Rhombohedral Graphene: A Tale of Many Crystals Ahmed Abouelkomsan, Filippo Gaggioli, Daniele Guerci, Liang Fu Abstract Experiments on rhombohedral graphene have uncovered an extraordinary wealth of correlated quantum phases—from chiral superconductors to electronic crystals—all within a single family of atomically thin materials. Here, we introduce a simple indicator, derived from the noninteracting band dispersion, that identifies strongly correlated regions in the phase diagram of rhombohedral graphene as a function of carrier density and displacement field. We develop a neural-network variational Monte Carlo method, combined with Hartree-Fock theory, to solve the interacting ground states. Our calculation reveals a variety of electron crystals with no classical analog. These include, at increasing density: Wigner crystal, self-doped Wiger crystal, as well as “anticrystal”—a lattice of holes in an electron liquid. We discuss their experimental manifestations and possible connection to superconductivity. Introduction — Rhombohedral graphene has emerged as a most versatile platform for strongly correlated quantum matter. Within a single family of atomically thin carbon multilayers, experiments have now uncovered an exceptional range of distinct quantum phases including integer and fractional topological phases [ 26 , 37 , 36 , 55 , 8 , 4 ] , chiral superconductors [ 24 , 14 , 41 , 45 , 27 , 40 ] , and insulating and metallic electronic crystals [ 23 , 59 ] . While the phase diagram tuned by carrier density and displacement field is remarkably rich, this richness also poses a central challenge: identifying the organizing principles that control the emergence, competition, and coexistence of these interaction-induced phases. A key feature of rhombohedral graphene is that the displacement field reshapes its low-energy band structure and can flatten the bottom of the band, leading to strong correlation effects. In particular, some of these chiral superconducting states appear near insulating states at lower density [ 24 ] , which are believed to be electron Wigner crystals. The complex phase diagram in rhombohedral graphene as a function of carrier density n and displacement field D , as well as its evolution with layer thickness, calls for theoretical understanding. In this work, we find a useful guide for navigating the phase diagram of rhombohedral graphene, and present a comprehensive study of its (n,D) phase diagram using state-of-the-art neural-network variational Monte Carlo and Hartree-Fock theory. Specifically, we introduce a simple indicator of strongly correlated regions in the phase diagram based on the noninteracting energy dispersion. Guided by this indicator, our large-scale numerical simulations find a plethora of crystalline states in different regions of the phase diagram, enabled by the tunable fermiology of rhombohedral graphene and its interplay with Coulomb interaction. Besides the standard Wigner crystal, we predict the existence of quantum crystalline phases without classical analog, in the regime of annular Fermi sea. These include i) a self-doped crystal with non-integer filling of the unit cell, and ii) an “anticrystal” featuring a triangular lattice of holes (rather than localized electrons). Fermiology of Rhombohedral Graphene — In rhombohedral multilayer graphene, near charge neutrality, the low-energy bands originate from sublattice polarized states localized on the top and bottom layers [ 32 , 57 ] . Importantly, these low-energy states are coupled through a higher-order tunneling process whose order is determined by the number ( L ) of interlayer hopping events required to connect them, resulting in a coupling that scales as |{\boldsymbol{k}}|^{L} . A perpendicular displacement field D induces a potential-energy imbalance between these layer-polarized states, opening a gap and providing a direct knob to tune the band dispersion. As shown in Fig. 1 (a), increasing D progressively flattens the bottom of the conduction band until a critical value D_{c} , at which the band minimum shifts from {\boldsymbol{k}}=0 (the Dirac point) to a finite momentum |{\boldsymbol{k}}|=k_{0} . Above D_{c} , the band is reshaped into a Mexican-hat dispersion. Consequently, at fixed carrier density, the Fermi surface changes from circular to annular, as