to $\Bbbk \,\mathcal {Y}_{\mathcal {O}}$ . Since the action of $S_m\times G$ on $\mathcal {Y}_{\mathcal {O}}$ is transitive, this homomorphism is surjective. By comparing the dimensions, we obtain that this homomorphism is an isomorphism.

As an immediate corollary of Propositions 4.2 and 4.4, we have the following corollary.

Example 4.6. Consider the extreme case $m=kl$ . In this case, $G=H$ is a subgroup of $S_{m}$ and hence we are interested in the multiplicities

$$ \begin{align*} [\mathrm{Ind}^{S_m}_{S_k\times S_l} (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu): \mathbf{S}^\nu]. \end{align*} $$

Note that these are not the Littlewood–Richardson coefficients as our $S_k\times S_l$ is not a Young subgroup of $S_m$ . This example appears in [Reference RybaRy21] and is studied in more detail in Section 6. The extreme example of this case is $k=1$ , when we get $H=G=S_m$ .

Example 4.8. Consider the case $m=(l-1)+(k-1)$ and take M such that the first row of M is $(0,1,1,\ldots ,1)$ and the first column of M is $(0,1,1,\ldots ,1)^t$ . Then the remaining entries of M are zero. In this case, $H\cong S_{k-1}\times S_{l-1}$ is both a Young subgroup of G and a Young subgroup of $S_m$ . Let us denote by $\to $ the usual branching relation on partitions (that is, $\xi \to \eta $ means that $\xi $ is obtained from $\eta $ by removing a removable node in the Young diagram). Then the multiplicities described in Proposition 4.4 are given by

$$ \begin{align*} \sum_{\alpha\to \lambda}\sum_{\beta\to \mu}\mathbf{c}_{\alpha,\beta}^\nu. \end{align*} $$

5.1. The semigroup $PI_n^*$

Consider the semigroup $PI_n^*$ defined in [Reference Kudryavtseva and MaltcevKMal11, Section 2.2]. The elements of $PI_n^*$ are all possible bijections between quotients of subsets of $\underline {n}$ (equivalently, between subsets of quotients of $\underline {n}$ ) and the operation is the natural version of composition for such bijections. Here is an example for $n=8$ , where composition is from right to left and the elements are presented using the two-row notation for functions:

$$ \begin{align*} \left(\begin{array}{@{}cccc@{}}\{1\}&\{3,5\}&\{6,7\}&\{8\}\\ \{1,2\}&\{3,4\}&\{5\}&\{6,8\} \end{array}\right) \circ \left(\begin{array}{@{}cccc@{}}\{2,3\}&\{4\}&\{5,6\}&\{7,8\}\\ \{1,2\}&\{3\}&\{5\}&\{6,7,8\}\end{array}\right) = \left(\begin{array}{@{}cc@{}}\{4,5,6\}&\{7,8\}\\\{3,4\}&\{5,6,8\}\end{array}\right). \end{align*} $$

The main point in this multiplication is that as soon some element x in the product hits an undefined part of the other element, all elements connected to x become undefined. In the above example, this happens for $x=2$ in the right factor. We do refer the reader to [Reference Kudryavtseva and MaltcevKMal11, Section 2.2] for the full definition which is very technical. The semigroup $PI_n^*$ is, in some sense, a mixture between $IS_n$ and $I_n^*$ .

Each element $\xi $ in $PI_n^*$ is thus uniquely given by the following data:

• • a set partition $\xi _d$ of some subset X of $\underline {n}$ , called the domain of $\xi $ ;

• • a set partition $\xi _r$ of some subset Y of $\underline {n}$ (with the same number of parts as $\xi _d$ ), called the codomain of $\xi $ ;

• • a bijection $\overline {\xi }$ from $X/\xi _d$ to $Y/\xi _r$ .

The cardinality of $X/\xi _d$ is called the rank of $\xi $ and denoted $\mathrm {rank}(\xi )$ . We also call the same number the rank of $\xi _d$ .

Let $\xi ,\zeta \in PI_n^*$ . Recall, see for example [Reference Kudryavtseva and MaltcevKMal11, Proposition 3], the following description of Green’s relations $\mathcal {L}$ , $\mathcal {R}$ , $\mathcal {H}$ and $\mathcal {D}=\mathcal {J}$ for $PI_n^*$ :

• • $\xi \mathcal {L}\zeta $ if and only if $\mathrm {dom}(\xi )=\mathrm {dom}(\zeta )$ ;

• • $\xi \mathcal {R}\zeta $ if and only if $\mathrm {cod}(\xi )=\mathrm {cod}(\zeta )$ ;

• • $\xi \mathcal {H}\zeta $ if and only if $\mathrm {dom}(\xi )=\mathrm {dom}(\zeta )$ and $\mathrm {cod}(\xi )=\mathrm {cod}(\zeta )$ ;

• • $\xi \mathcal {J}\zeta $ if and only if $\mathrm {rank}(\xi )=\mathrm {rank}(\zeta )$ .

