Combined sub-sampling and analytical integration for efficient large-scale $GW$ calculations for 2D systems

Accurate and efficient predictions of the quasiparticle properties of complex materials remain a major challenge due to the convergence issue and the unfavorable scaling of the computational cost with respect to the system size. Quasiparticle $GW$ calculations for two dimensional (2D) materials are especially difficult. The unusual analytical behaviors of the dielectric screening and the electron self-energy of 2D materials make the conventional Brillouin zone (BZ) integration approach rather inefficient and require an extremely dense $k$-grid to properly converge the calculated quasiparticle energies. In this work, we present a combined non-uniform sub-sampling and analytical integration method that can drastically improve the efficiency of the BZ integration in 2D $GW$ calculations. Our work is distinguished from previous work in that, instead of focusing on the intricate dielectric matrix or the screened Coulomb interaction matrix, we exploit the analytical behavior of various terms of the convolved self-energy $\Sigma(\mathbf{q})$ in the small $\mathbf{q}$ limit. This method, when combined with another accelerated $GW$ method that we developed recently, can drastically speed-up (by over three orders of magnitude) $GW$ calculations for 2D materials. Our method allows fully converged $GW$ calculations for complex 2D systems at a fraction of computational cost, facilitating future high throughput screening of the quasiparticle properties of 2D semiconductors for various applications. To demonstrate the capability and performance of our new method, we have carried out fully converged $GW$ calculations for monolayer C$_2$N, a recently discovered 2D material with a large unit cell, and investigate its quasiparticle band structure in detail.


I. INTRODUCTION
Two dimensional (2D) materials are at the center of materials research in recent years.
The intense research activities have resulted in the discovery of an impressive and growing list of 2D materials that were once considered rare and unstable. Among them, 2D semiconductors have received particular attention for their potential use in future electronics and energy related applications. With the increasing role that theory plays in the design and prediction of 2D semiconductors, the importance of accurate understanding of their electronic structures cannot be overstated. Although the GW approximation [1 -3] has been recognized as one of the most accurate theories for predicting the quasiparticle properties of a wide range of materials, straightforward applications of the GW method to 2D materials have been met with multiple computational challenges that make fully converged GW calculations (even at the G 0 W 0 level) rather difficult. These challenges are so grave that, if not properly addressed, they may lead to false theoretical predictions and confusions.
One of the difficulties of 2D GW calculations comes from the Brillouin zone (BZ) integration of the GW self-energy, which is often carried out using discrete summation on a uniform k-grid (N 1 × N 2 × 1 for 2D systems): where Σ nk (q, ω) is the contribution to the GW self-energy for state |nk from point q in the BZ, Ω is the volume of the BZ, and f q is the appropriate weight. This summation typically converges rather quickly with respect to the BZ sampling density for bulk (3D) semiconductors. For example, for silicon (diamond structure with a 2-atom unit cell), a 6 × 6 × 6 k-grid is sufficient to converge the calculated GW band gap to within 0.01 eV.
Although this slow convergence issue is now well understood, it was somewhat unexpected at first. Since the computational cost of GW calculations scales as O(N 2 k ), where N k is the number of the BZ integration points, the slow BZ integration convergence issue in 2D GW calculations has significantly hindered practical applications of the GW method for accurate 2D materials predictions.
Compounding matters further is the need to include a large vacuum layer in the modeling of 2D systems using the periodic supercell approach (to minimize the spurious interlayer interactions), resulting in a large cell volume even for relatively simple 2D materials with only a few atoms in the unit cell. This is particularly true for theories (such as the GW method) that involve the calculations of nonlocal interactions or response functions. The calculated quasiparticle energies converge extremely slowly with respect to the vacuum layer thickness d if unmodified long-range Coulomb interaction is used [8][9][10]. Although the use of truncated Coulomb interaction [9,10] greatly expedites the convergence with respect to d, the calculated results still depend on the layer separation (albeit on a much weaker degree), and one still need to include a sizable vacuum layer of about 20Å or greater for most 2D materials.
The large cell volume translates into the need to include a large number of electronic states in GW calculations. For example, it has been shown [4,5,11,12] that one may need to include up to 10,000 conduction bands in the conventional GW calculations even for simple 2D materials with a small unit cell of a few atoms. Note that in order to reach a similar level of convergence, this number scales linearly with the system size (i.e., number of atoms in the unit cell), making fully converged GW calculations for more complex 2D systems extremely difficult using the conventional band-summation approach.
