ELECTRICAL NETWORKS AND LIE THEORY

THOMAS LAM AND PAVLO PYLYAVSKYY

Abstract. We introduce a new class of “electrical” Lie groups. These Lie groups, or more
precisely their nonnegative parts, act on the space of planar electrical networks via com-
arXiv:1103.3475v2 [math.RT] 19 Mar 2011

binatorial operations previously studied by Curtis-Ingerman-Morrow. The corresponding

electrical Lie algebras are obtained by deforming the Serre relations of a semisimple Lie al-
gebra in a way suggested by the star-triangle transformation of electrical networks. Rather
surprisingly, we show that the type A electrical Lie group is isomorphic to the symplectic
group. The nonnegative part (EL2n )≥0 of the electrical Lie group is a rather precise ana-
logue of the totally nonnegative subsemigroup (Un )≥0 of the unipotent subgroup of SLn .
We establish decomposition and parametrization results for (EL2n )≥0 , paralleling Lusztig’s
work in total nonnegativity, and work of Curtis-Ingerman-Morrow and de Verdière-Gitler-
Vertigan for networks. Finally, we suggest a generalization of electrical Lie algebras to all
Dynkin types.

1. Introduction
In this paper we consider the simplest of electrical networks, namely those that consist
of only resistors. The electrical properties of such a network N are completely described
by the response matrix L(N), which computes the current that flows through the network
when certain voltages are fixed at the boundary vertices of N. The study of the response
matrices of planar electrical networks has led to a robust theory, see Curtis-Ingerman-Morrow
[CIM], or de Verdière-Gitler-Vertigan [dVGV]. In 1899, Kennelly [Ken] described a local
transformation (see Figure 3) of a network, called the star-triangle or Y − ∆ transformation,
which preserves the response matrix of a network. This transformation is one of the many
places where a Yang-Baxter style transformation occurs in mathematics or physics.
Curtis-Ingerman-Morrow [CIM] studied the operations of adjoining a boundary spike and
adjoining a boundary edge to (planar) electrical networks. Our point of departure is to
consider these operations as one-parameter subgroups of a Lie group action. The star-
triangle transformation then leads to an “electrical Serre relation” in the corresponding Lie
algebra, which turns out to be a deformation of the Chevalley-Serre relation for sln :
Serre relation: [e, [e, e! ]] = 0 electrical Serre relation: [e, [e, e! ]] = −2e
The corresponding one-parameter subgroups satisfy a Yang-Baxter style relation which is a
deformation of Lusztig’s relation in total positivity:
Lusztig’s relation: ui (a)uj (b)ui (c) = uj (bc/(a + c))ui (a + c)uj (ab/(a + c))
electrical relation: ui (a)uj (b)ui (c) = uj (bc/(a + c + abc))ui (a + c + abc)uj (ab/(a + c + abc)).

T.L. was partially supported by NSF grant DMS-0901111, and by a Sloan Fellowship.
P.P. was partially supported by NSF grant DMS-0757165.
1
We study in detail the Lie algebra el2n with 2n generators satisfying the electrical Serre
relation. An (electrically) nonnegative part (EL2n )≥0 of the corresponding Lie group EL2n
acts on planar electrical networks with n + 1 boundary vertices, and one obtains a dense
subset of all response matrices of planar networks in this way. This nonnegative part is a
rather precise analogue of the totally nonnegative subsemigroup (U2n+1 )≥0 of the unipotent
subgroup of SL2n+1 , studied in Lie-theoretic terms by Lusztig [Lus]. We show (Proposition
4.2) that (EL2n )≥0 has a cell decomposition labeled by permutations w ∈ S2n+1 , precisely
paralleling one of Lusztig’s results for (U2n )≥0 and reminiscent of the Bruhat decomposition.
This can be considered an algebraic analogue of parametrization results in the theory of
electrical networks [CIM, dVGV]. Surprisingly, EL2n itself is isomorphic to the symplectic
Lie group Sp2n (R) (Theorem 3.1). This semisimplicity does not in general hold for electrical
Lie groups: EL2n+1 is not semisimple. We also describe (Theorem 4.10) the stabilizer lie
algebra of the infinitesimal action of el2n on electrical networks.
While we focus on the planar case in this paper, we shall connect the results of this paper
to the inverse problem for electrical networks on cylinders in a future paper. There we
borrow ideas from representation theory such as that of crystals and R-matrices.

One obtains the types B and G electrical Serre relations by the standard technique of
“folding” the type A electrical Serre relation. These lead to a new species of electrical
Lie algebras eD defined for any Dynkin diagram D. Besides the above results for type A,
there appear to be other interesting relations between these eD and simple Lie algebras. For
example, eb2 := eB2 is isomorphic to gl2 . We conjecture (Conjecture 5.2) that for a finite type
diagram D, the dimension dim(eD ) is always equal to the dimension of the maximal nilpotent
subalgebra of the semisimple Lie algebra gD with Dynkin diagram D, and furthermore
(Conjecture 5.3) that eD is finite dimensional if and only if gD is finite-dimensional. We
give the electrical analogue of Lusztig’s relation for type B (see Berenstein-Zelevinsky [BZ])
where we observe similar positivity to the totally nonnegative case.
The most interesting case beyond finite Dynkin types is affine type A. The corresponding
electrical Lie (semi)-group action on planar electrical networks is perhaps even more natural
than that of EL2n , since one can obtain all (rather than just a dense subset of) response
matrices of planar networks in this way ([dVGV, CIM]), and furthermore the circular sym-
metry of the planar networks is preserved. We however do not attempt to address this case
in the current paper.