sketched in Fig. 1 (b). At finite doping, the non-interacting Fermi surface topology also changes with the carrier density n . In particular, in the Mexican hat regime, as n decreases the system changes from having a single electron pocket centered at {\boldsymbol{k}}=0 to an annular Fermi sea with an outer electron-like Fermi surface and an inner hole-like Fermi surface. Additional details can be found in the Supplementary Materials (SM) [ 1 ] . Figure 1: Microscopic band structure (a) and illustration of the simply-connected and annular Fermi sea regimes (blue shaded), with the black and red lines indicating electron- and hole-like Fermi surfaces, respectively. Ratio of interaction energy to kinetic energy in 4 (c) and 5 layer (d) rhombohedral graphene, with relative dielectric constant \epsilon=5 . Our goal is to explore the consequence of the underlying band structure on the interacting phase diagram. As a first step, we introduce a dimensionless quantity that measures the strength of Coulomb interactions relative to the kinetic energy in this system [ 21 ] . Specifically, we define r_{\rm eff}\equiv\frac{E_{\rm int}}{E_{\rm kin}}=\dfrac{e^{2}/(4\pi\epsilon\epsilon_{0}d_{e})}{\langle\varepsilon({\boldsymbol{k}})\rangle}, (1) where E_{\rm int} is the Coulomb energy evaluated at the average interparticle distance d_{e}=1/\sqrt{\pi n} , E_{\rm int}=e^{2}/(4\pi\epsilon_{0}\epsilon d_{e}) and \langle\varepsilon({\boldsymbol{k}})\rangle is the kinetic energy per-particle, averaged over all occupied states in the non-interacting limit [ 1 ] . For conventional two-dimensional electron gas with parabolic dispersion, r_{\rm eff} reduces to the Wigner–Seitz radius r_{s} determined by the carrier density ( r_{\rm eff}\propto n^{-1/2} ), and large r_{s} at low density leads to the emergence of an electron Wigner crystal, driven by Coulomb interaction. In contrast, in rhombohedral graphene the ratio r_{\rm eff} depends not only on the density, but also on the form of the underlying band dispersion tuned by the displacement field [ 21 ] . Figure 1 (c) and 1 (d) show r_{\rm eff} as a function of density n and displacement field D for four- and five-layer rhombohederal graphene, respectively, which allows us to identify the region of strong correlation. Remarkably, in the (n,D) plane, the region of large r_{\rm eff}\geq 20 —which corresponds to strongly correlated liquid and Wigner crystals in the case of two-dimensional electron gas—matches well the region of experimentally observed insulating and superconducting states. In particular, r_{\rm eff} generally evolves non-monotonically with the displacement field D and the density n . This trend mirrors the experimentally observed sequence in rhombohedral four-layer graphene [ 25 ] in the density range 0.2-0.6\times 10^{12}\,\mathrm{cm}^{-2} , where, at fixed n , the ground state evolves from a metal to a highly resistive state and then back to a metal as the displacement field is increased. It is instructive to compare the r_{\rm eff} indicator—which takes into consideration all states below the Fermi level—with the Stoner criterion: UD(E_{F})>1 , where D(E_{F}) is the noninteracting density of states at the Fermi level. In the Stoner picture, strong interaction effects are expected at the van-Hove singularity due to the divergent density of states. However, while the line of van-Hove singularity for rhombohedral graphene continues to very high density and displacement field [ 1 ] , the interaction-driven states are only observed below a certain density \sim 1\times 10^{12}\,\mathrm{cm}^{-2} , as correctly captured by r_{\rm eff} . Our r_{\rm eff} plot also correctly identifies strongly correlated region of N=5,6 layer rhombohedral graphene. As shown in Fig. 1 (d), large r_{\rm eff} appears at smaller displacement fields and extends over a broader range of electron densities compared to N=4 [ 25 , 23 ] , which reflects the flatter band dispersion – similar conclusions can be drawn for N=6 layers, see the SM [ 1 ] . Thus, our indicator r_{\rm eff} provides a useful guide for navigating the rich phase diagram of correlated electron states in rhombohedral graphene. The discussion above highlights two key ingredients of rhombohedral multilayer graphene: the nonmonotonic evolution of the correlation strength and the