Recently, we developed an accelerated GW approach that can drastically speed up GW calculations for large systems [11]. In this method, the computationally demanding bandsummation in conventional GW calculations is replaced by an energy-integration method, resulting in a speedup factor of up to two orders of magnitude for large and/or complex systems, including 2D materials. The slow BZ integration convergence issue, however, still poses a formidable challenge for 2D GW calculations. Considering the importance of accurate predictions of the quasiparticle properties of 2D materials, it is not surprising that there have been several proposed schemes that aim at addressing the slow BZ integration convergence issue, noticeably the work of Rasmussen et al. [13] and that of da Jornada et al. [7].
Motivated by these works, we present here an efficient and accurate yet simple-to-implement method that can significantly reduce the required BZ sampling density for well converged 2D GW calculations. We have tested our method for a range of 2D semiconductors [12,14], and, for most cases, the calculated GW quasiparticle energies converge to within 50 meV or less using a very coarse 6 × 6 × 1 k-grid. Combining these two new approaches, we are able to carry out fully converged 2D GW calculations with an overall speed-up factor of over show the smallest q (other than q = 0) included in a 6 × 6 × 1 (2D) or 6 × 6 × 6 (3D) grid. three orders of magnitude compared with the conventional approach.

II. RESULTS
A. Analytical behavior of the GW self-energy of 2D systems The slow BZ integration convergence issue in 2D GW calculations is a manifestation of the asymptotic behavior of Σ nk (q, ω) (defined in Eq. 1) in the long wavelength (small q) limit, which is related to the analytical properties of the dielectric function ǫ −1 GG ′ (q, ω), or equivalently, that of the screened Coulomb interaction W GG ′ (q, ω). These quantities vary rapidly as the wave vector q approaches zero, making a simple discrete summation using the uniform sampling scheme very difficult to converge. If we write the BZ summation of the GW self-energy into two parts, it becomes clear that most convergence error comes from the q = 0 term, or, more precisely, the contribution from the mini-BZ centered around the Γ point. In fact, even in conventional GW calculations using a uniform k-grid, the contribution from the q = 0 term has to be treated carefully due to the divergence of the Coulomb interaction. This is typically done by exploiting the analytical behavior of the dielectric matrix and the (truncated) Coulomb interaction at the small q limit and carrying out a mini-BZ averaging of the screened Coulomb matrix, as has been implemented in the BERKELEYGW package and has been discussed in great details in previous works [2,9,15]. The large black dots in the figure correspond a sampling point in a 6 × 6 × 1 k-grid for the monolayer hBN and that in a 6 × 6 × 6 k-grid for bulk cBN. Whereas ǫ −1 00 (q) of bulk cBN varies smoothly as q approaches 0, due to the diminishing 2D dielectric screening in the long wave length limit, there is a sharp upturn of this quantity at small q for the monolayer hBN system. Accurate capturing of such rapid variation would require an extremely dense k-grid if uniform sampling schemes were used. Note that, strickly speaking, the dielectric function for a 2D system calculated using periodic boundary conditions is not a truly 2D dielectrc function but that of the 3D model system. It has been shown [9], however, that if a truncated Coulomb potential is used, the calculated GW self-energy converges quickly with increasing interlayer separation. Therefore, it is compelling to exploit the analytical behavior of the dielectric function and that of the screened Coulomb interaction to achieve converged GW results without the need to use a very high density BZ sampling grid. Rasmussen et al. [13] proposed a well-motivated analytical model for the screened Coulomb interaction in the long wavelength limit for 2D systems and carried out the integration of the self-energy in the mini-BZ centered around Γ. Figure 1 (c) shows the BZ of a 2D hexagonal system. A 6 × 6 × 1 uniform sampling grid is shown with black dots in the figure; the shaded area is the mini-BZ centered around the Γ point. Using this method, Rasmussen et al. [13] showed that the calculated quasiparticle band gap of monolayer MoS 2 converges to about 0.1 eV using a 12 × 12 × 1 k -grid. Although this is a significant achievement, a 12 × 12 × 1 k-grid is still fairly dense, and it is desirable to further reduce the required BZ sampling density. Note that the computation cost of GW calculations scales as O(N 2 k ), where N k is the number of the BZ integration points, a small reduction in the k-grid density will result in significant saving of the computation time. For example, by reducing the 2D k-grid density from 12 × 12 × 1 to 6 × 6 × 1, the computational cost would be reduced by a factor of 16.