The idempotents of $PI_n^*$ are naturally identified with equivalence relations on subsets of $\underline {n}$ : given an equivalence relation $\rho $ on some subset X of $\underline {n}$ , the corresponding idempotent $\varepsilon _\rho $ is the identity map on $X/\rho $ . The maximal subgroup of $PI_n^*$ corresponding to $\rho $ is the symmetric group $S_\rho $ on the equivalence classes of $\rho $ .

The semigroup $PI_k^*$ appears in the Schur–Weyl duality for the k th tensor power of the direct sum of the natural and the trivial representations of $IS_n$ ; see [Reference Kudryavtseva and MazorchukKMaz09b, Theorem 2].

5.2. Cells and simple $PI_n^*$ -modules

Let $\rho $ be an equivalence relation on some subset X of $\underline {n}$ . Let $\mathbf {L}_\rho $ be a left cell (that is, an $\mathcal {L}$ -equivalence class) in $PI_n^*$ containing $\varepsilon _\rho $ .

The linearization $\Bbbk [\mathbf {L}_\rho ]$ is naturally a left $PI_n^*$ -module with the action given as follows, for $\xi \in PI_n^*$ and $\zeta \in \mathbf {L}_\rho $ :

$$ \begin{align*} \xi\cdot\zeta= \begin{cases} \xi\zeta& \text{ if } \xi\zeta\in \mathbf{L}_\rho,\\ 0& \text{ otherwise.} \end{cases} \end{align*} $$

This is the cell $PI_n^*$ -module corresponding to $\mathbf {L}_\rho $ .

The maximal subgroup $S_\rho $ acts on $\mathbf {L}_\rho $ by multiplication on the right. This equips $\Bbbk [\mathbf {L}_\rho ]$ with the natural structure of a $\Bbbk [PI_n^*]$ - $\Bbbk [S_\rho ]$ -bimodule.

For any $\lambda \vdash \mathrm {rank}(\varepsilon _\rho )$ , the $\Bbbk [PI_n^*]$ -module

$$ \begin{align*} \mathbf{N}^{\lambda}:=\Bbbk[\mathbf{L}_\rho]\bigotimes_{\Bbbk[S_\rho]} \mathbf{S}^\lambda \end{align*} $$

is simple. Up to isomorphism, it does not depend on the choice of $\mathbf {L}_\rho $ inside its $\mathcal {J}$ -cell (that is, for the fixed rank of $\varepsilon _\rho $ ). Moreover, up to isomorphism, each simple $\Bbbk [PI_n^*]$ -module has such a form for some $\rho $ and $\lambda $ . We refer to [Reference SteinbergSt16] for details.

5.3. Tensor product of cell modules

Let $\rho $ and $\sigma $ be two set partitions (also known as equivalence relations) of two subsets X and Y of $\underline {n}$ , respectively. The tensor product

$$ \begin{align*} \Bbbk[\mathbf{L}_\rho]\otimes_\Bbbk\Bbbk[\mathbf{L}_\sigma] \end{align*} $$

is a $PI_n^*$ -module via the diagonal action of $PI_n^*$ . In this subsection, we show that this module has a filtration whose subquotients are isomorphic to cell modules. Also, we describe a combinatorial object that determines the corresponding multiplicities (just like in the cases of $IS_n$ and $I_n^*$ , the fact that these multiplicities are well defined follows directly from the semi-simplicity of the semigroup algebra of $PI_n^*$ ).

Set $k=\mathrm {rank}(\varepsilon _\rho )$ , $l=\mathrm {rank}(\varepsilon _\sigma )$ . Let U be a subset of $\underline {n}$ of cardinality at least k, and $\tilde {\rho }$ a set partition of U of the same rank as $\rho $ . Let V be a subset of $\underline {n}$ of cardinality at least l, and $\tilde {\sigma }$ a set partition of V of the same rank as $\sigma $ . Consider the set partition $\tau $ of $U\cup V$ whose parts are exactly the following sets:

• • all nonempty intersections of a part of $\tilde {\rho }$ with a part of $\tilde {\sigma }$ ;

• • all nonempty intersections of a part of $\tilde {\rho }$ with $\underline {n}\setminus V$ ;

• • all nonempty intersections of a part of $\tilde {\sigma }$ with $\underline {n}\setminus U$ .

