Refined open intersection numbers and the Kontsevich-Penner matrix model

A study of the intersection theory on the moduli space of Riemann surfaces with boundary was recently initiated in a work of R. Pandharipande, J.P. Solomon and the third author, where they introduced open intersection numbers in genus 0. Their construction was later generalized to all genera by J.P. Solomon and the third author. In this paper we consider a refinement of the open intersection numbers by distinguishing contributions from surfaces with different numbers of boundary components, and we calculate all these numbers. We then construct a matrix model for the generating series of the refined open intersection numbers and conjecture that it is equivalent to the Kontsevich-Penner matrix model. An evidence for the conjecture is presented. Another refinement of the open intersection numbers, which describes the distribution of the boundary marked points on the boundary components, is also discussed.


Introduction
A compact Riemann surface is a compact connected complex manifold of dimension 1. Denote by M g,l the moduli space of all compact Riemann surfaces of genus g with l marked points. P. Deligne and D. Mumford defined a natural compactification M g,l ⊂ M g,l via stable curves (with possible nodal singularities) in [9]. The moduli space M g,l is a nonsingular complex orbifold of dimension 3g − 3 + l. It is defined to be empty unless the stability condition 2g − 2 + l > 0 (1.1) is satisfied. We refer the reader to [9,10] for the basic theory.
In his seminal paper [23], E. Witten initiated new directions in the study of M g,l . For each marking index i consider the cotangent line bundle L i → M g,l , whose fiber over a point

JHEP03(2017)123
[Σ, z 1 , . . . , z l ] ∈ M g,l is the complex cotangent space T * z i Σ of Σ at z i . Let ψ i ∈ H 2 (M g,l , Q) denote the first Chern class of L i , and write τ a 1 τ a 2 · · · τ a l c g := M g,l ψ a 1 1 ψ a 2 2 · · · ψ a l l . (1. 2) The integral on the right-hand side of (1.2) is well-defined, when the stability condition (1.1) is satisfied, all the a i are non-negative integers and the dimension constraint 3g−3+l = a i holds. In all other cases τ a i c g is defined to be zero. The intersection products (1.2) are often called descendent integrals or intersection numbers. Let t i , i ≥ 0, be formal variables and let τ a 1 τ a 2 · · · τ a l c g t a i l! .
The generating series F c is called the closed free energy. The exponent τ c := exp(F c ) is called the closed partition function. Witten's conjecture ( [23]), proved by M. Kontsevich ( [13]), says that the closed partition function τ c becomes a tau-function of the KdV hierarchy after the change of variables t n = (2n + 1)!!T 2n+1 . Integrability immediately follows [11] from Kontsevich's matrix integral representation In [15] the authors started to develop a parallel theory for Riemann surfaces with boundary. A Riemann surface with boundary is a connected 1 dimensional complex manifold with finite positive number of circular boundaries, each with a holomorphic collar structure. A compact Riemann surface is not viewed here as a Riemann surface with boundary. Given a Riemann surface with boundary (X, ∂X), we can canonically construct a double via Schwarz reflection through the boundary. The double D(X, ∂X) of (X, ∂X) is a compact Riemann surface. The doubled genus of (X, ∂X) is defined to be the usual genus of D(X, ∂X). On a Riemann surface with boundary (X, ∂X), we consider two types of marked points. The markings of interior type are points of X\∂X. The markings of boundary type are points of ∂X. Let M R g,k,l denote the moduli space of Riemann surfaces with boundary of doubled genus g with k distinct boundary markings and l distinct interior markings. The moduli space M R g,k,l is defined to be empty unless the stability condition 2g − 2 + k + 2l > 0