Finding a compact and reliable analytic model for the the response function for a wide range of 2D materials, even in the small q limit, is difficult. Instead of exploiting the analytical behavior of the dielectric function, recently, da Jornada et al. [7] proposed a non-uniform sub-sampling scheme to improve the quality of discreteel at BZ integration.
The screened Coulomb interaction matrix in the mini-BZ is approximated by a weighted summation of a few sub-sampling points in the mini-BZ as shown schematically in Fig. 1 (d).
Using the method, it was shown that the quasiparticle band gap of bilayer MoSe 2 converges to within 50 meV using a coarse 2D k-grid of 6 × 6 × 1 with 10 sub-sampling points in the mini-BZ. In both of these methods, a better convergence is achieved with more accurate evaluation of the rapid variation of the screened Coulomb interaction matrix W GG ′ (q) within the mini-BZ. However, instead of working on the screened Coulomb interaction matrix, we believe that it is more efficient to exploit the analytical behavior of q-dependent self-energy contribution Σ nk (q) directly.
The electron self-energy can be conveniently separated into two parts, a screened exchange (Σ SEX ) and a Coulomb hole (Σ COH ) part [2]; the screened exchange part can be further separated into a bare exchange (Σ X ) and a correction term (Σ SX ) arising from the screening potential: (3) Figure 2 shows these self-energy terms (solid dots) for the valence band maximum (VBM, left panels) and conduction band minimum (CBM, right panels) states of monolayer MoS 2 as a function of wave vector q. Note that it is the integration of these contributions over the BZ that gives the self-energy correction for the electronic state (e.g., VBM or CBM) of interest.
Interestingly, all these quantities show well-behaved asymptotic properties. Therefore, it is very important to analyze and exploit the analytical behavior of these quantities in the small q limit.
We first examine the screened exchange energy for state |nk : where M vn (q, k, G) = v, k + q|e i(q+G)·r |n, k are plane-wave matrix elements between the two states |v, k + q and |n, k . It is convenient to write the screened Coulomb potential as the summation of the bare Coulomb potential v b and the screening potential v scr , i.e., Correspondingly, the screened exchange energy can be separated into the bare exchange and a correction coming from the screening potential, i.e., Σ SEX = (Σ X + Σ SX ), as mentioned earlier.
The analytical behavior of the bare exchange energy Σ X (q) can then be understood by examining the truncated 2D Coulomb potential [9] in the momentum space: where q is the wave vector within the 2D BZ, L z is periodicity along the z direction, G ( G z ) denotes the G vectors that are parallel (perpendicular) to the 2D atomic layer. The truncated Coulomb potential then approaches 2πL/|q | in the small q limit. For simplicity, we will drop the parallel sign ( ) for wave vectors q within the 2D BZ if there are no confusions. Therefore, it is straightforward to speculate that the leading term of the bare exchange energy for the valence (occupied) states has the same asymptotic expression as the bare Coulomb potential. Extending the expression to finite q, for 2D isotropic systems, we The solid curve of the top-left panel of Fig. 2 shows a perfect 3-parameter fitting of Σ X VBM (q) of monolayer MoS 2 calculated on four q points indicated with large black dots. With this fitted expression, the integration of Σ X nk (q) within the mini-BZ can be carried out analytically. Due to the absence of the self-exchange, the bare exchange for conduction (unoccupied) states is much smaller than that for occupied states, and is basically featureless, as shown in the top-right panel of Fig. 2, which can be well fitted with a 2-or 3-parameter function, i.e., using values calculated on 4 q points as shown with solid curve in the top-right panel of Fig.   2.