Let $\xi \in \mathbf {L}_\rho $ and $\zeta \in \mathbf {L}_\sigma $ be such that $\mathrm {cod}(\xi )=\tilde {\rho }$ and $\mathrm {cod}(\zeta )=\tilde {\sigma }$ . Then $\tau $ is the maximum element in the set of all set partitions $\omega $ of $U\cup V$ such that $\varepsilon _\omega \xi \in \mathbf {L}_\rho $ and $\varepsilon _\omega \zeta \in \mathbf {L}_\sigma $ . Note also that the action of $S_\tau $ on $\xi \otimes \zeta $ is, clearly, free. All this means exactly that the element $\xi \otimes \zeta $ contributes to a cell subquotient $\Bbbk [\mathbf {L}_\tau ]$ of $\Bbbk [\mathbf {L}_\rho ]\otimes _\Bbbk \Bbbk [\mathbf {L}_\sigma ]$ .

To provide a combinatorial object for counting the multiplicities, we need to invert our problem. Given some $\tau $ , we need to find the number of all possible $\tilde {\rho }$ and $\tilde {\sigma }$ as above, which are connected to $\tau $ also in the above sense.

Set $m:=\mathrm {rank}(\varepsilon _\tau )$ . Let $\tilde {Q}_{k,l}^{m}$ be the set of all $0/1$ -matrices whose rows are indexed by the elements in $\{0,1,2,\ldots ,k\}$ and whose columns are indexes by the elements in $\{0,1,2,\ldots ,l\}$ satisfying the following conditions:

• • the $(0,0)$ -entry is $0$ ;

• • exactly m entries are equal to $1$ ;

• • all rows with positive indices are nonzero;

• • all columns with positive indices are nonzero.

The group $G:=S_k\times S_l$ acts on $\tilde {Q}_{k,l}^m$ by (independent) permutation of rows and columns with positive indices. Consider the corresponding orbit set $\tilde {Q}_{k,l}^m/G$ , and set $\tilde {\mathbf {p}}_{k,l}^m:=|\tilde {Q}_{k,l}^m|$ and $\tilde {\mathbf {q}}_{k,l}^m:=|\tilde {Q}_{k,l}^m/G|$ . Note that for $\tilde {\mathbf {q}}_{k,l}^m$ to be nonzero, we must have the inequality $\max \{k,l\}\leq m\leq (k+1)(l+1)-1$ . It would be interesting to have a combinatorial formula for $\tilde {\mathbf {q}}_{k,l}^m$ .

Proposition 5.1. The module $\Bbbk [\mathbf {L}_\rho ]\otimes _\Bbbk \Bbbk [\mathbf {L}_\sigma ]$ has a filtration whose subquotients are isomorphic to cell modules. Moreover, the multiplicity of $\Bbbk [\mathbf {L}_\tau ]$ in $\Bbbk [\mathbf {L}_\rho ]\otimes _\Bbbk \Bbbk [\mathbf {L}_\sigma ]$ equals

$$ \begin{align*} \tilde{\mathbf{p}}_{k,l}^m= \sum_{i=0}^k\sum_{j=0}^l (-1)^{k+l-i-j}\binom{k}{i}\binom{l}{j}\binom{(i+1)(j+1)-1}{m}. \end{align*} $$

5.4. Tensor product of simple modules

Similarly to the case of the dual symmetric inverse monoid, the results of Section 5.3 suggest what should be done to understand tensor products of simple $PI_n^*$ -modules. This seems to be a rather difficult problem. Below we outline how to apply Section 5.3 to this problem in more detail.

Let $\rho $ be a set partition of $X\subset \underline {n}$ with k parts, and $\sigma $ a set partition of $Y\subset \underline {n}$ with l parts. Let $\tau $ be the set partition of $X\cup Y$ given by intersecting the parts of $\rho $ with the parts of $\sigma $ and, additionally, the parts of $\rho $ with $\underline {n}\setminus Y$ and the parts of $\sigma $ with $\underline {n}\setminus X$ . Then we have $\varepsilon _{\tau } (\varepsilon _{\rho }\otimes \varepsilon _{\sigma })=\varepsilon _{\rho }\otimes \varepsilon _{\sigma }$ and, in fact, $\varepsilon _\tau $ is the maximum (with respect to $\preceq $ ) element in the set of all idempotents in $PI_n^*$ that do not annihilate $\varepsilon _{\rho }\otimes \varepsilon _{\sigma }$ . Assume that $\tau $ has exactly m parts.

