text
stringlengths 0
643k
⌀ | meta
stringlengths 137
151
⌀ |
|---|---|
# Introduction
With the increase in world population and the advancement of technology, the need for energy resources is increasing day by day, but fossil fuel resources continue to decline. More than \(60 \%\) of the input energy of industrial processes is lost as waste heat. The thermal energy is also wasted from vehicles, electronic devices like computers and mobile phones and even the human body. If this huge amount of wasted energy is converted back into usable energy, the problems of energy crisis will be solved to some extent. Thermoelectric power generation is considered as one of the best solutions for remarkably reducing the use of fossil fuels and preventing an energy crisis. Thermoelectric power generators (TEGs) capture thermal energy from the environment and convert it directly into electricity. TEGs are reliable and maintenance-free solid-state devices that cause no noise and environmental pollution. TEGs have been studied for many years, but their use was practically very limited until the last decade. Despite the many advantages of TEG over conventional power generators, TEGs suffer from inadequate output power and low efficiency. Many studies have shown that using low-dimensional materials for thermoelectric purposes is an efficient solution to overcome the problem of low efficiency in TEGs. The Improved efficiency in low-dimensional thermoelectric modules has two main reasons. First is the existence of novel quantum effects related to the confinement of carriers. The next reason is the increase of boundaries in the system, which increases the scattering of phonons and reduces the thermal conductivity. Two-dimensional (2D) materials refer to classes of low-dimensional crystalline solids involving a single or few layers with weak inter-layer bondings and strong intra-layer interactions in a layer. 2D materials have been widely investigated because of their special structures and unique properties which are very different from their corresponding 3D materials. This sort of materials have extensive applications due to their novel transport and optical properties. Recently, the few-layer form of black phosphorus, namely phosphorene, has been exfoliated successfully, and attracted huge interest due to its outstanding optical, mechanical, electronic, and thermoelectric properties. Unlike graphene, which has a zero-energy bandgap, phosphorene is a semiconductor with a large energy bandgap that can be adjusted by defects, electric field, and strain. The energy bandgap can also be modulated by the number of layers in a stack that makes bilayer phosphorene more interesting than its monolayer . In addition to the high carrier mobility and large band-gap, bilayer phosphorene is easy to fabricate, which makes it more attractive to study. The presence of edges in the ribbons considerably affects the electronic structure of phosphorene. It has been pointed out a prominent feature of phosphorene nanoribbons with zigzag edges is the presence of quasi-flat edge modes in the middle of the bandgap. Mid-gap states have distinctive properties from the bulk band and play an important role in conduction. Each phosphorene layer grants a band of two-fold degenerate edge modes. So, two and four branches are appearing in the bandgap of energy for monolayer and bilayer zigzag phosphorene nanoribbons correspondingly. One way to improve thermoelectric properties is to filter electrons at specific energies by means of a boxcar transmission function. Previous studies have shown that such a transmission function can block electrons that are responsible for the reduction of electrical output power and then increase thermoelectric efficiency. The mid-gap edge states transmission function of some nanoribbons probably is the ideal boxcar function. There are usually two issues with using this transmission function. The first drawback is the small energy gap and the overlap of the edge states with the valence or conduction bands. The second issue for most cases is that the width of this transmission function is much larger than that which is suitable for use for thermoelectric applications. The appropriate width for the transmission function is determined in proportion to the operating temperatures of the leads. In this study, we have proposed that the application of periodic on-site potentials at the edges of zigzag bilayer phosphorene nanoribbons can tune the width of edge states' transmission function and introduce this structure as a worthy candidate for thermoelectric purposes. This article is organized as follows. In section [\[sec:model\]](#sec:model){reference-type="ref" reference="sec:model"}, we generally introduce the bilayer phosphorene nanoribbons crystal structure with inter-layer and intra-layer hopping parameters and introduce the TB Hamiltonian in order to study band structure. The rest of section [\[sec:model\]](#sec:model){reference-type="ref" reference="sec:model"} is devoted to discussion on zigzag bilayer phosphorene nanoribbons wave functions using the analogy between two-dimensional S-S-H ladder and bilayer phosphorene nanoribbons. In the following, we deal with the effect of all hopping parameters including intra-layer and inter-layer on wave function and dispersion of edge modes. In section [3](#sec:thermoelectric){reference-type="ref" reference="sec:thermoelectric"}, methods for calculating thermoelectric properties are introduced and in section [\[sec:results\]](#sec:results){reference-type="ref" reference="sec:results"}, we presente the results of numerical study of the thermoelectric properties of zigzag bilayer phosphorene in the presence of periodic on-site potentials at both edges using the Landauer-B\(\mathrm{\ddot{u}}\)ttiker formalism. Lastly, we finish the article with the conclusion in section [\[sec:Conclusion\]](#sec:Conclusion){reference-type="ref" reference="sec:Conclusion"}.