topological transition of the Fermi surface from a single electron pocket to multiple pockets. In the following, we develop state-of-art neural network methods to solve many-body ground states variationally, focusing on the region of large r_{\rm eff} and annular Fermi sea. Figure 2: (a)–(d) Ground-state charge density obtained from NN-VMC, illustrating four representative crystalline states. Blue denotes regions of high electron density, while white denotes regions of low density. The parameters are (\alpha/E_{C},\gamma/E_{C},n)=(0.582,0.740,0.86\times 10^{12}~{\rm cm}^{-2}) for (a), (0.740,0.740,0.659\times 10^{12}~{\rm cm}^{-2}) for (b), (0.232,0.702,0.293\times 10^{12}~{\rm cm}^{-2}) for (c), and (0.544,0.739,0.293\times 10^{12}~{\rm cm}^{-2}) for (d). E_{C}=e^{2}/(4\pi\epsilon\epsilon_{0}\ell) is the Coulomb energy scale where \ell=\hbar v_{F}/t_{1} is the characteristic length scale of graphene with the Fermi velocity v_{F} and t_{1} is inter-layer opposite sublattice hopping strength [ 1 ] . The parallelogram indicates the simulation supercell. Calculations are performed with relative dielectric constant \epsilon=5 . Top panel in each plots shows the non-interaction dispersion: \epsilon({\boldsymbol{k}})=-\alpha|{\boldsymbol{k}}|^{2}+\gamma|{\boldsymbol{k}}|^{4} measured relative to the band minimum. The dashed line denotes the Fermi level. Neural Network VMC — In recent years, neural network variational Monte Carlo (NN-VMC) techniques have become an extremely powerful tool for strongly correlated quantum matter [ 5 , 6 , 38 , 43 ] , enabling accurate variational studies of continuum electron gases [ 7 , 42 , 47 , 53 , 20 , 19 ] , superconducting states [ 33 ] , fractionalized phases [ 51 , 44 , 39 , 15 , 2 , 60 ] , and other interacting many-body problems [ 30 , 17 , 35 , 3 , 18 ] . In particular, the ability of self-attention architectures [ 54 , 22 ] to represent diverse quantum phases within a single expressive ansatz makes them well suited for rhombohedral graphene, where metallic, crystalline, and possibly other competing phases appear in close proximity. Our self-attention–based wavefunction consists of a sum of N_{\rm det} determinants of many-body orbitals, and is given by \Psi_{\{\theta\}}(\{{\boldsymbol{r}}_{i}\})=\sum_{n=1}^{N_{\rm det}}{\rm det}[\phi_{j}^{(n)}({\boldsymbol{r}}_{i};\{{\boldsymbol{r}}_{\neq i}\})] (2) where \phi_{j}^{n}({\boldsymbol{r}}_{i};\{{\boldsymbol{r}}_{\neq i}\}) is a generalized many-body orbital for the i -th particle, that also depends on the coordinates of the remaining particles {\boldsymbol{r}}_{\neq i} . \{\theta\} denotes the variational parameters of the ansatz. The many-body orbitals are output entirely from a self-attention deep neural network [ 54 , 22 ] . The neural network parameters \{\theta\} are optimized by minimizing the total energy E=\langle\Psi|H|\Psi\rangle/\langle\Psi|\Psi\rangle . For the NN-VMC, we focus on the role of the conduction band dispersion on interaction physics using a simplified dispersion: \varepsilon({\boldsymbol{k}})=-\alpha|{\boldsymbol{k}}|^{2}+\gamma|{\boldsymbol{k}}|^{4} , which captures the transition from single pocket ( \alpha<0 ) to Mexican hat ( \alpha>0 ) regime. We stress that our NN architecture itself is completely system agnostic: the only input information about rhombohedral graphene lies in the loss function, which is the variational energy. Details of NN-VMC implementation can be found in the Supplementary Material. Our NN-VMC discovers four distinct types of crystalline states (Fig. 2 ) in different regions of the (n,D) phase diagram, where the band dispersion is Mexican hat and the noninteracting Fermi sea is annular. Particularly interesting are two unconventional quantum crystals that occur at relatively high electron densities. First, at lower displacement fields where the Mexican hat is more shallow, i.e., smaller values k_{0}=\sqrt{\alpha/2\gamma} , we find a triangular lattice of density minima (holes) in an otherwise nearly uniform electronic background, as shown in Fig. 2 (a). The period of this crystal is set by the density of holes, which occupy the top of the Mexican hat above the Fermi level. We refer to this crystalline state as “anticrystal”. The origin of the anticrystal phase can be qualitatively understood by considering the non-interacting