Fix some linear orderings for the parts of $\rho $ and for the parts of $\sigma $ . Then we have the matrix $M\in \tilde {Q}_{k,l}^m$ defined as follows:

• • for positive i and j, the $(i,j)$  th entry of M is $1$ if and only if the intersection of the i th part of $\rho $ with the j th part of $\sigma $ is not empty;

• • the $(0,j)$  th entry of M is $1$ if and only if the intersection of the j th part of $\sigma $ with $\underline {n}\setminus X$ is not empty;

• • the $(i,0)$  th entry of M is $1$ if and only if the intersection of the i th part of $\rho $ with $\underline {n}\setminus Y$ is not empty.

Consider the G-orbit $G\cdot M$ of M inside $\tilde {Q}_{k,l}^m$ and denote by $\mathbf {M}$ the set of all possible bijective decorations of the ‘ $1$ ’-entries of the elements in $G\cdot M$ by the elements of $\underline {m}$ . Clearly, both $S_m$ and G act on $\mathbf {M}$ and these two actions commute. Hence, the linearization $\Bbbk [\mathbf {M}]$ of $\mathbf {M}$ becomes an $S_m$ -G-bimodule.

Let $\lambda \vdash k$ , $\mu \vdash l$ and $\nu \vdash m$ . To understand tensor products of simple $PI_n^*$ -modules, Proposition 5.1 (more precisely, it’s proof based on the proof of Proposition 4.2) implies that we need to understand the multiplicity of $\mathbf {S}^\nu $ in the module

$$ \begin{align*} \Bbbk[\mathbf{M}]\otimes_{\Bbbk[G]} (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu). \end{align*} $$

Let H be the stabilizer of M in G. Consider the set A of all possible bijections between the elements of $\underline {m}$ and the ‘ $1$ ’-entries in M. Then each element of H gives rise to a permutation of the ‘ $1$ ’-entries in M, defining a homomorphism from H to $S_m$ . Moreover, the permutation in $S_m$ corresponding to a nontrivial element of H is, obviously, nontrivial. Therefore, this homomorphism is injective and thus we can canonically identify H as a subgroup of $S_m$ .

Proposition 5.3. We have

$$ \begin{align*} [\Bbbk[\mathbf{M}]\otimes_{\Bbbk[G]} (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu): \mathbf{S}^\nu]= [\mathrm{Ind}^{S_m}_H\mathrm{Res}^G_H (\mathbf{S}^\lambda\otimes_\Bbbk\mathbf{S}^\mu) :\mathbf{S}^\nu]. \end{align*} $$

Example 5.6. Consider the case $k=l$ and $m=3k$ , where M is the matrix in which all ‘1’-entries are either in the $0\,\mathrm{th}$ row or in the $0\,\mathrm{th}$ column or on the main diagonal. In this case, H is isomorphic to $S_k$ and is diagonally embedded both into G and into the Young subgroup of $S_{m}$ isomorphic to $S_k\times S_k\times S_k$ . Since these embeddings are diagonal, the restriction from G to H and the induction from H to $S_k\times S_k\times S_k$ are computed using Kronecker coefficients. After that, the induction from ${S_k\times S_k\times S_k}$ to $S_m$ is computed using Littlewood–Richardson coefficients. Consequently, the multiplicities in Proposition 5.3 are given by the following expression:

$$ \begin{align*} \sum_{\alpha\vdash k}\sum_{\beta\vdash k} \sum_{\gamma\vdash k}\sum_{\delta\vdash k} \sum_{\epsilon\vdash k}\sum_{\kappa\vdash 2k} \mathbf{c}_{\epsilon,\kappa}^\nu \mathbf{c}_{\gamma,\delta}^\kappa \mathbf{g}_{\delta,\epsilon}^\beta \mathbf{g}_{\beta,\gamma}^\alpha \mathbf{g}_{\lambda,\mu}^\alpha.\end{align*} $$

6.1. Setup

In this section, we look at Example 4.6 in more detail. Let $m=kl$ and let M be the $k\times l$ matrix in which all entries are equal to $1$ . We fix the bijection between $\underline {m}$ and the entries of this matrix that reads the entries left-to-right and then top-to-bottom. The group $S_k$ acts by permuting the rows of M, which identifies it as a subgroup of $S_m$ . The group $S_l$ acts by permuting the columns of M, which identifies it as a subgroup of $S_m$ . Put together, we have identified $S_k\times S_l$ as a subgroup of $S_{kl}$ .