JHEP03(2017)123
is satisfied. The moduli space M R g,k,l may have several connected components depending upon the topology of (X, ∂X) and the cyclic orderings of the boundary markings. Foundational issues concerning the construction of M R g,k,l are addressed in [14]. The moduli space M R g,k,l is a real orbifold of real dimension 3g − 3 + k + 2l, it is in general not compact and may be not orientable when g > 0.
Since interior marked points have well-defined cotangent spaces, there is no difficulty in defining the cotangent line bundles L i → M R g,k,l for each interior marking, i = 1, . . . , l. Naively, one may want to consider a descendent theory via integration of products of the first Chern classes ψ i = c 1 (L i ) ∈ H 2 (M R g,k,l , Q) over a compactification M R g,k,l of M R g,k,l . Namely, τ a 1 τ a 2 · · · τ a l σ k o g := 2 − g+k−1 2 M R g,k,l ψ a 1 1 ψ a 2 2 · · · ψ a l l , (1.4) and in all other cases τ a 1 τ a 2 · · · τ a l σ k o g := 0. Note that, in particular, g +k must always be odd in order to get non-zero numbers. The new insertion σ corresponds to the addition of a boundary marking. The coefficient in front of the integral on the right-hand side of (1.4) appears to be useful for the description of the new intersection numbers, that are called the open intersection numbers, in terms of integrable systems.
In genus 0 the moduli M 0,k,l := M R 0,k,l is canonically oriented for k odd, and one can calculate an integral of the form M 0,k,l ψ a 1 1 ψ a 2 2 · · · ψ a l l , given boundary conditions for the line bundles L i . More precisely, given nowhere vanishing boundary conditions s ∈ C ∞ (E → ∂M 0,k,l ), for E = L ⊕a i i , one may define the integral (1.4) by τ a 1 τ a 2 · · · τ a l σ k o where e(E, s) is the relative Euler class. The result depends on the boundary conditions. In [15] a family of boundary conditions, called canonical boundary conditions for each bundle L i is constructed. It is proven that for a generic choice of canonical boundary conditions, s ij ∈ C ∞ m (L i → ∂M 0,k,l ), i ∈ [l], j ∈ [a i ], the boundary conditions s = s ij is nowhere vanishing along ∂M 0,k,l , assuming 2 a i = 3g − 3 + k + 2l. Here we use the notation [l] for a set {1, 2, . . . , l} and the subscript m indicates that multi-valued section, rather than sections, are used. It is then shown that any two generic choices of canonical boundary conditions give rise to the same integral (1.5). In [15] all open intersection numbers for doubled genus 0 were calculated, and the authors proposed a conjectural description of the open intersection numbers in all genera. Let s be a formal variable. Define

JHEP03(2017)123
The generating series F o is called the open free energy and the exponent is called the open partition function. The conjecture of R. Pandharipande, J. P. Solomon and the third author ( [15]) says that the generating series F o satisfies a certain system of partial differential equations that is called in [15] the open KdV equations. In higher genus the construction of open intersection numbers needs some refinement. Firstly, the moduli space M g,k,l is in general non-orientable for g > 0. In order to overcome this issue, J. P. Solomon and the third author define graded spin surfaces, which are open surfaces with a spin structure and some extra structure. In [18] the moduli of graded spin surfaces M g,k,l is defined and is proved to be canonically oriented. When g = 0 it coincides with M R 0,k,l . Canonical boundary conditions are then constructed for the line bundles L i , and again it is proven that one can define where e(E, s) is the relative Euler with respect to the canonical boundary conditions. As in g = 0, generic choices of canonical boundary conditions give rise to the same integrals. It should be stressed that, although [18] has not appeared yet, the moduli and boundary conditions mentioned above are fully described in section 2 of [21]. A combinatorial formula for the open intersection numbers was found in [21]. The conjecture of R. Pandharipande, J. P. Solomon and the third author was proved in [6].
Properties of the open free energy F o were intensively studied in [2-4, 7, 8, 16]. In particular, in [7,8] the second author introduced a formal power series  [7,8] the new variables s i , i ≥ 1, appeared naturally from the point of view of integrable systems. The second author suggested to consider them as descendents of the boundary marked points. A geometric construction of the descendent theory for the boundary marked points, a derivation of the combinatorial formula for it, and a geometric proof of the conjecture of [7] regarding the extended theory, will appear in [19,22]. In [8] the second author found a simple relation of the extended open partition function τ o,ext to the wave function of the Kontsevich-Witten tau-function. In [3] the first author proved that both extended open partition function and closed partition function belong to the same family of tau-functions, described by the matrix integrals of Kontsevich type. Namely, the Kontsevich-Penner integral for N = 0 coincides with Kontsevich's integral (1.3). In [3] it was shown that for N = 1 it describes the extended open partition function. From this matrix integral representation JHEP03(2017)123 In [18] the authors defined open intersection numbers over any connected component of the moduli space M g,k,l . To be precise, they proved the following result. Theorem 1.1. Let a 1 , . . . , a l , k be non-negative integers satisfying 2 a i = 3g − 3 + k + 2l, and let E = k i=1 L ⊕a i i . Then for any connected component C of M g,k,l there exist nowhere vanishing canonical boundary conditions s in the sense of [15,18]. Thus one may define the integral C e(E, s). Moreover, any two nowhere vanishing choices of the canonical boundary conditions give rise to the same integral.
The theorem allows us to define refined open intersection numbers as the integrals of monomials in psi-classes over the components M g,k,l,b of M g,k,l and very refined open JHEP03(2017)123 intersection numbers as the corresponding integrals over the components M g,k,l : where a 1 , . . . , a l , k, E are as in Theorem 1.1 and s is a nowhere vanishing canonical multisection. These new intersection numbers are rational numbers. Let N be a positive integer.
Introduce the refined open free energy F o,N by Of course, the function F o,N can be easily expressed in terms of the function F o : In this paper we generalize the result of the third author from [21] and find a combinatorial formula for the very refined open intersection numbers. We also derive matrix models for the refined and the very refined open partition functions. We then show that the form of our matrix model for the refined open partition function τ o N suggests a natural way to add the variables s i , i ≥ 1, in it. We denote the extended function by τ o,ext