The correction to the exchange energy arising from the dielectric screening of the Coulomb potential can also be analyzed. The 2D dielectric function takes the form [13,16] ǫ(q) ≈ 1 + 2πα 2D q (8) in the long wavelength limit, where α 2D is the 2D polarizability. Therefore, the screening potential v scr takes the form v scr ≈ 4π 2 α 2D /(1 + 2πα 2D q). Considering the dynamical screening effects, we propose the following analytical form for Σ SX nk (q, ω) for both valence and conduction states: The middle panels of Fig. 2 show the 3-parameter fittings for Σ SX (q) for the VBM and CBM states of monolayer MoS 2 calculated at their respective DFT energies. Finally, we find that the Coulomb hole self-energy can also be well fitted with the same analytical form, i.e., for both the valence and conduction states as shown in the bottom panels of Fig. 2.
We have implemented a nonlinear fitting algorithm (the Levenberg-Marquardt algorithm) in our code. We monitor the fitting quality, i.e., the residual error, so one can easily spot possible issues with the fitting procedure. For all systems we have studied, the fitting procedure converges quickly with a reasonable initial guess (e.g., by setting all inititial parameters to 1.0). In order to demonstrate the reliability and quality of the proposed fitting functions and that of the implemented fitting algorithm, we have calculated the selfenergy at a few additional q points and have added these data points (red dots) to Fig.   2. These additional data points agree well with the functions fitted using the original data (black dots).

B. Combined sub-sampling and analytical integration approach
Putting these results together, we propose an approach that have the advantages of both of the previously proposed schemes [7,13], a combined sub-sampling and analytical integration of the self-energy within the mini-BZ, to tackle the convergence issue of the BZ integration in 2D GW calculations. The BZ is sampled with a coarse uniform k-grid as usual; a 6 × 6 × 1 grid is sufficient for most 2D systems with small unit cells. For complex 2D materials with large unit cells, an even coarser k-grid may be used as we will discuss later. We then carry out a few additional sampling points inside the mini-BZ. The three qdependent GW self-energy terms, namely, Σ X (q), Σ SX (q, ω), and Σ COH (q, ω) are calculated on these additional sampling points and the results are fitted using the analytical functions discussed in the previous section. The BZ integration of the GW self-energy is separated into two parts, a conventional weighted summation over all k-points except the Γ point, and an integration of the fitted analytical functions over the mini-BZ: where Ω 0 is the area of the 2D mini-BZ as shown in Fig. 1 (d).
The self-energy Σ nk (q, ω) is tyipcally calculated at two energy points, ω = ǫ DFT nk , and ω = ǫ DFT nk +∆ǫ. A linear expansion [2] of the self-energy is then carried out to obtain the selfenergy evaluated at the quasiparticle energy, i.e., Σ nk (ω = E QP nk ). Since the integration over the mini-BZ is carried out using the fitted analytical functions as opposed to the weighted summation approach, we need only a small number of sub-sampling points. In fact, for all isotropic 2D systems we have studied, 4 additional sampling points are sufficient to converge the calculated quasiparticle energy to within 0.01 eV for a given N × N × 1 BZ sampling grid, as discussed in the next section. We mention that our method can be extended to treat anisotropic 2D materials. In this case, the sub-sampling calculations within the mini-BZ have to be carried out along the two reciprocal lattice directions b 1 and b 2 , and some of the fitting coefficients are vectors instead of scalars. We will report results for anisotropic 2D systems in a separate publication.

C. Convergence behavior of the GW band gap of monolayer MoS 2
We first demonstrate the performance of our method using monolayer MoS 2 as an example. The quasiparticle properties of monolayer MoS 2 have been investigated by several groups [4-6, 11, 17, 18]. Therefore, this system serves as a good model for testing our methods. Figure 3  We include in Fig. 3 (a) the results of Rasmussen et al. [13] (black solid curve) for comparision. It should be mentioned that our result seems to agree with that of Rasmussen et al. [13] calculated with a 24×24×1 k-grid. This is a coincidence rather than a confirmation considering various differences (e.g., pseudopotential, crystal structure, and several cutoff parameters) in the two calculations.
We have tested the calculated band gap with respect to the number of sub-sampling points in the mini-BZ, as shown in Fig. 3 (b). The result essentially converges with 3 subsampling q points. The extremely small error (< 5 meV) likely comes from numerical errors instead of from the systematic convergence error. We have also tested the sensitivity of the results on the choice of the sub-sampling q-points, and we can confirm that the results are fairly insensitive. Different choices of the sub-sampling q-points within a given N × N × 1 k-grid give practical identical results; the difference is usually within a few meV. This is expected since the calculated self-energy can be fitted extremely well with the proposed functional forms as shown in Fig. 2.