2. Electrical networks
For more background on electrical networks, we refer the reader to [CIM, dVGV].

2.1. Response matrix. For our purposes, an electrical network is a finite weighted undi-
rected graph N, where the vertex set is divided into the boundary vertices and the interior
vertices. The weight w(e) of an edge is to be thought of as the conductance of the corre-
sponding resistor, and is generally taken to be a positive real number. A 0-weighted edge
would be the same as having no edge, and an edge with infinite weight would be the same
as identifying the endpoint vertices.
2
 We define the Kirchhoff matrix K(N) to be the square matrix with rows and columns
labeled by the vertices of N as follows:
! "
− w(e) for i #= j
Kij = " e joins i and j
e incident to i w(e) for i = j.

We define the response matrix L(N) to be the square matrix with rows and columns
labeled by the boundary vertices of N, given by the Schur-complement
L(N) = K/KI
where KI denotes the submatrix of K indexed by the interior vertices. The response matrix
encodes all the electrical properties of N that can be measured from the boundary vertices.

2.2. Planar electrical networks. We shall usually consider electrical networks N embed-
ded into a disk, so that the boundary vertices, numbered 1, 2, . . . , n + 1, lie on the boundary
of the disk. Given an odd integer 2k − 1, for k = 1, 2, . . . , n + 1, and a nonnegative real

1 1 1

n+1 2 n+1 2 n+1 2

N v4 (t)(N ) t v5 (t)(N )

3 3 1/t
... ... ...
3
4 4 4

Figure 1. The operations vi (t) acting on a network N.

number t, we define v2k−1 (t)(N) to be the electrical network obtained from N by adding a
new edge from a new vertex v to k, with weight 1/t, followed by treating k as an interior
vertex, and the new vertex v as a boundary vertex (now named k).
Given an even integer 2k, for k = 1, 2, . . . , n + 1, and a nonnegative real number t, we
define v2k (t)(N) to be the electrical network obtained from N by adding a new edge from k
to k + 1 (indices taken modulo n + 1), with weight t.
The two operations are shown in Figure 1. In [CIM], these operations are called adjoining
a boundary spike, and adjoining a boundary edge respectively. Our notation suggests, and
we shall establish, that there is some symmetry between these two types of operations.
In [CIM, Section 8], it is shown that L(vi (a)·N) depends only on L(N), giving an operation
vi (t) on response matrices. Denote by xij the entries of the response matrix, where 1 ≤ i, j ≤
n + 1. In particular, we have xij = xji . Then if δij denotes the Kronecker delta, we have
txik xkj
(1) v2k−1 (t) : xij %→ xij − and v2k (t) : xij %→ xij + (δik − δ(i+1)k )(δjk − δ(j+1)k )t.
txkk + 1
We caution that the parameter ξ in [CIM] is the inverse of our t in the odd case.
3
 a
a+b a b ab/(a + b)
b

Figure 2. Series-parallel transformations.

2.3. Electrically-equivalent transformations of networks. Series-parallel transforma-

tions are shown in Figure 2. The following proposition is well-known and can be found for
example in [dVGV].
Proposition 2.1. Series-parallel transformations, removing loops, and removing interior
degree 1 vertices, do not change the response matrix of a network.

a b
B A
c

Figure 3. The Y − ∆, or star-triangle, transformation.

The following theorem is attributed to Kennelly [Ken].

Proposition 2.2 (Y − ∆ transformation). Assume that parameters a,b,c and A,B,C are
related by
bc ac ab
A= , B= , C= ,
a+b+c a+b+c a+b+c
or equivalently by
AB + AC + BC AB + AC + BC AB + AC + BC
a= , b= , c= .
A B C
Then switching a local part of an electrical network between the two options shown in Figure
3 does not change the response matrix of the whole network.