JHEP03(2017)123
Remark 1.2. In [16] the author conjectured that there exists a refinement of the extended open partition function τ o,ext that distinguishes contributions from Riemann surfaces with different numbers of boundary components and that coincides with the Kontsevich-Penner tau-function τ N . Since we construct this refinement, our conjecture can be considered as a stronger version of the conjecture of B. Safnuk from [16]. Remark 1.3. Another approach to refined open intersection numbers was recently suggested by B. Safnuk in [17]. His approach is quite different to ours, because, in particular, he does not consider boundary marked points and, moreover, he uses a different compactification of M g,0,l . His intersection numbers are given as integrals of some specific volume forms. B. Safnuk also has a combinatorial formula for his refined open intersection numbers and it directly gives the Kontsevich-Penner matrix model. It would be interesting to obtain a direct relation between the two approaches.

Organization of the paper
In section 2 we show that the construction of [18] admits a refinement that allows to define the products (1.8) and (1.9). We also prove combinatorial formulas for the refined and 2 Very refined open intersection numbers 2.1 Reviewing the proof of the combinatorial formula of [21] In order to prove a combinatorial formula for the refined open intersection numbers, we first review the proof technique in the rather long paper [21]. Throughout this subsection we shall address to places in [21].
Step 1. The starting point of [21] is the following well known fact. Let M be an orbifold with boundary or even corners, of real dimension 2n. Suppose E → M is a vector bundle of real rank 2n, and s a nowhere vanishing (possibly multi-valued) section of E → ∂M. Let π : S → M be the sphere bundle associated to E, Φ an angular form and Ω an Euler form on M. In other words, Φ is a 2n − 1 form on the total space S with