# Edge states in bilayer Phosphorene [\[sec:model\]]{#sec:model label="sec:model"}
## Bilayer phosphorene structure
Bilayer Phosphorene can be generally described as a two-dimensional material in which two puckered structure monolayer of Phosphorene are stacked in parallel. The layers of covalently bonded phosphorus atoms are stuck weakly to each other by Van der Waals interactions. Bilayer phosphorene can be considered with different stacking orders. In the A-B stacking which is considered in this study, the top layer is moved by half of a unit cell along x or y with respect to the bottom layer, figures [\[fig.bilayer\]](#fig.bilayer){reference-type="ref" reference="fig.bilayer"}(a) and (b). In this study, we focus on nanoribbons with zigzag edges. A monolayer of phosphorene includes two puckered-structure sublayers. The unit cell of the phosphorene monolayer contains four atoms and for an extra layer, four additional atoms are added to the unit cell. The tight-binding Hamiltonian model is a proper technique to investigate the qualitative electronic properties of phosphorus based systems: \[\begin{aligned} H={}&\sum_{\langle i,j\rangle \langle i^\prime,j^\prime \rangle}t^{\parallel}_\mathbf{i}(\hat{c}^{\dag A_{1(2)}(B_{1(2)})}_{i,j}\hat{c}^{B_{1(2)}(A_{1(2)})}_{i^\prime,j^\prime}\\ &+\sum_{\langle i,j \rangle \langle i^\prime,j^\prime \rangle}t^{\perp}_\mathbf{i}(\hat{c}^{\dag A_1(B_2)}_{i,j}\hat{c}^{B_2(A_1)}_{i^\prime,j^\prime}\\ &+\sum_{\langle i,j \rangle}\epsilon_{i,j}(\hat{c}^{\dag A_{1(2)}(B_{1(2)})}_{i,j}\hat{c}^{A_{1(2)}(B_{1(2)})}_{i,j}), \end{aligned} \label{HAM}\] where the first and second terms describe the intra-layer and inter-layer Hamiltonian respectively. In order to avoid confusion let us label each site by two indexes \((i, j)\) where \(i\) denotes the row number along the \(x\)-direction and \(j\) denotes the position of the site along the \(y\)-direction. In Hamiltonian of Eq. ([\[HAM\]](#HAM){reference-type="ref" reference="HAM"}), \(t_\mathbf{i}\) describe hopping parameters between the \(\langle i,j \rangle\) and \(\langle i^\prime,j^\prime \rangle\) sites which are shown in figure [\[fig.bilayer\]](#fig.bilayer){reference-type="ref" reference="fig.bilayer"}(a). \(\epsilon_{i,j}\) is the on-site energy and there are set to zero for all lattice sites when no defect and external field exist in the lattice. The band structure of a ZBPNR can be well described by considering 15 adequate tight-binding hopping integrals (Table[1](#tab.hopping){reference-type="ref" reference="tab.hopping"}), 5 inter-layer and 10 intra-layer. In what follows we consider a ZBPNR with periodic-boundary condition in the x-direction and open boundary condition in the y-direction. Therefore, we can use the wave number in the \(x\)-direction, \(k_x\), as a good quantum number to represent the band structure of the system. Figure [\[fig.bilayer\]](#fig.bilayer){reference-type="ref" reference="fig.bilayer"} (d) illustrates the energy band structure of the ZBPNR with \(N_w=50\) zigzag chains along the width of ribbon using the above-mentioned hopping parameters.