For $\lambda \vdash k$ , $\mu \vdash l$ and $\nu \vdash kl$ , denote by $\mathbf {b}^\nu _{\lambda ,\mu }$ the multiplicity of $\mathbf {S}^\lambda \otimes _\Bbbk \mathbf {S}^\mu $ in the restriction of $\mathbf {S}^\nu $ from $S_{kl}$ to $S_k\times S_l$ .

6.2. Partition algebra

For $n\geq 0$ and $x\in \Bbbk \setminus \{0\}$ , consider the corresponding partition algebra $\mathcal {P}_n(x)$ , as defined in [Reference JonesJo, Reference MartinMar94]. It is a $\Bbbk $ -algebra that has a basis consisting of set partitions of two (the upper and the lower) copies of $\underline {n}$ . Such set partitions are usually represented by certain diagrams (partition diagrams) whose connected components, viewed as graphs, represent the parts of the partition, for example:

Note that such a diagram representing a partition is not unique. Algebra multiplication is defined in terms of concatenation of such diagrams followed by a certain straightening procedure that results in a new diagram and a nonnegative integer r (the number of parts removed during the straightening procedure). The product is then this new diagram times $x^r$ . Here is an example:

In this example, the part $\{2\}$ in the bottom of the top diagram and the part $\{2\}$ in the top of the bottom diagram together form a unique part that is removed in the straightening procedure. We refer the reader to [Reference Halverson and RamHR05] for details.

Given a partition diagram $\mathbf {d}$ , the parts that intersect both copies of $\underline {n}$ are called propagating lines and the number of such parts is called rank. This rank is an integer between $0$ and n. For $0\leq i\leq n$ , all partition diagrams of rank at most i span a two-sided ideal in $\mathcal {P}_n(x)$ , which we denote by $I_i$ . Given a simple $\mathcal {P}_n(x)$ -module L, there is a minimal i such that $I_i$ does not belong to the annihilator of L. Simple $\mathcal {P}_n(x)$ -modules with this property are in a natural bijection with partitions of i and are constructed as follows.

Let $\mathbf {d}$ be a partition diagram of rank i. Consider the set $\mathbf {L}_{\mathbf {d}}$ of all partition diagrams of rank i that can be obtained from $\mathbf {d}$ by left concatenation with other diagrams. Then the linearization $\Bbbk \mathbf {L}_{\mathbf {d}}$ is naturally a $\mathcal {P}_n(x)$ - $\Bbbk [S_i]$ -bimodule, which is free as a $\Bbbk [S_i]$ -module. For $\lambda \vdash i$ , the module

$$ \begin{align*} \mathbf{C}^\lambda:=\Bbbk\mathbf{L}_{\mathbf{d}} \bigotimes_{\Bbbk[S_i]}\mathbf{S}^\lambda \end{align*} $$

has a simple top that we denote by $\mathbf {N}^\lambda $ . The set

$$ \begin{align*} \{ \mathbf{N}^\lambda\,:\, \lambda\vdash i, \, 0\leq i\leq n \} \end{align*} $$

is a full and irredundant set of representatives of isomorphism classes of simple $\mathcal {P}_n(x)$ -modules.

We also note that $\mathbf {C}^\lambda \cong \mathbf {N}^\lambda $ provided that $x\not \in \{0,1,2,\ldots ,2n-2\}$ since in this case, the partition algebra is semi-simple; see [Reference Martin and SaleurMarS94] or [Reference Halverson and RamHR05, Theorem 3.27]. Note that for $x\in \{0,1,2,\ldots ,2n-2\}$ , the isomorphism $\mathbf {C}^\lambda \cong \mathbf {N}^\lambda $ might still hold for some $\lambda $ . In particular, under the assumption $x\not \in \{0,1,2,\ldots ,2n\}$ , the module $\mathbf {N}^\lambda $ has a basis that does not depend on x and in which the action of each partition diagram is given by a matrix in which all coefficients are polynomials in x.

We call the modules $\mathbf {C}^\lambda $ standard. A standard filtration of some $\mathcal {P}_n(x)$ -module is a filtration whose subquotients are standard modules. This terminology reflects the fact that the algebra $\mathcal {P}_n(x)$ is quasi-hereditary (here our assumption $x\neq 0$ is important).