JHEP03(2017)123
Step 2. In [21], section 4 using the theory of Jenkins-Strebel differential [20], with the required modifications for graded surfaces with boundary, a combinatorial stratification of M g,k,l is constructed. The stratification, given a choice of positive perimeters p = {p 1 , . . . , p l }, consists of cells parameterized by metric graded ribbon graphs (G, z). These are ribbon graphs with a (positive) metric on edges, l+b holes, where the last b holes, called boundaries correspond to boundary components, the i th hole for 1 ≤ i ≤ l is called a face and is of perimeter p i , and there are k boundary vertices which correspond to boundary marked points. z is an index for the graded structure, whose description is not important at the moment. The topology of the cells is defined in the natural way using the metric. A cell M (G ′ ,z ′ ) is a face of a cell M (G,z) if G ′ is obtained from G by contracting some edges and z ′ is the degenerated graded structure. The edge contraction operation allows a compactification of the combinatorial moduli, which is a quotient of M g,k,l , generically 1 : 1. Denote this compactification by M comb g,k,l (p). Write also M comb g,k,l = p 1 ,...,p l >0 M comb g,k,l (p), and endow it with the natural topology and piecewise linear structure obtained by the graphs description. For later uses, write Not only the moduli, but also the S 1 bundles associated to the line bundles L i have a combinatorial counterpart, first obtained in [13]. Using these, in [21], subsection 4.3, a combinatorial S 2n−1 bundle S = S(E) is constructed for any vector bundle E = L a i i , where n = a i . It is then shown, in Proposition 4.39, that canonical multisections used to calculate the open intersection numbers can be taken to be pull backs of canonical multisections over M comb g,k,l . Call multisections of S whose pull back is canonical combinatorial canonical. where s is a canonical multisection which is a pull back of the combinatorial canonical multisection s ′ .
Step 3. In [13] a combinatorial angular form α i and a combinatorial curvature form ω i were constructed, and using them a combinatorial formula for the closed numbers was obtained, by integration over highest dimensional cells, those parameterized by trivalent ribbon graph. The main result of [21], section 3 is an explicit formula for the angular form Φ of a bundle which is a direct sum of complex line bundles L i , in terms of their angular forms α i and curvature forms ω i , such that dΦ is the pull back of − ∧ ω i . Plugging this and (2.1) in Lemma 2.1 we get where Φ is the explicit angular form for L ⊕a i i . Finally, this equation can be simplified by noting that only highest dimensional cells of the combinatorial moduli and its boundary contribute to the integrals. The highest JHEP03(2017)123 dimensional cells of M comb g,k,l are those parameterized by trivalent graded ribbon graphs. Denote their set by SR 0 g,k,l . For any such graph, (G, z) write Br(G) for the set of bridges, that is, edges which are either internal edges between two boundary vertices or boundary edges between boundary marked points. The highest dimensional cells in ∂M comb g,k,l are exactly those obtained from contracting a bridge in a cell of SR 0 g,k,l . Putting all together we obtain ( [21],Lemma 4.45) where s ′ is combinatorial canonical.
Step 4. The expression (2.2) has a complicated part, the integral of (s ′ ) * Φ, since it involves the multisection s ′ . However, it turns out that the properties of canonical sections allow computing the right-hand side of (2.2) using iterative integrations by parts. The result is the integral version of the combinatorial formula. To this end, one must first have an explicit description of the contributing graded ribbon graphs.
Definition 2.2. Let g, k, l be non-negative integers such that 2g − 2 + k + 2l > 0, A be a finite set and α : [l] → A a map. α, A will be implicit in the definition. A (g, k, l)ribbon graph with boundary is an embedding ι : G → Σ of a connected graph G into a (g, k, l)-surface with boundary Σ such that • The degree of any vertex v ∈ V (G) \ {x i } is at least 3.
where each D i is a topological open disk, with z i ∈ D i . We call the disks D i faces.
The genus g(G) of the graph G is the genus of Σ. The number of the boundary components of G or Σ is denoted by b(G) and v I (G) stands for the number of the internal vertices. We denote by Faces(G) the set of faces of the graph G, and we consider α as a map Two ribbon graphs with boundary ι : , and an isomorphism of graphs φ : G → G ′ , such that JHEP03(2017)123 where α, α ′ are the labelings of G, G ′ respectively and f ∈ Faces(G) is any face of the graph G.
Note that in this definition we do not require the map Φ to preserve the numbering of the internal marked points. A ribbon graph is critical, if • Boundary marked points have degree 2.
• All other vertices have degree 3.
A (0, 3, 0)−ribbon graph with boundary is called a ghost. Consider maps K from the set of directed edges of G to Z 2 which satisfy • K(e) + K(ē) = 1, whereē is e with opposite orientation.
• For any face f i of the graph G we have K(e) = 1, where the sum is taken over the directed edges of f i , whose direction agree with the orientation of f i .
• Any directed edge of a boundary component has K = 0.
A grading of a critical ribbon graph is the equivalence class of such maps modulo the relations obtained by vertex flips. That is, K, K ′ are identified if they differ by a sequence of moves which flip all the edge assignments for the edges which touch a vertex v. Write [K] for the equivalence class of K. A graph together with a grading is called a graded graph.
A metric graded graph is a graded graph (G, [K]) together with a metric ℓ : Edges(G) → R + . Let M (G,[K]) be the moduli of such metrics.
From now on the explicit object [K] will replace the abstract index z used so far. In figure 1 two critical ribbon graphs are shown, the right one is a ghost. We draw internal edges as thick (ribbon) lines, while boundary edges are usual lines. Note that not all boundary vertices are boundary marked points. We draw parallel lines inside the ghost, to emphasize that the face bounded by the boundary is a special face, without a marked point inside.
• G is a connected graph, • Elements of N are disjoint as sets (without ordering).
After the identification of the vertices v 1 and v 2 the corresponding point in the graph is called a node. The vertex v 1 is called the legal side of the node and the vertex v 2 is called the illegal side of the node.
The set of edges Edges(G) is composed of the internal edges of the G i 's and of the boundary edges. The boundary edges are the boundary segments between successive vertices which are not the illegal sides of nodes. For any boundary edge e we denote by m(e) the number of the illegal sides of nodes lying on it. The boundary marked points of G are the boundary marked points of G i 's, which are not nodes. The set of boundary marked points of G will be denoted by V BM (G) also in the nodal case.
• All of its components G i are critical.
• Ghost components do not contain the illegal sides of nodes.
It is called odd critical if it is critical and any boundary component of G i has an odd number of points that are the boundary marked points or the legal sides of nodes. A graded (odd) critical nodal graph (G, [K]) is a critical (odd) ribbon graph with gradings associated to each component G i .
A nodal ribbon graph with boundary is naturally embedded into the nodal surface Σ = ( i Σ i ) /N . The (doubled) genus of Σ is called the genus of the graph. The notions of an isomorphism and metric are also as in the non-nodal case. Write M (G,[K]) for the moduli of metrics on (G, [K]).
Remark 2.4. The genus of a closed, and in particular doubled, nodal surface Σ is the genus of the smooth surface obtained by smoothing all nodes of Σ.
In figure 2 there is a critical nodal graph of genus 0, with 5 boundary marked points, 6 internal marked points, three components, one of them is a ghost, two nodes, where a plus sign is drawn next to the legal side of a node and a minus sign is drawn next to the illegal side.   ), which has one node N more than G has, can also be made critical nodal by declaring the side of N which corresponds to e to be legal, if K(e) = 0, and otherwise declare the other side of N to be legal. Denote the resulting graph by B∂ e (G, [K]). The operation B is called the base operation.