3. The electrical Lie algebra el2n

Let el2n be the Lie algebra el2n generated over R by ei , i = 1, . . . , 2n subject to relations
• [ei , ej ] = 0 for |i − j| > 1;
• [ei , [ei , ej ]] = −2ei , |i − j| = 1.
Theorem 3.1. The Lie algebra el2n is isomorphic to the real symplectic Lie algebra sp2n .
Proof. We identify the symplectic algebra sp2n with the space of 2n × 2n matrices which in
block form # $
A B
C D
4
satisfy A = −D T , B = B T and C = C T (see [FH, Lecture 16]). Note that the subLie algebra
consisting of the matrices where B = C = 0 is naturally isomorphic to gln .
Let εi ∈ Rn denote the standard basis column vector with a 1 in the i-th position. Let
a1 = ε1 , a2 = ε1 + ε2 , a3 = ε2 + ε3 , . . . , an = εn−1 + εn , and let b1 = ε1 , b2 = ε2 , . . . , bn = εn .
For 1 ≤ i ≤ n, define elements of sp2n by the formulae
0 ai · aTi
# $ # $
0 0
e2i−1 = e2i = .
0 0 bi · bTi 0
We claim that this gives a symplectic representation φ of el2n which is an isomorphism
of el2n with sp2n . We first check that the relations of el2n are satisfied. It is clear from the
block matrix form that [e2i−1 , e2j−1 ] = 0 = [e2i , e2j ] for any i, j. Now,
(ai · aTi )(bj · bTj )
# $
0
[e2i−1 , e2j ] = .
0 −(bj · bTj )(ai · aTi )
But by construction, bTj · ai = 0 = aTi · bj unless j = i, or j = i − 1. Thus [ek , el ] = 0 unless
|k − l| = 1. Finally,
# $ # $
0 −2(ai · aTi )(bi · bTi )(ai · aTi ) 0 −2ai · aTi
[e2i−1 , [e2i−1 , e2i ]] = = = −2e2i−1
0 0 0 0
using the fact that aTi · bi = 1 = bTi · ai . Similarly, one obtains [ek , [ek , ek±1]] = −2ek . Thus
we have a symplectic representation φ : el2n → sp2n .
Now we show that this map is surjective. First we verify # that the $ gln ⊂ φ(el2n ). The
A 0
non-zero commutators [ei , ej ] gives matrices of the form where A is a scalar
0 −AT
multiple of one of the matrices E1,1 , or Ei,i + Ei+1,i or Ei+1,i + Ei+1,i+1 . Here Ei,j denotes
the n × n matrix with a single non-zero entry equal to one in the (i, j)-th position. All
the matrices of the above form occur. It is easy to see that φ(el2n ) must then contain the
matrices where A = Ei,i , Ei,i+1 , or Ei+1,i for each i. But these matrices generate gln as a Lie
algebra. However, it is known [Hum, Proposition 8.4(d)] that for a semisimple Lie algebra
L, one has [Lα , Lβ ] = Lα+β when α, β, α + β are all roots, and Lα denotes a root subspace.
It follows easily from the explicit description of the root system of sp2n and the definition of
ei that every root subspace of sp2n is contained in φ(el2n ), completing the proof.
To see that the map φ : el2n → sp2n is injective, we note:
Lemma 3.2. The dimension of el2n is n(2n + 1).
Proof. According to Lemma 5.1 the dimension of el2n is at most n(2n + 1). On the other
hand, we just saw that the map el2n → sp2n is surjective. The statement follows. !