6.3. Schur–Weyl dualities for partition algebras

Consider the natural $S_k$ -module $\Bbbk ^k$ and its n-fold tensor power $(\Bbbk ^k)^{\otimes n}$ . It is an $S_k$ -module via the diagonal action of $S_k$ . The algebra $\mathcal {P}_n(k)$ acts naturally on $(\Bbbk ^k)^{\otimes n}$ , see for example, [Reference Halverson and RamHR05, (3.2)], this action commutes with the action of $S_k$ and, moreover, the two commuting actions generate each other’s centralizers. This is known as the Schur–Weyl duality for partition algebras; see [Reference Halverson and RamHR05, Section 3]. As a $\Bbbk [S_k]$ - $\mathcal {P}_n(k)$ bimodule, the space $(\Bbbk ^k)^{\otimes n}$ decomposes into a direct sum of simple bimodules indexed by those partitions $\lambda =(\lambda _1,\lambda _2,\ldots )$ of k for which $k-\lambda _1\leq n$ . The corresponding simple $\Bbbk [S_k]$ - $\mathcal {P}_n(k)$ bimodule is then $\mathbf {S}^\lambda \otimes _\Bbbk \mathbf {N}^{\lambda _{>1}}$ , where $\lambda _{>1}$ is the partition of $k-\lambda _1$ obtained from $\lambda $ by deleting the first part $\lambda _1$ .

We note that neither the action of $\Bbbk [S_k]$ nor the action of $\mathcal {P}_n(k)$ on $(\Bbbk ^k)^{\otimes n}$ is faithful, in general.

Now let $m=kl$ . Then we have the natural $S_k$ -module $\Bbbk ^k$ and the natural $S_l$ -module $\Bbbk ^l$ . Their tensor product $\Bbbk ^k\otimes _{\Bbbk }\Bbbk ^l$ is isomorphic to the natural $S_m$ -module $\Bbbk ^m$ . For any $n\geq 0$ , we also have the obvious isomorphism

$$ \begin{align*} (\Bbbk^k\otimes_{\Bbbk}\Bbbk^l)^{\otimes n} \cong (\Bbbk^k)^{\otimes n}\otimes_{\Bbbk}(\Bbbk^l)^{\otimes n}. \end{align*} $$

By the Schur–Weyl duality, the $S_{kl}$ -endomorphisms of the left-hand side are given by the image of $\mathcal {P}_n(kl)$ . Similarly, the $S_{k}\times S_l$ -endomorphisms of the right-hand side are given by the image of $\mathcal {P}_n(k)\otimes _{\Bbbk }\mathcal {P}_n(l)$ . This suggests that our original problem of restriction from $S_{kl}$ to $S_{k}\times S_l$ should be interpretable, via the Schur–Weyl duality, in terms of some induction problem for partition algebras. This is what we investigate in the next subsections.

6.4. Generalized comultiplication for partition algebras

We do not know of any interesting comultiplication for partition algebras. However, the following proposition describes a generalized version of such comultiplication.

6.5. Partition algebra and $\mathbf {b}^\nu _{\lambda ,\mu }$

Proposition 6.1 implies that, given a $\mathcal {P}_n(a)$ -module X and a $\mathcal {P}_n(b)$ -module Y, the tensor product $X\otimes _{\Bbbk } Y$ is, naturally, a $\mathcal {P}_n(ab)$ -module.

Proof. By additivity, it is enough to prove the claim under the assumption that both X and Y are cell modules.

Let $\mathbf {L}_1$ and $\mathbf {L}_2$ be two left cells such that $X=\Bbbk [\mathbf {L}_1]$ and $X=\Bbbk [\mathbf {L}_2]$ . Then X has the canonical basis consisting of all partition diagrams in $\mathbf {L}_1$ and Y has the canonical basis consisting of all partition diagrams in $\mathbf {L}_2$ . Therefore, $X\otimes _{\Bbbk } Y$ has the canonical basis consisting of all elements of the form $\mathbf {d}_1\otimes \mathbf {d}_2$ , where $\mathbf {d}_i\in \mathbf {L}_i$ for $i=1,2$ .

Consider the oriented graph $\Gamma $ whose vertices are all these elements in the basis of $X\otimes _{\Bbbk } Y$ . For two vertices $\mathbf {d}_1\otimes \mathbf {d}_2$ and $\mathbf {d}^{\prime }_1\otimes \mathbf {d}^{\prime }_2$ , put an oriented edge from $\mathbf {d}_1\otimes \mathbf {d}_2$ to $\mathbf {d}^{\prime }_1\otimes \mathbf {d}^{\prime }_2$ provided that there is a partition diagram $\tilde {\mathbf {d}}$ which, when applied to $\mathbf {d}_1\otimes \mathbf {d}_2$ , outputs $\mathbf {d}^{\prime }_1\otimes \mathbf {d}^{\prime }_2$ , up to a nonzero scalar. Note that $\Gamma $ does not depend on $x\in \{a,b\}$ , since we assume $x\neq 0$ . Also note that all vertices have loops as we can choose $\tilde {\mathbf {d}}$ to be the identity partition diagram.