JHEP03(2017)123
Definition 2.8. Any l−set defines a vector bundle E L := L L(i) , defined both on the moduli and on the combinatorial moduli. Let S L be the associated combinatorial sphere bundle. Let Φ L be the associated explicit angular form, mentioned in Step 3 above, and defined in [21], section 3. Its curvature form is ω L = i∈[n] ω L(i) .
Let C ⊆ SR m g,k,l be a set of graphs and let C ′ ⊆ SR m+1 g,k,l be the set of graphs obtained by applying for any graph in C and any effective bridge e of it, first the edge contraction ∂ e and then the base operation B. Suppose C is closed in the following sense: for any graph (G, [K]) in SR m g,k,l \ C and any effective bridge e of it we have B(∂ e G) / ∈ C ′ . Then This lemma is the global version of the combination of Lemmas 6.7 and 6.8 of [21] (there a local version is given, in terms of a single graph, rather than a set C, and in terms of a single l−subset of it, rather than summing over all subsets).
Applying Lemma 2.9 iteratively to C = SR m g,k,l , and using some parity observation (Proposition 6.13 in [21]) give the integrated form of the combinatorial formula, [21], Theorem 6.12. , let L be any A straightforward corollary is (equation (35) in [21]) Note that in the last theorem and corollary there is no more dependence on the choice of the multisection.
Step 5. The last step is to perform Laplace transform to the integrated formula described above. This is the content of [21], sections 6.
and to understand the signs. Here G i are the components of G. This is the content of section 6.2 in [21]. After understanding the sign and the ratio of forms, the Laplace transform calculations are straightforward and give if e is an internal edge between faces i and j; Summing over the different gradings K, and using the results of section 6.2 regarding the signs give λ(e).

A combinatorial formula for the refined and very refined numbers
In order to write a combinatorial formula for the more refined numbers, first note The proof is immediate, the operation B remembers the cyclic order of the illegal nodes on each boundary edge, hence remembers the topology of the graph on which B was applied. The edge contraction is easily inverted on the level of graphs, and the value of K on the contracted bridge can be read from knowing which side of the node the B operation declared to be illegal. The second part of the observation follows from the fact that the different gradings on G ′ do not change the way we invert ∂ e .
Note that Steps 1-3 of the previous section work without change for the more refined numbers, giving us (2.7) where SR m g,b,k,l is the subset of SR m g,k,l made of graphs whose smoothing has b boundary components, and s ′ is again combinatorial canonical. Define similarly SR m g,b,k,l , R m g,b,k,l and R m g,b,k,l . Define SR m g,k,l , SR m g,k,l , R m g,k,l , R m g,k,l , accordingly, for graphs which correspond to a partitionk of boundary marked points. Then acting similarly for the very refined numbers yields Step 4 requires some modification. Observation 2.13 allows us to apply Lemma 2.9 to the sets C obtained by taking an arbitrary (G, [K]) ∈ SR 0 g,k,l and creating all elements of SR m g,k,l obtained from it by contracting bridges and applying B. Using Lemma 2.9 iteratively now gives Theorem 2.14. For integers a 1 , . . . , a l ≥ 0 which sum to n = k+2l+3g−3

2
, let L be any

JHEP03(2017)123
Similarly, under the same assumptions, Step 5 follows without change, since the Laplace transform is performed cell-by-cell, and then summed over gradings, we see that for the refined numbers it holds that Theorem 2.16. Fix g, k, l ≥ 0 such that 2g − 2 + k + 2l > 0. Let λ 1 , . . . , λ l be formal variables. Then we have Moreover, suppose that b ≥ 1 and k ∈ P (k, b). Then we have λ(e).