4. The electrical Lie group EL2n

Let EL2n be the simply-connected (real) Lie group with Lie algebra el2n . Let ui (a) =
exp(aei ) denote the the one-parameter subgroups corresponding to the generators of el2n .
Theorem 4.1. The elements ui (a) satisfy the relations
(1) ui (a)ui (b) = ui (a + b),
(2) ui (a)uj (b) = uj (b)ui (a), if |i − j| > 1;
(3) ui (a)uj (b)ui (c) = uj (bc/(a + c + abc))ui (a+c+abc)uj (ab/(a + c + abc)), if |i−j| = 1.
5
Proof. The first two relations are clear. For the third, observe that for each 1 ≤ i ≤ 2i − 1,
the elements ei , ei+1 , and [ei , ei+1 ] are the usual Chevalley generators of a Lie subalgebra
isomorphic to sl2 . Thus we can verify the relation inside SL2 (R):
# $# $# $ # $
1 a 1 0 1 c 1 + ab a + c + abc
= =
0 1 b 1 0 1 b 1 + bc
# $# $# $
1 0 1 a + c + abc 1 0
.
bc/(a + c + abc) 1 0 1 ab/(a + c + abc) 1
!
Remark 4.1. There is a one-parameter deformation which connects the relation Theorem
4.1(3) with Lusztig’s relation [Lus] in total positivity:
(2) ui (a)uj (b)ui (c) = uj (bc/(a + c))ui (a + c)uj (ab/(a + c)).
Consider the associative algebra Uτ (el2n ) where the generators ε1 , ε2 , . . . , ε2n satisfy the fol-
lowing deformed Serre relations:
• εi εj = εj εi if |i − j| > 1;
• εi εj εi = τ εi + (ε2i εj + εj ε2i )/2, if |i − j| = 1.
This is a one-parameter family of algebras which at τ = 0 reduces to U(n+ ), where sl2n+1 =
n− ⊕ h ⊕ n+ is the Cartan decomposition, while at τ = 1 it gives the universal enveloping
algebra U(el 2n ) of the electrical Lie algebra. For Uτ (el2n ), the “braid move” for the elements
exp(aεi ) then takes the form
ui (a)uj (b)ui (c) = uj (bc/(a + c + τ abc))ui (a + c + τ abc)uj (ab/(a + c + τ abc)) if |i − j| = 1.
At τ = 0 this reduces to (2), and at τ = 1 it is the relation Theorem 4.1(3).
4.1. Nonnegative part of EL2n . Define the nonnegative part (EL2n )≥0 of EL2n , or equiv-
alently the electrically nonnegative part of Sp2n , to be the subsemigroup generated by the
ui (a) with nonnegative parameters a.
For a reduced word i = i1 . . . i$ of w ∈ S2n+1 denote by E(w) ⊂ (EL2n )≥0 the image of the
map
ψi : (a1 , . . . , a$ ) ∈ R$>0 %→ ui1 (a1 ) . . . ui! (a$ ).
It follows from the relations in Theorem 4.1 that the set E(w) depends only on w, and not
on the chosen reduced word. The following proposition is similar to [Lus, Proposition 2.7]
w
which gives a decomposition U≥0 = *w U≥0 of the totally nonnegative part of the unipotent
group.
Proposition 4.2. The sets E(w) are disjoint and cover the whole (EL2n )≥0 . Each of the
maps ψi : R$>0 → E(w) is a homeomorphism.
Proof. Using Theorem 4.1, we can rewrite any product of generators ui (a) by performing
braid moves similar to those in the symmetric group S2n+1 . Any product of the generators
can be transformed into a product that corresponds to a reduced word in S2n+1 , and thus it
belongs to one of the E(w).
If the map ψi is not injective for some reduced word i, then we can find two reduced
products
ui1 (a1 ) . . . ui! (a$ ) = ui1 (b1 ) . . . ui! (b$ )
6
for two )-tuples of positive numbers such that a1 #= b1 . Without loss of generality we can
assume a1 > b1 , and thus
ui1 (a1 − b1 )ui2 (a2 ) . . . ui! (a$ ) = ui2 (b2 ) . . . ui! (b$ ).
This shows that two different E(w)-s have non-empty intersection. Thus it suffices to show
that the latter is impossible.
Furthermore, it suffices to prove that the top cell corresponding to the longest element w0
does not intersect any other cell. Indeed, if any two cells intersect, by adding some extra
factors to both we can lift it to the top cell intersecting one of the other cells. Assume we
have a product of the form
(3) u = [u1 (a1 )][u2 (a2 )u1 (a3 )] . . . [uk (a$−k+1) . . . u1 (a$ )]
%k &
where ) = 2 . Let Φ(u) ∈ Sp2n be the image of u in Sp2n under the natural map EL2n →
Sp2n induced by the map φ : el2n → sp2n of Theorem 3.1. We argue that the positive
parameters a$−k+1 , . . . , a$ can be recovered uniquely from just looking at the n + k/2-th row
of Φ(u) for k even, or the (k + 1)/2-th row for k odd. Furthermore, if exactly one of them
is equal to zero, then the same calculation will tell us that.
Indeed, assume k is odd. Then the n + (k + 1)/2-th entry in the (k + 1)/2-th row is just
equal to a$−k+1 . Next, once we know a$−k+1, we can use (k − 1)/2-th entry of the same row
to recover a$−k+2, after which we can use n + (k − 1)/2 entry to recover a$−k+3, and so on.
For each step we solve a linear equation where the parameter we divide by is strictly positive
as long as all previous parameters ai recovered are positive. For example, let n = 2 and take
the product
u1 (a1 )u2(a2 )u1 (a3 )u3(a4 )u2 (a5 )u1 (a6 ).
Then the second row of this product is (a4 a5 , 1, a4 + a4 a5 a6 , a4 ). The last entry tells us a4 ,
then from the first entry we solve for a5 , then from the third entry we solve for a6 . The case
of even k is similar, the order in which we have to read the entries of the n + k/2-th row in
this case is k/2-th, n + k/2-th, k/2 − 1-th, n + k/2 − 1-th, etc. For example, in the product
u1 (a1 )u2 (a2 )u1 (a3 )u3 (a4 )u2 (a5 )u1 (a6 )u4 (a7 )u3 (a8 )u2 (a9 )u1 (a10 ),
the last row is (a7 a8 a9 , a7 , a7 a8 + a7 a8 a9 a10 , 1 + a7 a8 ). By looking at second, last, first and
third entries we can solve for a7 , a8 , a9 and a10 one after the other.
Now we are ready to complete the argument. Assume that E(w0 ) intersects some other
E(v). For any reduced word of w0 , one can find a subword which is a reduced word for v, so
that we have
u = [u1 (a1 )][u2 (a2 )u1 (a3 )] . . . [uk (a$−k+1) . . . u1 (a$ )]
= [u1 (b1 )][u2 (b2 )u1(b3 )] . . . [uk (b$−k+1 ) . . . u1 (b$ )],
where the ai are all positive, but the bi are nonnegative, with at least one zero. The above
algorithm of recovering the ai ’s will arrive at a contradiction. Thus the E(w0 ) is disjoint
from all other E(v).
It remains to show that φ−1i is continuous for any reduced word. If i is the reduced word
of w0 used in (3), this follows from the algorithm above: the ai depend continuously on the
matrix entries of Φ(u). But any two reduced words are connected by braid and commutation
moves, so it follows from the (continuously invertible) formulae in Theorem 4.1 that φ−1 i is
continuous for any reduced word of w0 . But for any other v ∈ S2n+1 , a reduced word j for v
7
can be found as an initial subword of some reduced word i for w0 . The map φ−1 j can then
be expressed as a composition of: right multiplication by a fixed element of EL2n , the map
$(w0 ) $(v)
φ−1
i , and the projection from R>0 to R>0 , all of which are continuous. !
2
4.2. Action on electrical networks. Let P(n + 1) ⊂ R(n+1) denote the set of response
matrices of planar electrical networks with n + 1 boundary vertices. In [CIM], it is shown
that P(n + 1) is exactly the set of symmetric, circular totally-nonnegative (n + 1) × (n + 1)
matrices. In [dVGV, Théorème 4], it is shown that P(n + 1) can be identified with the set
of planar electrical networks with n + 1 boundary vertices modulo the local transformations
of Section 2.3. Let N0 denote the empty network (with n + 1 boundary vertices) and let
L0 = L(N0 ) denote the zero matrix.
Theorem 4.3. The nonnegative part (EL2n )≥0 of the electrical group acts on P(n + 1) via:
ui (a) · L(N) = L(vi (a)(N)).
Proof. For a single generator ui (a), the stated action is well-defined because it can be de-
scribed explicitly on the level of response matrices. The formulae for L(vi (a)(N)) in terms
of L(N) is given by the equation (1).
We first show that the relations of Theorem 4.1 hold for this action. Relation (1) follows
from the series-parallel relation for networks. Relation (2) is immediate: the corresponding
networks are identical without any transformations. Relation (3) follows from the Y − ∆
transformation (see Example 4.4).
But now suppose we have two different expressions for u ∈ (EL2n )≥0 in terms of generators.
Then using Relations (1)-(3) of Theorem 4.1, we may assume that both expressions are
products corresponding to a reduced word. By Proposition 4.2, the two products come
from reduced words for the same w ∈ S2n+1 . It follows that they are related by relations
(1)-(3). !
bc ab
Example 4.4. The products u3 (a)u4 (b)u3 (c) and u4 ( a+c+abc )u3 (a + c + abc)u4 ( a+c+abc ) act
on a network in exactly the same way, as shown in Figure 4.