dispersion and Fermi surfaces, shown in Fig. 1 (a) and (b). Away from the band bottom at |{\boldsymbol{k}}|=k_{0} , the kinetic energy changes slowly around {\boldsymbol{k}}=0 , while increasing rapidly at large momenta. As a result, the electron- and hole pockets have very different masses such that, for sufficiently high electron densities, the heavier hole mass favors crystallization, while the lighter electrons prefer a liquid state. At higher displacement fields, i.e., in the deep Mexican hat regime, we instead find a self-doped crystal (Fig. 2 (b)) in which the electron density is periodically modulated, forming a triangular crystal. However, this crystalline state contains more density peaks than the number of particles, for example 31 peaks for N=25 particles. This means that the emergent crystal unit cell contains less than one electron on average, distinguishing this state from a conventional Wigner crystal that has one localized electron per unit cell. Since the number of density maxima does not match the number of electrons, we refer to this state as a “phantom” crystal. The appearance of the phantom crystal is also tied to the Mexican-hat dispersion: in addition to the mean interparticle spacing set by the density, the annular Fermi sea introduces another length scale set by k_{0}^{-1} . As a result, the dominant charge-ordering wave vector minimizing the total energy is set by the interplay of these two length scales. In our NN-VMC calculations, we find the magnitude of the charge-ordering wave vector Q to compromise between the conventional Wigner-crystal wave vector set by the particle density and the characteristic wave vectors of the Mexican-hat dispersions that produce prominent features in the non-interacting static density-density susceptibility [ 1 ] . This is consistent with the fact that the self-doped crystal appears in an intermediate-coupling regime at high electron density, where neither weak-coupling fermiology nor strong-coupling arguments alone are quantitatively accurate. Moving to lower densities, we encounter two more types of electron crystals. The first, realized in the shallow Mexican hat regime (smaller D ), is the ordinary triangular Wigner crystal (WC), shown in Fig. 2 (c), with one electron per unit cell. Moving to deeper Mexican hat dispersions (higher D ), this state transforms into a nodal Wigner crystal (Fig. 2 (d)). The nodal WC is characterized by a momentum space occupation n({\boldsymbol{k}}) that is concentrated around the non-interacting annular Fermi sea centered at |{\boldsymbol{k}}|\approx k_{0} [ 1 ] . In real space, the amplitude of the wavefunction is depleted along a ring centered around the maxima at each triangular lattice site. The peculiar shape of nodal crystal orbitals can be quantitatively understood by considering an effective single-particle problem with Mexican-hat dispersion, in which the particle moves in an approximately harmonic potential generated by interactions with neighboring localized electrons [ 29 , 28 , 1 ] . When the momentum scale k_{0} is large enough, the electron wavepacket is then built from plane waves with |{\boldsymbol{k}}|\approx k_{0} . The resulting ground state wavefunction consists of a Gaussian factor, typical of the harmonic oscillator ground state, times an oscillatory Bessel function that first vanishes along a ring with radius \simeq 2.4048/k_{0} , in agreement with our numerical findings (cf SM [ 1 ] ). The phantom crystal obtained in our calculations is consistent with recent Hartree-Fock results [ 10 , 16 ] . Additionally, our calculations predict the presence of an anticrystal at higher densities. Figure 3: Hartree Fock phase diagram for electrons with dispersion \epsilon({\boldsymbol{k}}) of the conduction band of rhombohedral tetralayer graphene, interacting via the long-range Coulomb interaction with dielectric constant \epsilon=12 . The color code indicates the filling factor \nu of the ground state crystal unit cell. The dashed line denotes the Lifshitz line where the Fermi surface topology chances. Hartree–Fock Phase Diagram — Having presented our NN-VMC results on finite-size systems, we now provide complementary evidence for the observed crystalline states within the Hartree–Fock (HF) framework. To this end, we perform self-consistent calculations