JHEP03(2017)123
It will be useful for the future to rewrite formula (2.10) in the following way. For a graph G = ( i G i ) /N ∈ R * g,k,l introduce a combinatorial constant c(G) by c(G) : and e I (G i ) denotes the number of internal edges in G i , v B3 (G i ) is the number of boundary trivalent vertices and v BM (G i ) is the number of boundary marked points in G i . Then for any g, k, l ≥ 0, b ≥ 1 and k ∈ P (k, b) we have We consider the standard volume form on H M . In [6] the second and the third authors proved that where P a,b,m is a polynomial of degree m in expressions of the form tr(HΛ −d 1 HΛ −d 2 · · · HΛ −dr ), r ≥ 1.
Here the degree is introduced by putting deg(tr(HΛ −d 1 HΛ −d 2 · · · HΛ −dr )) := r + 2 r i=1 d i . Note that the integral is zero, if m is odd, and is a rational function in λ 1 , . . . , λ N of degree − m 2 , if m is even. The integral on the right-hand side of (3.3) is understood as the term-wise integral of (3.4) with respect to our Gaussian probability measure on H M . We refer the reader to [6] for a more detailed discussion.
Let us briefly recall the derivation of formula (3.3). It is obtained from the combinatorial formula (2.6), rewritten similarly to (3.2), using the standard matrix models technique. An odd critical nodal ribbon graph with boundary can be obtained from the disjoint union of critical non-nodal ribbon graphs with boundary by gluing boundary marked points. Since the sides of each node of the nodal graph are marked by plus or minus, we should assign pluses and minuses to the boundary marked points of the critical non-nodal ribbon graphs with boundary. A collar neighborhood of a boundary component of a critical nonnodal ribbon graph with boundary, that is not a ghost, is a circle with ribbon half-edges attached to it and also with boundary marked points (see figure 4). Such a circle with a configuration of ribbon half-edges and marked points will be called a boundary piece. We see that our odd critical nodal ribbon graph with boundary is obtained by • gluing a set of trivalent stars (see figure 5) and boundary pieces,  Remember also that, according to the definition of an odd critical nodal ribbon graph with boundary, the number of boundary marked points coming with plus on each boundary piece should be odd. We obtain that the trivalent stars give the contribution e The ghost components give the factor e Proof. We now use the combinatorial formula (3.2). As we explained in the previous section, an odd critical nodal ribbon graph with boundary is obtained by gluing trivalent stars (figure 5) and boundary pieces (figure 4), adding ghost components (figure 5) and then gluing boundary points to create nodes. The problem now is to control the distribution of the boundary marked points on the boundary components in a smoothing the resulting nodal ribbon graph with boundary. Our idea is the following. Consider the nodal surface with boundary that is associated with our nodal ribbon graph with boundary. Consider a small neighborhood of a boundary node of this surface. At this node two small pieces of boundary components meet. Then, instead of gluing these two pieces at one point, we connect them by a small ribbon edge (see figure 6). The new ribbon edge will be called an external ribbon edge. In figure 6 we fill the external ribbon edges by dots in order to distinguish them with the usual internal ribbon edges. Doing this procedure at each node, we obtain a non-nodal surface with boundary, which is a smoothing of the initial nodal surface. Note that each half of an external ribbon edge is marked by plus or minus. Note that the resulting non-nodal surface can be glued from elementary pieces in the following way. Again we have trivalent stars (figure 5). Then we have boundary pieces similar to what we have in the previous section, but now we want to replace each boundary marked point by an external ribbon half-edge, marked by plus or minus (see figure 7). In the same way we replace the ghost component from figure 5 by the ghost component with external ribbon half-edges (see figure 8). In order to have marked points we have to introduce an external ribbon half-edge marked by minus (see figure 9). Now, in order to JHEP03(2017)123   obtain our non-nodal surface with external ribbon edges, we glue a set of elementary pieces of four types (figures 5,7,8,9) according to the following rules: • An internal ribbon half-edge should be glued to an internal ribbon half-edge.
• An external ribbon half-edge with some sign should be glued to an external ribbon half-edge with an opposite sign.
For a polynomial P (Z) ∈ C[z ij , z kl ] let Then we have , we go around the boundary piece in the clockwise direction and look at the external ribbon half-edges that we meet. If an external ribbon half-edge is marked by plus then we assign to it the matrix Z and if it is marked by minus then we assign to it the matrix Z t . Then the function on Mat N,N (C) is the trace of the product of these matrices taken according to their order in the clockwise direction. So, the resulting function on H M × Mat N,N (C) is the product of two traces. Note that the product of the traces of two matrices is the trace of their tensor product. We obtain that all boundary pieces with external ribbon half-edges give the following contribution to the matrix model for τ o : where id M and id N are the identity matrices in the spaces H M and Mat N,N (C), respectively, and We see that the expression (3.8) is equal to Finally, the trivalent stars give the contribution e Using this theorem, we can obtain a matrix model for the refined open partition function τ o N in the following way: JHEP03(2017)123

Extended refined open partition function
The extended open partition function τ o,ext ∈ Q[[t 0 , t 1 , . . . , s 0 , s 1 , . . .]], introduced in [7,8], is uniquely determined by the following equations: Note that equation (3.9) gives a formula for τ o that is slightly different to the initial formula (3.3), where d 2 z := d(Re z)d(Im z). Formulas (3.10) and (3.11) imply that It is easy to see that So, we get This formula together with equation (3.9) motivates us to introduce a formal power series

JHEP03(2017)123
The uniqueness of a power series with this property is obvious. However, the existence of such a series is not trivial. In order to prove it we will define a formal power series τ o,ext N using the function τ o and then prove that it satisfies equation (3.14).  Proof. Note that .