Since $\Gamma $ is finite, we can choose a filtration of $\Gamma $ of the form

$$ \begin{align*} \varnothing =\Gamma_0\subset \Gamma_1\subset \cdots\subset \Gamma_k=\Gamma \end{align*} $$

by nonempty full subgraphs such that the following two conditions are satisfied, for each i:

• • if $\alpha $ is an edge from u to v and $u\in \Gamma _i$ , then $v\in \Gamma _i$ ;

• • the full subgraph of $\Gamma _i$ whose vertices are all vertices outside $\Gamma _{i-1}$ is strongly connected.

Then the linearization $\Bbbk [\Gamma _i]$ of the vertex set of each $\Gamma _i$ is a submodule of $X\otimes _{\Bbbk } Y$ , so we have the induced filtration of $X\otimes _{\Bbbk } Y$ by submodules

$$ \begin{align*} 0\subset \Bbbk[\Gamma_1]\subset \Bbbk[\Gamma_2]\subset\cdots \subset X\otimes_{\Bbbk} Y \end{align*} $$

and this filtration, by construction, does not depend on x.

We claim that each $\mathcal {P}_n(ab)$ -module $N_i:=\Bbbk [\Gamma _i]/\Bbbk [\Gamma _{i-1}]$ has a filtration by standard modules. Let j be the minimal rank for which there is a partition diagram that does not annihilate $\Bbbk [\Gamma _i]/\Bbbk [\Gamma _{i-1}]$ . Choose some idempotent partition diagram $\tilde {\mathbf {d}}$ of rank j and fix some $\mathbf {d}_1\otimes \mathbf {d}_2\in \Gamma _i\setminus \Gamma _{i-1}$ that is fixed by $\tilde {\mathbf {d}}$ (this exists as $\tilde {\mathbf {d}}$ is an idempotent and it does not annihilate $N_i$ ). Applying the left cell of $\tilde {\mathbf {d}}$ to $\mathbf {d}_1\otimes \mathbf {d}_2$ gives a set of vertices that is invariant under the action of all partition diagrams (modulo $\Gamma _{i-1}$ ). Hence, this must coincide with the whole of $\Gamma _i\setminus \Gamma _{i-1}$ due to the strong connectivity of the latter. In reality, we just proved that $N_i$ is a quotient of $\Bbbk [\mathbf {L}_{\tilde {\mathbf {d}}}]$ since we have a natural surjective homomorphism from the latter to the former that sends $\tilde {\mathbf {d}}'\in \mathbf {L}_{\tilde {\mathbf {d}}}$ to $\tilde {\mathbf {d}}'\cdot (\mathbf {d}_1\otimes \mathbf {d}_2)$ .

Let G be the $\mathcal {H}$ -class of $\tilde {\mathbf {d}}$ . Then $\mathbf {L}_{\tilde {\mathbf {d}}}$ is a free right G-act. Let H be the stabilizer of $\mathbf {d}_1\otimes \mathbf {d}_2$ in G. Then the above gives a surjection from the module

(6-1) $$ \begin{align} \Bbbk[\mathbf{L}_{\tilde{\mathbf{d}}}]\otimes_{\Bbbk[G]}\Bbbk[G/H] \end{align} $$

to $N_i$ . Note that the module (6-1) has a filtration by standard modules by construction.

Everything above is independent of x, as long as $x\neq 0$ . For generic x, standard modules are simple and hence the surjection in the previous paragraph is, in fact, an isomorphism between the module (6-1) and $N_i$ . However, then the fact that this surjection is an isomorphism does not depend on x, completing the proof.

From the theory of quasi-hereditary algebras, we know that the multiplicity of a standard subquotient in a standard filtration of some module that admits a standard filtration does not depend on the choice of such a filtration.

Let us now go back to the setup as described in Section 6.1.

Corollary 6.5. Let $\lambda \vdash k$ , $\mu \vdash l$ and $\nu \vdash kl$ . Then, for

$$ \begin{align*} n\geq \max(kl-\nu_1,k-\lambda_1,l-\mu_1), \end{align*} $$

the number $\mathbf {b}^\nu _{\lambda ,\mu }$ coincides with the multiplicity of the $\mathcal {P}_n(kl)$ -module $\mathbf {N}^{\nu _{>1}}$ in the tensor product of the $\mathcal {P}_n(k)$ -module $\mathbf {N}^{\lambda _{>1}}$ and the $\mathcal {P}_n(l)$ -module $\mathbf {N}^{\mu _{>1}}$ .