2 2
1 1 1 1/a
a+c+abc
1/c
n+1 n+1
bc
a+c+abc
b
ab
a+c+abc
... 3 ... 3

4 4

Figure 4.

A permutation w = w(1)w(2) · · · w(2n + 1) ∈ S2n+1 is efficient if

(1) w(1) < w(3) < · · · < w(2n + 1) and w(2) < w(4) < · · · < w(2n)
(2) w(1) < w(2), w(3) < w(4), . . ., w(2n − 1) < w(2n).
8
Recall the left weak order of permutations is given by w + v if and only if there is a u
so that uw = v and )(v) = )(u) + )(w). It is a standard fact that w + v if and only if
whenever w(i) > w(j) and i < j then v(i) > v(j). Thus the set of efficient permutations
has a maximum in left weak order, namely w = 1(n + 2)2(n + 3) · · · (n)(2n + 1)(n + 1) with
length n(n − 1)/2.
Theorem 4.5. Let w ∈ S2n+1 . The map Θw : E(w) → P(n + 1) given by Θw (u) = u · L0 is
injective if and only if w is efficient. We have image(Θw ) ∩ image(Θv ) = ∅ for w #= v both
efficient. If w is not efficient, there is a unique efficient v such that image(Θw ) = image(Θv ).
Proof. Let w ∗ ∈ S2n+1 be the efficient permutation of maximal length. Then a possible
reduced word for w ∗ is (n + 1) · · · (46 · · · (2n − 2)) (35 · · · (2n − 1)) (246 · · · (2n)). The graph
obtained by taking the corresponding ui (a) and acting on the empty network is exactly the
“standard graph” of [CIM, Section 7] or the graph CN of [dVGV]. In particular, Θw∗ is
injective by [CIM, Theorem 2] or [dVGV, Théorème 3]. Suppose w is an arbitrary efficient
permutation. Then since w ∗ = uw for some u ∈ S2n+1 , if Θw is not injective then Θw∗ is not
injective as well, so we conclude that Θw is injective.
For a pair (i, j) with 1 ≤ i < j ≤ n + 1, let us say that a network N is (i, j)-connected
if we can find a disjoint set of paths p1 , p2 , ..., p%(j−i+1)/2& so that for each k, the path pk
connects boundary vertex i + k − 1 to boundary vertex j − k + 1 without passing through
any other boundary vertex. This is a special case of the connections of circular pairs (P, Q)
of [CIM]. Let Nw be a graph constructed from some reduced word of an efficient w. We
observe that Nw is (i, j)-connected if and only if w(2i) > w(2j − 1). For example, the first