(3.16)
Then the lemma follows from Theorem 3.1 and the elementary formula: For a finite value of N the transform, defined by the right hand side of (3.15) is not invertible. However, if we know τ o,ext N for all N ≥ 1, we can find τ o . Let us consider the space U N of unitary N × N matrices. Then we introduce the volume form on U N , which is proportional to the Haar measure and normalized by Let p 1 , p 2 , . . . and p ′ 1 , p ′ 2 , . . . be formal variables.

JHEP03(2017)123
where the polynomials h k (p 1 , p 2 , . . .), k ∈ Z, are defined by i≥0 h i z i = e i≥1 p i z i for k ≥ 0 and by h k := 0 for k < 0. Thus, it is enough to prove the lemma for g = s λ , where l(λ) ≤ N . (If l(λ) > N then both g| p i = 1 i tr Z i and f N are equal to zero.) For any matrix A ∈ Mat N,N (C) let For any partition µ and matrices A, B ∈ Mat N,N (C) we have the following formula ([1], eq. (39)), Then for any unitary matrix U we can compute On the other hand, for any partition µ and matrices A, B ∈ Mat N,N (C) we have [1], eq. (31) Therefore, we obtain This completes the proof of the lemma.
In particular, we have Equations
We see that this matrix model is obtained from (3.9) simply by adding the factor e i≥1 in the integrand. Doing the Feynman diagram expansion of (3.18) one can easily see that there is a combinatorial formula for the intersection numbers (3.17) similar to (2.9), where we allow odd critical nodal ribbon graphs with boundary to have certain exceptional components. Let us formulate it precisely.
Recall that a (g, k, l)-ribbon graph with boundary is called critical, if • Boundary marked points have degree 2.
• All other vertices have degree 3.
We will call a (g, k, l)-ribbon graph with boundary exceptional, if g = l = 0 and k ≥ 1.
Obviously, for each k ≥ 1 there exists a unique such graph up to an isomorphism, see figure 10. Note, that we step back a little bit from the original definition of a ribbon graph with boundary, because exceptional graphs with k = 1 or k = 2 are strictly speaking not stable. Note also that a critical (0, 3, 0)-ribbon graph with boundary, that we call a ghost, coincides with an exceptional graph with k = 3. However, our idea is to distinguish them. Speaking formally, to a (0, 3, 0)-ribbon graph with boundary we additionally assign a type: JHEP03(2017)123 it can be a ghost or an exceptional graph. Using this terminology, the set of critical ribbon graphs with boundary does not intersect the set of exceptional graphs. A nodal ribbon graph with boundary G = ( i G i ) /N will be called extended critical, if • It does not have boundary marked points.
• All of its components G i are critical or exceptional.
• Ghost components do not contain the illegal sides of nodes.
• Exceptional components do not contain the legal sides of nodes.
The fact, that we do not allow boundary marked points now, may look surprising, but one can note that an exceptional component with k = 1 can be easily interpreted as a boundary (3.19)

String equation
Proposition 3.4. We have the string equation  Proof. We will use formula (3.18). Denote Our approach is a modification of the diagrammatic method of E. Witten ([24]) that he used for a proof of the Virasoro equations for the closed partition function τ c . First of all, note that  The only non-trivial step in the proof is to express the derivative , as a matrix integral. Let us prove that (3.23) JHEP03(2017)123 Figure 11. Graphs that dominate for λ 0 → ∞.
The t 0 derivative corresponds to an extra insertion of τ 0 on the left-hand side of (3.17). We want to consider the generating function from the left-hand side of (3.19) with an extra insertion of τ 0 . In order to get it from the right-hand side of (3.19), we have to sum over graphs G = ( i G i ) /N ∈ R ext l+1 with a distinguished face, which we call C 0 , labeled with a variable λ 0 , then consider the behavior for λ 0 → ∞ and extract the coefficient of 1 λ 0 . The coefficient of 1 λ 0 comes precisely from graphs, where the face C 0 has only one edge. The structure of the neighborhood of the distinguished face in such graphs is indicated in figure 11. We see that there are two cases. In the first case, the edge of our face is internal. In the second case, the edge of the face is boundary. Then, automatically, the face belongs to a component G i of type (0, 1, 1). The first picture in figure 11 already appeared in [24] in the diagrammatic proof of the string equation for τ c . The contribution of this picture in our situation is computed in exactly the same way, as in [24], and it gives the first term tr H in the brackets on the right-hand side of (3.23). Consider the second picture in figure 11. A graph outside the dotted lines can be an arbitrary odd extended critical nodal ribbon graph with an additional distinguished illegal "half" of a node. The part inside the dotted lines gives 1 λ 0 . So, in order to get the contribution of the second picture, we should sum over all exteriors. It is easy to see that this sum gives the second term tr Z in the brackets on the right-hand side of (3.23).
Computations (3.21), (3.22) and (3.23) show that the string equation (3.20) is equivalent to the equation We want to compute the generating series from the left-hand side of (3.19) with an extra insertion of τ 1 . In order to get it from the right-hand side of (3.19), we have to sum over graphs G = ( i G i ) /N ∈ R ext l+1 with a distinguished face, which we call C 0 , labeled with a variable λ 0 , and then pick out the coefficient of 1