Although the main focus of this work is to address the slow convergence issue of the BZ integration in GW calculations for 2D materials, we would like to discuss other convergence issues in GW calculations. These issues may become another bottleneck for GW calculations  shows the actual number the integration points using our method [11].
Our calculations also benefit from the energy-integration method [11,19] we developed to speed up the band summation in GW calculations. As we have mentioned earlier, conventional GW calculations for 2D materials require to include a large number of conduction bands, making highly converged calculations even for simple 2D materials containing a few atoms very difficult. The total number of the empty states in our calculations for the monolayer MoS 2 is about 25,000, and one needs to include about 10,000 bands to properly converge the band gap (to within 0.01 eV) as shown in Fig. 5 (a). Using our energy integration method, we only need about 740 integration (sampling) points to achieve the same level of convergence. Similar to what we have discussed earlier, the change in the calculated band gap with respect to the number of bands included in the GW calculations is much smaller than the change in the VBM (or CBM) quasiparticle energy due to error cancellation, as shown in Fig. 5 (b). However, there are situations in which highly converged results for the quasiparticle energyies (not just the band gap) are required. One of the advantages of our method is that we can afford to include (effectively) all empty states in our GW (both for the dielectric matrix and the self-energy) calculations without the need to concern about the band summation convergence issue. In order to further demonstrate the capability and performance of our method, we now investigate the quasiparticle band structure of C 2 N [20], an interesting 2D carbon nitride that is distinguished from other 2D systems by its unique holey structure as shown in Fig.   6. The theoretically optimized lattice constant is 8.29Å; the three unique bond lengths are shown in the figure. The structure has a large unit cell of 18 atoms, making fully converged GW calculations a real challenge. In fact, C 2 N has a 2D unit cell area that is equivalent to that of a 24-atom graphene supercell.
The basic electronic structure of monolayer C 2 N has been studied by several groups [21][22][23][24]. One interesting feature of the band structure of monolayer C 2 N is that the top valence bands are nearly dispersion-less if local or semilocal energy functionals within DFT are used.
These flat valence bands are primarily derived from nitrogen and carbon p x and p y orbitals as shown in Fig. 7 (a) and (c). The two valence bands immediately below the top two (at the Γ point), in contrast, are mostly derived from carbon p z orbitals with small nitrogen p z components as shown in Fig. 7 (b) and (d). Since the in-plane (p x and p y ) states may experience significantly different quasiparticle self-energy corrections compared with the out-of-plane p z states, the ordering of these closely spaced valence bands may change after including GW self-energy corrections, which would have important consequences on the calculated optical and transport properties of this material. In the following, we first discuss the converged quasiparticle band structure of monolayer C 2 N and discuss its important features compared with that calculated using the LDA. We then discuss several important convergence issues of the GW results. within the LDA drop below the p z derived states after including the GW correction. The p z derived states become the top-most valence states and are more dispersive.
To better illustrate the change in the band ordering, we show the zoomed-in band structure around the Γ point in the right panels of Fig. 8. We note that a similar valence band ordering change was observed earlier [23] with the use of HSE06 hybrid functional [25,26]. Interestingly, we find that the band ordering change also occur to the conduction bands (although not as significant as that of valence bands) as shown in the right panels of Fig. 8. These changes in the ordering of the band edge state will have profound impact of the calculated optical and transport properties of this material, which deserve further investigations.
We now discuss several important convergence issues of GW calculations of this material. k-point density using the uniform sampling approach and the current method. Due to its relatively large unit cell (thus a small BZ), the calculated band gap converges to within 0.02 eV using a very coarse 3 × 3 × 1 k-grid, or within 0.01 eV using a 4 × 4 × 1 k-grid, with our BZ integration method. Quasiparticle GW calculations of monolayer C 2 N have been reported earlier [22]. The authors used a very small cutoff energy (5 Ry) for the dielectric matrix and included only a few hundred bands in the calculations of the dielectric matrix and the self-energy. The reported GW band gap of monolayer C 2 N was 3.75 eV [22], to be compared with our result of 3.54 eV.