1 1

4 2 4 2

3 3

Figure 5.
network in Figure 5 corresponds to w = s5 s3 s6 s4 s2 = (1, 3, 5, 2, 7, 4, 6). We see that it is
(2, 4)-connected, which agrees with w(4) = 6 > 5 = w(7). If the dashed edge is not there,
we have w = s3 s6 s4 s2 = (1, 3, 5, 2, 4, 7, 6) and w(4) = 5 < 6 = w(7) in agreement with the
network not being (2, 4)-connected. Similarly, the second network in Figure 5 corresponds
to w = s4 s5 s3 s6 s4 s2 = (1, 3, 5, 7, 2, 4, 6) with the dashed edge and to w = s5 s3 s6 s4 s2 =
(1, 3, 5, 2, 7, 4, 6) without. In the first case it is (1, 4)-connected, in the second it is not, in
agreement with relative order of w(2) and w(7).
Note that the inversions w(2i) > w(2j − 1) are exactly the possible inversions of an
efficient permutation. Since w is determined by its inversions, it follows that the set of
(i, j)-connections of Nw determines w, and that image(Θw ) ∩ image(Θv ) = ∅ for w #= v both
efficient.
9
 Suppose w is not efficient. Then it is easy to see that w has a reduced expression si1 si2 · · · si!
where either (1) i$ is odd, or (2) i$ is even and i$ = i$−1 ± 1. But ui (a) · L0 = L0 for odd
i since all the boundary vertices are still disconnected in ui (a) · N0 , and ui±1 (a)ui (b) · L0 =
ui (1/a + b)L0 , using the series-transformation (Proposition 2.1). It follows that Θw is not
injective. Furthermore, image(Θw ) = image(Θv ), where v is obtained from w by recursively
(1) removing i$ from a reduced word of w if i$ is odd, or (2) changing the last two letters
(i$ ± 1)i$ to i$ when i$ is even. An efficient v obtained in this way must be unique, since
image(Θv ) ∩ image(Θv" ) = ∅ for efficient v #= v ! . !

%2n+2& 4.6. The number of efficient w ∈ S2n+1 is equal to the Catalan number Cn+1 =
Corollary
1
n+2 n+1
.
Proof. It is clear that (2i, 2j + 1) is an inversion of w only if (2i, 2j −1) and (2i+ 2, 2j + 1) are
also inversions, and this characterizes inversion sets of efficient w. Thus the set of efficient
w ∈ S2n+1 is in bijection with the lower order ideals of the positive root poset of sln+1 which
is well known to be enumerated by the Catalan number [FR]. !
Corollary 4.7. (EL2n )≥0 · L0 is dense in P(n + 1).
Proof. Follows from Theorem 4.5 and [dVGV, Théorème 5]. !

4.3. Stabilizer.
Lemma 4.8. The stabilizer subsemigroup of (EL2n )≥0 acting on the zero matrix L0 is the
subsemigroup generated by u2i+1 (a).
Proof. It is clear that u2i+1 (a) lies in the stabilizer. But the action of any u2i (a) will change
the connectivity of the network, and it is impossible to return to trivial connectivity by
adding more edges, or by relations. !

The semigroup stabilizer is too small in the sense that it does not detect the relations
ui±1 (a)ui (b) · L0 = ui (1/a + b)L0 used in the proof of Theorem 4.5. We shall calculate the
stabilizer subalgebra of the corresponding infinitesimal action of the Lie algebra el2n , which
will in particular give an algebraic explanation of these relations. The reason we do not work
with the whole Lie group EL2n is threefold: (1) non-positive elements of EL2n will produce
networks that are “virtual”, that is, have negative edge weights; (2) the topology of EL2n
means that to obtain an action one cannot just check the relations of Theorem 4.1; (3) when
the parameters are non-positive, the relation Theorem 4.1(3) develops singularities.
To describe the infinitesimal action of el2n , we give derivations of R[xij ], the polynomial
ring in (n + 1)(n + 2)/2 variables xij where 1 ≤ i, j ≤ n + 1 and we set xij = xji .
Proposition 4.9. The electrical Lie algebra el2n acts on R[xij ] via derivations as follows:
(4) e2i %→ ∂ii + ∂i+1,i+1 − ∂i,i+1
'
(5) e2i−1 %→ − xip xiq ∂pq
1≤p≤q≤n+1

Proof. These formulae can be checked by directly verifying the defining relations of el2n .
Alternatively, they can be deduced by differentiating the formulae (1). !
10
 We calculate that
'
(6) [e2i , e2i−1 ] %→ −xii ∂ii + xi,i+1 ∂i+1,i+1 + (xi+1,p ∂i+1,p − xip ∂ip ) .
1≤p≤n+1

Theorem 4.10. The stabilizer subalgebra el02n , at the zero matrix L0 , of the infinitesimal
action of el2n on the space of response matrices is generated by ei for i odd, and [e2i−1 , e2i ]
for i = 1, 2, . . . , n.
Proof. The fact the stated elements lie in el02n follows from (5) and (6), since xij = 0 at the
zero matrix. By Lemma 3.2, the elements eα in the proof of Lemma 5.1 form a basis of el2n .
Write αij := αi + αi+1 + · · · + αj for 1 ≤ i ≤ j ≤ 2n to denote the positive roots of sl2n+1 .
Then we know that eαi,i ∈ el02n for i odd, and eαi,i+1 ∈ el02n for each i. It follows easily that
eαi,j ∈ el02n for every pair 1 ≤ i ≤ j ≤ 2n where at least one of i and j are odd. It follows
that
dimR (el2n ) − dimR (el02n ) ≤ #{(i, j) | 1 ≤ i ≤ j ≤ 2n and i and j are even} = n(n + 1)/2.
By Theorem 4.5 the action of (EL2n )≥0 on the zero matrix L0 gives a space of dimension
n(n + 1)/2. It follows that the above inequality is an equality, and that el02n is generated by
the stated elements. !
Note that a basis for el2n /el02n is given by the eαi,j ’s where i and j are both even. These
αi,j ’s are exactly the inversions of efficient permutations (cf. proof of Theorem 4.5).