Therefore, equation (3.24) is equivalent to
can only come from graphs, where the face C 0 has at most three edges. The structure of such graphs in indicated in figure 12 and figure 13. We see that there are 10 cases and we divide them in two types. Graphs of internal type are those graphs where all the edges of the face C 0 are internal and graphs of boundary type are those graphs where at least one edge of the face C 0 is boundary. The diagrams inside the dotted lines in the top row in figure 12 are pieces of arbitrary larger graphs, while the graphs in the bottom row in figure 12 are special ribbon graphs corresponding to closed Riemann surfaces of genus 0 and 1 respectively. The five pictures in figure 12 already appeared in [24] in the diagrammatic proof of the dilaton equation for τ c . The contribution of these pictures in our situation is computed in exactly the same way, as in [24], and it gives the five terms in the integrand in line (3.29).
Let us consider graphs of boundary type. Let us look at the first picture in figure 13. Suppose that the face adjacent to C 0 is labeled with λ i . Then the diagram inside the dotted lines gives So, the coefficient of 1 Note that the graph outside the dotted lines can be an arbitrary odd extended critical ribbon graph with boundary with a distinguished face labeled by λ i and having a boundary edge. Now it is easy to see that the first picture in figure 13 gives the term − λ i to the Kontsevich-Penner tau-function τ N from (1.7). In the case N = 1 we show how to relate directly our matrix model (3.14) to the Kontsevich-Penner matrix model. We also discuss more evidence for the conjecture. In particular, we show that the conjecture is true in genus 0 and 1. Remark 4.1. Note that in [2] our variables T k are denoted by t k . Note also that we write the Kontsevich-Penner matrix integral in a way slightly different from [2] (see formula (1.1) there). In order to identify formula (1.1) from [2] with the right-hand side of (1.7), one has to make the shift Φ → Φ + Λ and then the variable change Φ = −H.
In [2,3] the first author proved that We propose the following conjecture.

Case N = 1
In [2,3] the relation between τ o,ext and τ 1 was established with the help of some properties of the integrable hierarchies. In this section we prove directly that for N = 1 the integral representation Because of the Dirac delta-function, the last integral reduces to the one over the Hermitian matrices of the formΦ where H is an M × M Hermitian matrix and C is a complex vector. Since Remark 4.4. We expect that a similar argument can be applied for any positive integer N .

String and dilaton equations
String and dilaton equations for the Kontsevich-Penner model were derived in [2,5] (In a more general setup of the Generalized Kontsevich Model the string equation in terms of the eigenvalues of the external matrix was derived already in [12]). They coincide with the equations for the extended refined open partition function, derived in sections 3.5 and 3.6.

(4.3)
Let us insert genus parameters on the both sides of this equation. Let us look at the combinatorial formula (3.19). An elementary computation shows that for a graph G ∈ R ext  Let us now look at the numbers θ a 1 · · · θ an KP,N . For n ≥ 1 denote by R KP n the set of critical ribbon graphs with boundary, but with no boundary marked points and n internal faces together with a bijective labeling α : Faces(G)  We see that, similarly to the intersection numbers (3.17), the number θ a 1 · · · θ an KP,N is a polynomial in N with rational coefficients. It is easy to see that a graph G ∈ R KP n contributes only to intersection numbers θ a 1 · · · θ an KP,N with a i = 3(g(G) − 1 + n). So, for a non-negative integer g we define θ a 1 · · · θ an KP,N g to be equal to θ a 1 · · · θ an KP,N , if a i = 3(g − 1 + n), and to be equal to 0 otherwise. The combinatorial formula (4.6) immediately implies that θ a i KP,N g is an odd polynomial in N of degree ≤ g + 1, if g is even, an even polynomial in N of degree ≤ g + 1, if g is odd. Properties (4.4) and (4.7) agree with the conjectural equation (4.3). Also these properties together with equation (4.1) imply that Conjecture 4.2 is true for N = −1. Equations (4.5) and (4.8) together with (4.1) imply that the equation τ a 1 · · · τ a l σ c 1 · · · σ c k o,ext,N g = θ 2a 1 +1 · · · θ 2a l +1 θ 2c 1 +2 · · · θ 2c k +2 KP,N g i (2a i + 1)!! j 2 c j +1 (c j + 1)! (4.9) is true for g = 0 and g = 1.
We have also checked equation (4.9) in several cases in genus 2.