that minimize the Hartree–Fock energy for a prescribed real-space pattern directly in the thermodynamic limit, while further optimizing over the unit-cell lattice vectors ({\boldsymbol{a}}_{1},{\boldsymbol{a}}_{2}) , which determine the filling factor \nu=n|{\boldsymbol{a}}_{1}\times{\boldsymbol{a}}_{2}| . In this optimization, we restrict our attention to triangular unit cells with a fixed aspect ratio of 1:1 and consider a wide range of possible filling factors \nu . Additional fully unrestricted HF calculations, performed for periodic boundary conditions at particular system sizes, are discussed in the SM [ 1 ] . The filling factor \nu of the crystal unit cell in the ground state is shown in Fig. 3 as a function of the electron density and displacement field. Notably, our Hartree-Fock calculations based on the full band structure of rhombohedral graphene—which includes trigonal warping—reproduce all the unconventional crystalline states identified with our NN-VMC simulations. In addition to the ordinary and nodal Wigner crystals, both of which have \nu=1 and therefore one electron per unit cell, the HF calculations reveal extended regions of the (n,D) plane in which the optimal filling factor deviates from unity. At larger D and intermediate densities, we find triangular phantom crystals with \nu<1 . This metallic crystalline state can be understood as arising from hole doping of the insulating Wigner crystal at \nu=1 , with a hole density equal to approximately 10\% of the commensurate electron density. The corresponding band structure and hole-like Fermi surface are shown in Fig. 4 (a). Upon further increasing n , however, we find a distinct region in which the optimal filling factor is concentrated near \nu=2 , corresponding to two electrons per crystal unit cell. This region hosts the anticrystal phase, characterized by a triangular lattice of holes. In the anticrystal regime, the system first forms an insulating Wigner crystal of holes with exact filling factor \nu=2 . Upon changing the carrier density or the displacement field, this state evolves into the metallic anticrystal. Unlike the phantom crystal, which emerges at incommensurate fillings near \nu=1 , the metallic anticrystal originates in the vicinity of an insulating commensurate crystal with two electrons per unit cell. Figure 4 (b) shows the bands of the metallic anticrystal obtained at n=0.8\times 10^{12}\,\mathrm{cm}^{-2} and D=60\,\mathrm{meV} . For these parameters, we find the metallic anticrystal optimal doping to be \nu\approx 1.9 electrons per unit cell, corresponding to a hole doping of \delta n\approx 0.05\,n=0.04\times 10^{12}\,\mathrm{cm}^{-2} relative to \nu=2 . Thus, the system is hole-doped by approximately 5\% relative to the commensurate filling. Interestingly, the hole quasiparticles form a Fermi surface around \Gamma , distorted by trigonal warping, as shown in the inset of Fig. 4 (b). Figure 4: Hartree-Fock band structure of the self-doped Wigner crystal (a) and the metallic anticrystal (b). The inset of panels (a) and (b) show the Fermi surface of hole-like quasiparticles. Panel (a) is obtained at n=0.5\times 10^{12} cm -2 and D=65 meV ( \nu\approx 0.9) , and panel (b) at n=0.8\times 10^{12} cm -2 and D=60 meV ( \nu\approx 1.9 ), with dielectric constant \epsilon=12 . Discussion — The interplay between strong electronic correlations and a tunable, nontrivial band structure makes rhombohedral graphene an exceptional platform for exploring strongly correlated quantum matter. In this work, we advance the understanding of this fascinating material by introducing the dimensionless ratio r_{\rm eff} as a useful guide for navigating its complex phase diagram. Combining state-of-the-art NN-VMC with Hartree–Fock calculations, we uncover a rich variety of crystalline states enabled by the tunable fermiology of rhombohedral graphene. In particular, the anticrystal and self-doped crystal phases emerge from the competition among the distinct microscopic length scales associated with the annular Fermi-sea regime, and could be distinguished experimentally through direct imaging using scanning tunneling microscopy [ 34 , 52 ] . Furthermore, our results are consistent with recent experimental reports of metallic crystals [ 23 , 59 ] with 15\% hole doping at electron densities