5. Electrical Lie algebras of finite type

5.1. Dimension. Let D be a Dynkin diagram of finite type and let A = (aij ) be the asso-
ciated Cartan matrix. To each node i of D associate a generator ei , and define eD to be the
Lie algebra generated over R by the ei modulo for each i #= j the relations ad(ei )1−aij (ej ) = 0
for aij #= −1 and ad(ei )2 (ej ) = −2ei for aij = −1.
These “electrical Serre relations” can be deduced from the type A electrical Serre relations
of Section 3 by folding. Namely, the relation for an edge of multiplicity two (aij = −2) can
be obtained by finding the relation for the elements e2 and e1 + e3 inside el3 . Similarly, the
eG2 relation (aij = −3) can be obtained by finding the relation for the elements e1 + e2 + e3
and e4 inside eD4 , where e4 corresponds to the node of valency three in D4 .
Lemma 5.1. The dimension of the Lie algebra eD does not exceed the number of positive
roots in the finite root system associated to D.
Proof. Consider the nilpotent subalgebra uD of the usual Lie algebra associated to D. It
is well known that it is generated by Chevalley generators ẽi subject to the Serre relations:
ad(ẽi )1−aij (ẽj ) = 0 for each i #= j.
The Lie algebra uD has a basis ẽα labeled by positive roots α ∈ R+ of the root system,
where the ẽi correspond to simple roots. Let us fix an expression
ẽα = [ẽi1 , [ẽi2 , [. . . , [ẽil−1 , ẽil ] . . .]]]
of shortest possible length for each ẽα , and define the corresponding
(7) eα = [ei1 , [ei2 , [. . . , [eil−1 , eil ] . . .]]]
11
in eD . We claim that the eα span eD . Indeed, it is enough to show that any expression
[ej1 , [ej2 , [. . . , [ej!−1 , ej! ] . . .]]] lies in the linear span of the eα -s. Assume otherwise, and take
a counterexample of smallest possible total length ). Let us define êα in the free Lie algebra
fD with generators êi using (7). Then in fD we have a relation of the form
'
[êj1 , [êj2 , [. . . , [êj!−1 , êj! ] . . .]]] − cα êα = x̂ ∈ I
α

where I denotes the ideal generated by the Serre relations, so that uD = fD /I. Now fD is
naturally Z-graded, and I is a graded ideal, so we may assume all terms in the relation are
homogeneous with the same degree. Now replacing each instance of the Serre relation in x
with the corresponding electrical Serre relation gives us a relation
'
[ej1 , [ej2 , [. . . , [ej!−1 , ej! ] . . .]]] − cα eα = x
α

in eD , where x is a sum of terms of the form [ek1 , [ek2 , [. . . , [ek!−1 , ek!" ] . . .]]] where )! < ), and
thus by assumption lies in the span of the eα -s. The statement of the lemma follows. !
Conjecture 5.2. The dimension of eD coincides with the number of positive roots in the
root system of D.
One can also define the Lie algebras eD for any Dynkin diagram D, finite type or not.
Conjecture 5.3. The Lie algebra eD is finite dimensional if and only if D is of finite type.
5.2. Examples. We illustrate Conjecture 5.2 with some examples. For electrical type A2n
it has already been verified in Theorem 3.1.

5.2.1. Electrical B2 . Consider the case when D has two nodes connected by a double edge.
In that case we denote by eD = eb2 the Lie algebra generated by two generators e and f
subject to the relations
[e, [e, [e, f ]]] = 0, [f, [f, e]] = −2f.
Lemma 5.4. The Lie algebras eb2 and gl2 are isomorphic.
Proof. Consider the map
# $ # $
1 1 0 0
e %→ , f→
% .
0 1 1 0
One easily checks that it is a Lie algebra homomorphism, and that it is surjective. By Lemma
5.1 we know the dimension of eb2 is at most four, so the map must be an isomorphism. !