As we have discussed earlier, for many systems, the calculated GW band gap may appear to converge while the absolute quasiparticle energies for the valence and conduction bands are still not converged. This is because the valence and conduction bands may have the similar convergence behavior, and their difference (which defines the band gap) may appear to converge quickly. In fact, a fairly high kinetic energy cutoff for the dielectric matrix and a large amount of conduction bands are still needed in this case to achieve highly converged results for the quasiparticle energies of this system.

III. METHODS
We use the crystal structures optimized using the Perdew-Burke-Ernzerhof (PBE) functional [27] for subsequent electronic structure calculations. The optimized lattice constant for MoS 2 is 3.18Å, and the layer thickness (i.e., the S-S interlayer distance) is 3.16Å.
These values are in reasonable agreement with published theoretical results. The detail of the crystal structure of C 2 N will be discussed later. The monolayer systems are modeled with periodic cells with an interlayer separation of 25Å. The mean-field electronic structure calculations are carried out using the pseudopotential plane-wave-based density functional theory (DFT) method within the local density approximation (LDA) as implemented in a local version of the PARATEC package [28][29][30]. The Perdew-Zunger [31] parametrization of the Ceperley-Alder result [32] for the electron correlation energy is used. We use the Troullier-Martins norm-conserving pseudopotential [33]. Semicore 4s and 4p of Mo are included in the calculation. The plans wave cutoff for the DFT and GW calculations for MoS 2 is set at 125 Ry; for C 2 N, it is 70 Ry.
The GW quasiparticle calculations are carried out within the G 0 W 0 (i.e., one-shot GW ) approach [2] using a local version of the BERKELEYGW package [15] in which the method described in this work and a recently developed energy-integration method [11,19] are implemented. The summation over the conduction bands in GW calculations is carried out using the energy-integration approach [11,19]. Using this method, we can effectively include all conduction bands in the calculations at a fraction of the computational cost compared with the conventional band-by-band summation. The kinetic energy cutoff for the dielectric matrix is set at 75 Ry for MoS 2 and 40 Ry for C 2 N. These cutoffs are sufficient to converge the calculated quasiparticle band gap to within 0.02 eV. We use the Hybertsen-Louie generalized plasmon-pole model (HL-GPP) [2] to extend the static dielectric function to finite frequencies.

IV. DISCUSSION
Accurate and efficient GW calculations for 2D materials are met with a multitude of computational challenges. The computational cost of fully converged GW calculations for 2D materials, even for simple materials with small unit cells of a few atoms, can be very expensive, making reliable GW calculations for large and/or complex 2D systems a daunting task. The formidable computational demand has significantly held back the widespread adoption of this otherwise highly successful method for 2D materials predictions.
By carefully investigating the analytical behavior of the GW self-energy, we proposed a combined sub-sampling and analytical integration method that can greatly improve the efficiency of 2D GW calculations, enabling fast and accurate quasiparticle calculations for complex 2D systems. For most simple 2D materials with a small unit cell of a few atoms, a 6 × 6 × 1 2D BZ sampling grid is sufficient to converge the calculated quasiparticle band gap to within 0.02 ∼ 0.05 eV, resulting in a speed-up factor of over two orders of magnitude compared with the conventional uniform sampling approach. This method, when combined with another method that we developed earlier [11], results in a speed-up factor of well over three orders of magnitude for fully converged GW calculations for 2D materials.
To demonstrate the capability and performance of our method, we have carried out fully converged GW calculations for monolayer C 2 N, a recently discovered 2D material with a large unit cell of 18 atoms, and investigated its quasiparticle band structure in detail. Our calculations not only provide most converged results but also reveal interesting features of the near-edge electronic properties of this interesting 2D material.
With these development, we can carry out fully converged GW calculations for complex and/or large 2D materials with moderate computational resources that are available to most research groups. We believe that our developments will greatly facilitate future high throughput screening of the quasiparticle properties of 2D semiconductors for various applications. Note that our method only works for 2D semiconductors since the dielectric function and the electron self-energy for 2D metallic systems have different analytical behaviors. In addition, capturing the intra-band transitions in metallic systems may still require a fairly dense k-grid. It would be interesting to find out if current approach can be extended to metallic systems.

DATA AVAILABILITY
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

CODE AVAILABILITY
The code developed in this work will be made available from the corresponding author after optimization and on reasonable request.

COMPETING INTERESTS
The authors declare no Competing Financial or Non-Financial Interests.