6.6. Stability phenomenon for $\mathbf {b}^\gamma _{\mu ,\nu }$

For a partition $\lambda \vdash n$ and a positive integer a, we denote by $\lambda ^{(a)}$ the partition of $n+a$ obtained from $\lambda $ by increasing $\lambda _1$ to $\lambda _1+a$ . The following claim was also proved in [Reference RybaRy21, Theorem 4.2] by completely different arguments.

Proof. Let $n\geq \max (kl-\nu _1,k-\lambda _1,l-\mu _1)$ and note that

$$ \begin{align*} \max(kl-\nu_1,k-\lambda_1,l-\mu_1)= \max((kl+la)-\nu^{(la)}_1,(k+a)-\lambda^{(a)}_1,l-\nu_1) \end{align*} $$

for all $a>0$ . This allows us to use the interpretation of $\mathbf {b}^{\nu ^{(la)}}_{\lambda ^{(a)},\mu }$ given by Corollary 6.5. We have $\nu ^{(la)}_{>1}=\nu _{>1}$ and $\lambda ^{(a)}_{>1}=\lambda _{>1}$ for all $a>0$ . In particular, on the partition algebra side, all involved modules have exactly the same corresponding indices, regardless of $a>0$ .

Next we note that for all $a\gg 0$ , both algebras $\mathcal {P}_n((k+a)l)$ and $\mathcal {P}_n(k+a)$ are semi-simple; moreover, each simple module $\mathbf {N}^\gamma $ for each of these algebras coincides with the corresponding $\mathbf {C}^\gamma $ .

The multiplicity of a simple module over a semi-simple finite dimensional algebra can be computed as the rank of a certain primitive idempotent. Generic primitive idempotents for partition algebras were constructed in [Reference Martin and WoodcockMW99]. Outside of a finite number of values of the parameter, each such idempotent is given as a linear combination of diagrams whose coefficients are rational functions in the parameter of the partition algebra. We can take $a\gg 0$ such that the construction of these primitive idempotents applies (meaning that the denominators are nonzero). If we fix such a primitive idempotent for $\mathcal {P}_n((k+a)l)$ , then the action of this idempotent on the tensor product of some $\mathcal {P}_n(k+a)$ -module $\mathbf {C}^\gamma $ (in which we fix the standard diagram basis) with a fixed $\mathcal {P}_n(l)$ -module is given by a matrix whose coefficients are rational functions in a. Hence, this matrix has maximal rank for all but finitely many values of a. The claim follows.

6.7. Tensor product of cell modules for partition algebras

Unlike the cases of the (dual) symmetric inverse semigroups, the tensor product of cell modules for partition algebras does not have to have a filtration whose subquotients are cell modules. In this subsection, we give an explicit example illustrating how this property fails already for $n=2$ .

The algebra $\mathcal {P}_2(x)$ has dimension $15$ (the bell number $\mathbf {B}_4$ ). As usual, we assume $x\neq 0$ . The possible ranks for partition diagrams are $0,1$ and $2$ . We have one left cell $\mathcal {L}_2$ of rank two, consisting of the following two diagrams:

We have three left cells $\mathcal {L}_1^{(i)}$ , where $i=1,2,3$ , of rank one, each containing three elements, given by the following rows:

Finally, we have two left cells $\mathcal {L}_0^{(i)}$ , where $i=1,2$ , of rank zero, each containing two elements, given by the following rows:

In particular, we have, up to isomorphism, three cell modules $\Bbbk [\mathcal {L}_2]$ , $\Bbbk [\mathcal {L}_1^{(1)}]$ and $\Bbbk [\mathcal {L}_0^{(1)}]$ , of respective dimensions $2$ , $3$ and $2$ .

Now we claim that $\Bbbk [\mathcal {L}_1^{(1)}]\otimes _\Bbbk \Bbbk [\mathcal {L}_0^{(1)}]$ does not have a cell filtration. Indeed, this module has dimension six with the following basis:

It is easy to check that $v_1$ , $v_2$ and $w_3$ span in $\Bbbk [\mathcal {L}_1^{(1)}]\otimes _\Bbbk \Bbbk [\mathcal {L}_0^{(1)}]$ a submodule isomorphic to $\Bbbk [\mathcal {L}_1^{(1)}]$ . In the quotient, $w_1$ and $w_2$ span a submodule isomorphic to $\Bbbk [\mathcal {L}_2]$ , while $v_3$ only spans a submodule of $\Bbbk [\mathcal {L}_2]$ , as the latter has dimension two and we only have one vector left.