5.2.2. Electrical G2 . Consider the Lie algebra eg2 generated by two generators e and f sub-
ject to the relations
[e, [e, [e, [e, f ]]]] = 0, [f, [f, e]] = −2f.
Lemma 5.5. The Lie algebra eg2 is six-dimensional.
12
Proof. According to Lemma 5.1 the elements e, f, [ef ], [e[ef ]], [e[e[ef ]]], [f [e[e[ef ]]]] form a
spanning set for eg2 . Thus it remains to check that they are linearly independent. This is
easily done inside the following faithful representation of eg2 in gl4 :
   
1 1 0 1 0 0 0 0
0 1 1 0 0 0 0 0
0 0 1 0 , f %→ 0 0 0 0 .
e %→    
0 0 0 1 1 0 0 0
!
5.2.3. Electrical C3 . Consider the Lie algebra ec3 generated by three generators e, f and g
subject to the relations
[e, [e, [e, f ]]] = 0, [f, [f, e]] = −2f, [f, [f, g]] = −2f, [g, [g, f ]] = −2g, [e, g] = 0.
Lemma 5.6. The Lie algebra ec3 is nine-dimensional.
Proof. We consider two representations of ec3 : one inside sl9 and one inside sl2 . Let Eij de-
note a matrix with a 1 in the (i, j)-th position and 0’s elsewhere. For the first representation,
we define:
e %→ E42 + E54 + E76 + E87 + E89
f %→ 2E24 − 2E26 + 2E41 + 2E45 − E47 − 2E63 + E67 + E98
g %→ −E36 − E62 − E74 − E85 + E89
For the second representation we define
# $ # $ # $
1 1 0 0 0 1
e %→ , f %→ , g %→ .
0 1 1 0 0 0
One verifies directly that these are indeed representations of ec3 , and the direct sum of
these two representations is faithful, from which the dimension is easily calculated. In fact,
the first representation is simply the adjoint representation of ec3 , in the basis
e, f, g, [ef ], [e[ef ]], [f g], [e[f g]], [e[e[f g]]], [f [e[e[f g]]]],
and ec3 has a one-dimensional center which acts non-trivially in the second representation.
!
5.3. Y − ∆ transformation of type B. Let e and f be the generators of eb2 as before, and
denote by u(t) = exp(te) and v(t) = exp(tf ) the corresponding one-parameter subgroups.
The following proposition is a type B analog of the star-triangle transformation (Proposition
2.2).
Proposition 5.7. We have
u(t1 )v(t2 )u(t3 )v(t4 ) = v(p1 )u(p2)v(p3 )u(p4 ),
where
t2 t23 t4 π2 π2 t1 t2 t3
p1 = , p2 = , p3 = 1 , p4 = ,
π2 π1 π2 π1
where
π1 = t1 t2 + (t1 + t3 )t4 + t1 t2 t3 t4 , π2 = t21 t2 + (t1 + t3 )2 t4 + t1 t2 t3 t4 (t1 + t3 ).
13
Proof. Direct calculation inside GL2 , using Lemma 5.4. We have
# t t$ # $
e et 1 0
u(t) = , v(t) = ,
0 et t 1
and both sides of the equality are equal to
# t +t $
e 1 3 (1 + t3 t4 + t1 (t2 + t4 + t2 t3 t4 )) et1 +t3 (t1 + t3 + t1 t2 t3 )
.
et1 +t3 (t2 + t4 + t2 t3 t4 ) et1 +t3 (1 + t2 t3 )
!
Remark 5.1. One can consider a one-parameter family of deformations of the above formulas
by taking
π1 = t1 t2 + (t1 + t3 )t4 + τ t1 t2 t3 t4 , π2 = t21 t2 + (t1 + t3 )2 t4 + τ t1 t2 t3 t4 (t1 + t3 ).
At τ = 0 this specializes to the transformation found by Berenstein and Zelevinsky in [BZ,
Theorem 3.1]. Furthermore, just like the transformation in [BZ], this deformation is given
by positive rational formulae.
References
[BZ] A. Berenstein and A. Zelevinsky: Total positivity in Schubert varieties, Comment. Math. Helv.,
72 (1997), no. 1, 128–166.
[CIM] E.B. Curtis, D. Ingerman, and J.A. Morrow: Circular planar graphs and resistor networks,
Linear Algebra Appl., 283 (1998), no. 1-3, 115–150.
[dVGV] Y.C. de Verdière, I. Gitler, and D. Vertigan: Réseaux électriques planaires. II, Comment.
Math. Helv., 71(1) (1996), 144–167
[FH] W. Fulton and J. Harris: Representation theory. A first course., Graduate Texts in Mathematics,
129, Springer-Verlag, New York, 1991. xvi+551 pp.
[Hum] J.E. Humphreys: Reflection groups and Coxeter groups, Cambridge Studies in Advanced Mathe-
matics, 29, Cambridge University Press, Cambridge, 1990. xii+204 pp.
[FR] S. Fomin and N. Reading: Root systems and generalized associahedra, Geometric combinatorics,
63–131, IAS/Park City Math. Ser., 13.
[Ken] A.E. Kennelly: Equivalence of triangles and stars in conducting networks, Electrical World and
Engineer, 34 (1899), 413–414.
[Lus] G. Lusztig: Total positivity in reductive groups, Lie theory and geometry, 531–568, Progr. Math.,
123, Birkhäuser Boston, Boston, MA, 1994.
E-mail address: [email protected]

Department of Mathematics, University of Michigan, 530 Church St., Ann Arbor, MI

48109 USA

E-mail address: [email protected]

Department of Mathematics, University of Minnesota, 206 Church St. SE, Minneapolis,

MN 55455 USA

14