n\sim 0.4\times 10^{12}\, cm -2 , which we identify as the self-doped Wigner crystal. More importantly, we highlight the resistive state observed at higher density n\sim 0.6-0.8\times 10^{12}\, cm -2 , which intersects two chiral superconductivity domes at lower and higher displacement fields respectively. Our work identifies this resistive state, which is located in a region with large r_{\rm eff} [ 21 ] , as an anticrystal. Given its close proximity to superconductivity, the interplay of crystalline order and superconductivity may hold the key for a pairing mechanism in rhombohedral graphene. An important direction for future work is to study the effects of band topology on crystallization. In the presence of finite Berry curvature, crystalline order may also coexist with both quantized or unquantized anomalous Hall response [ 46 , 11 , 13 , 58 , 53 , 48 , 50 , 9 , 56 , 49 , 12 ] . Incorporating these Berry curvature effects beyond band dispersion in the strongly correlated regime of rhombohedral graphene is an important subject for further study. Note Added — A related study of crystalline phases in rhombohedral graphene will appear in a forthcoming preprint [ 31 ] . Acknowledgments — We acknowledge helpful discussions with Long Ju, Tonghang Han, Max Geier, Margarita Davydova, Luke Kim and Andrea Young. We thank Xiaomeng Liu for interesting and fruitful discussions. This work was supported by a Simons Investigator Award from the Simons Foundation. We acknowledge MIT SuperCloud and the Lincoln Laboratory Supercomputing Center, IAIFI (the NSF Cooperative Agreement PHY-2019786) and MIT Engaging cluster for providing the computational resources that have contributed to the results reported in this paper. Supplemental materials for: “” Ahmed Abouelkomsan 1∗ , Filippo Gaggioli 1∗ , Daniele Guerci 1∗ and Liang Fu 1 1 Department of Physics, Massachusetts Institute of Technology, Cambridge, MA-02139, USA ∗ These authors contributed equally to this work. Appendix A Microscopic Model and Correlation Effects in Rhombohedral Multilayer Graphene For rhombohedral L -layer graphene, we consider the single-particle Hamiltonian: h_{L}=\begin{pmatrix}h^{(0)}_{1}&h^{(1)}&h^{(2)}&0&\cdots&0\\ h^{(1)\dagger}&h^{(0)}_{2}&h^{(1)}&h^{(2)}&\ddots&\vdots\\ h^{(2)\dagger}&h^{(1)\dagger}&h^{(0)}_{3}&h^{(1)}&\ddots&0\\ 0&h^{(2)\dagger}&h^{(1)\dagger}&h^{(0)}_{4}&\ddots&h^{(2)}\\ \vdots&\ddots&\ddots&\ddots&\ddots&h^{(1)}\\ 0&\cdots&0&h^{(2)\dagger}&h^{(1)\dagger}&h^{(0)}_{L}\end{pmatrix}, (1) where, for simplicity, we dropped the momentum dependency, u_{\ell}=D(\ell-(N+1)/2) with D displacement field and: h^{(0)}_{l}({\boldsymbol{k}})=\begin{pmatrix}u_{l}&-t_{0}f_{{\boldsymbol{k}}}\\ -t_{0}f^{*}_{{\boldsymbol{k}}}&u_{l}\end{pmatrix},\quad h^{(1)}({\boldsymbol{k}})=\begin{pmatrix}t_{4}f_{{\boldsymbol{k}}}&t_{3}f^{*}_{{\boldsymbol{k}}}\\ t_{1}&t_{4}f_{{\boldsymbol{k}}}\end{pmatrix},\quad h^{(2)}=\begin{pmatrix}0&t_{2}/2\\ 0&0\end{pmatrix},\quad f_{{\boldsymbol{k}}}=\sum^{3}_{j=1}e^{i\bm{u}_{j}\cdot{\boldsymbol{k}}}. (2) In the latter expression, we have introduced the vectors connecting the two sublattices \bm{u}_{j}=a_{G}R^{j-1}_{2\pi/3}(0,1)/\sqrt{3} , R_{2\pi/3} rotation of 2\pi/3 around z and a_{G}=0.246\, nm. The model parameters are taken from Ref. [ 11 ] : t_{0}=3.1\,{\rm eV} , t_{\perp}=t_{1}=380\,{\rm meV} , and (t_{2},t_{3},t_{4})=(-21,290,141)\,{\rm meV} , while the layer-dependent onsite potential u_{\ell} accounts for the applied displacement field. The model features a characteristic length scale \ell=\hbar v_{F}/t_{1}=a_{G}\sqrt{3}t_{0}/(2t_{1})\approx 1.738\,{\rm nm} , which we use as the unit of length throughout our calculations. A.1 A Minimal Model for Interacting Phases in Rhombohedral Graphene We investigate interaction-driven phases in rhombohedral graphene within a minimal model of spin- and valley-polarized electrons. We focus on the interplay between the band dispersion and strong electronic interactions, setting aside band topological effects to reduce the complexity of the problem. Within our model, the electrons reside in a single band with dispersion relation \epsilon({\boldsymbol{k}}) and interact through a two-body potential V({\boldsymbol{r}}) . In first quantization, the N -particle Hamiltonian reads H=\sum_{j=1}^{N}\epsilon(-i\hat{\nabla}_{j})+\sum_{i