## Analogy between effective SSH ladder model and bilayer phosphorene edge states
In this section, we introduce a correspondence between an effective SSH ladder model and ZBPNR system to understand the behavior of edge states analytically. We discuss how the analogy between such systems may be used in order to find the dispersive properties of the edge bands in ZBPNRs which will be used in the following sections. Due to translational invariant along the \(x\)-direction, we can make the following Fourier transform: \[\hat{c}^\dag_{(i,j)}=\sum_{k_x}e^{ik_xi}\hat{c}^\dag_{k_x,j}\] \[\hat{c}_{(i,j)}=\sum_{k_x}e^{-ik_xi}\hat{c}_{k_x,j}\] Since our main goal is to obtain the the edge-states behavior, we restrict our focus on the hopping parameters of the problems which do not break the particle-hole symmetry of the system. In this regard, the most important hopping parameters are \(t^{\parallel}_1\) and \(t^{\parallel}_2\) and \(t^{\perp}_{11}\) and hence, we can write our effective Hamiltonian of the ZBPNR system as: \[H=H_1+H_2+H_{11},\] in which \[H_1=\sum_{<i,j>}t^{\parallel}_1\hat{c}^{\dag A}_{(i,j+1)}\hat{c}^{B}_{(i,j)}+t^{\parallel}_1\hat{c}^{\dag A}_{(i+1,j+1)}\hat{c}^{B}_{(i,j)},\] \[H_2=\sum_{<i,j>}t^{\parallel}_2\hat{c}^{\dag A}_{(i,j)}\hat{c}^{B}_{(i,j)},\] and \[H_{11}=\sum_{<i,j>}t^{\perp}_{11}\hat{c}^{\dag A_1}_{(i,j)}\hat{c}^{B_2}_{(i-1,j)}+t^{\perp}_{11}\hat{c}^{\dag A_2}_{(i,j)}\hat{c}^{B_1}_{(i,j)}.\] After performing the Fourier transformation the resulting Hamiltonian \(H(k_x)=H_1(k_x)+H_2(k_x)+H_{11}(k_x)\) will be obtained as: \[H_1(k_x)=\sum_{k_x}\sum_{j}\tau^{\parallel}_1(k_x)\hat{c}^{\dag A}_{(k_x,j+1)}\hat{c}^{B}_{(k_x,j)}+\tau^{\parallel}_1\hat{c}^{\dag A}_{(k_x,j+1)}\hat{c}^{B}_{(k_x,j)}e^{ik_x}),\] \[H_2(k_x)= \sum_{k_x}\sum_{j}\tau^{\parallel}_2(k_x)\hat{c}^{\dag A}_{(k_x,j)}\hat{c}^{B}_{(k_x,j)},\] and \[H_3(k_x)=\sum_{k_x}\sum_{j}(\tau^{\perp}_{11}(k_x)\hat{c}^{\dag A_1}_{(k_x,j)}\hat{c}^{B_2}_{(k_x,j)}e^{-ik_x}+\tau^{\perp}_{11}\hat{c}^{\dag A_2}_{(k_x,j)}\hat{c}^{B_1}_{(k_x,j)}),\] where \[\tau^{\parallel}_1(k_x)=2t^{\parallel}_1\cos(\frac{k_x}{2}),\] \[\tau^{\parallel}_2(k_x)=t^{\parallel}_2,\] and \[\tau^{\perp}_{11}(k_x)= t^{\perp}_{11}e^{(\frac{-ik_x}{2})}.\] Here we used gauge transformation\((\hat{c}^B_{k_x,j}\rightarrow \hat{c}^B_{k_x,j} e^{(\frac{-ik_x}{2})})\) and our results indicate that new hopping parameters \(\tau^{\parallel}_1(k_x)\) and \(\tau^{\perp}_{11}(k_x)\) are now momentum dependent. The most interesting property of the resulting Hamiltonian \(H(k_x)\) is its equivalence to an effective SSH ladder with the effective hopping parameters shown in Fig [\[ssh1\]](#ssh1){reference-type="ref" reference="ssh1"} (a). In order to illustrate this analogy, we have plotted the resulting band structures of both ZBPNR system and its corresponding SSH ladder in Fig. [\[ssh1\]](#ssh1){reference-type="ref" reference="ssh1"} (b). As it is shown, the resulting band structure derived from the effective SSH ladder coincides with one obtained for the ZBPNR system with width \(N_w=50\). This means that our SSH ladder can be used as an analogy to represent the edge states of ZBPNR at the middle of its energy spectrum. Here it is appropriate to mention that we found a similar analysis to introduce an effective SSH ladder for the case of graphene nanoribbons in Ref. when we were writing the paper.
## Wave functions of ZBPNR edge states
We can now proceed to obtain an analytical expression for the wave functions associated with the zero-energy state in ZBPNR. For simplicity, we generalize the notation of Refs. for the ZBPNR to show the components of the the wave functions in real space. As we know, each unit cell consists of two different atoms, denoted as \(A\) and \(B\) which can be used to label the wave function of the atom on the outermost sites by \(\psi_{n,A}\) and \(\psi_{n,A}\) where \(n\) shows the cell index on the edge of ribbon. We start our analysis by writing the Schrödinger equation as: \[\hat{H}\ket{\psi}=E\ket{\psi}\] There are \(N_w\) atoms across the width of ribbon in each layer and total wave function is labeled as \(\psi=\{\psi_{1,A_1},\psi_{1,A_2},\psi_{1,B_1},\psi_{1,B_2},\ldots\}\). We are interested in the edge states at \(E=0\) which means that we need to solve the equation \(\hat{H}\ket{\psi}=0\) with the following Hamiltonian: \[\hat{H}(k) =\begin{psmallmatrix} 0 &0& \tau^{\parallel}_1(k_x)&0&0&0&\ldots \\ 0 & 0&0&\tau^{\parallel}_1(k_x)&0&0&\ldots \\ (\tau^{\parallel}_1(k_x))^{\ast}&0&0&0&(\tau^{\parallel}_2(k_x))^{\ast}&(\tau^{\perp}_{11}(k_x))^{\ast}&\ldots\\ 0&(\tau^{\parallel}_1(k_x))^{\ast}&0&0&(\tau^{\perp}_{11}(k_x))^{\ast}&(\tau^{\parallel}_2(k_x))^{\ast}&\ldots\\ 0&0&\tau^{\parallel}_2(k_x)&\tau^{\perp}_{11}(k_x)&0&0&\ldots\\ 0&0&\tau^{\perp}_{11}(k_x)&\tau^{\parallel}_2(k_x)&0&0&\ldots\\ \vdots&\vdots&\vdots&\vdots&\vdots&\vdots& \end{psmallmatrix} \quad\] We propose the following ansatz for the edge states wave function: \[\begin{split} \ket{\psi(k_x)}_{edge}&=\sum_n(\psi_{(n,A_1)(k_x)}\ket{n,A_1}+\psi_{(n,B_1)(k_x)}\ket{n,B_1}\\ &+\psi_{(n,A_2)(k_x)}\ket{n,A_2}+\psi_{(n,B_2)(k_x)}\ket{n,B_2}), \end{split}\] and try to solve the eigenvalue problem using the recursive procedure mentioned in Ref. . It is clear that the amplitudes of the wave functions which are localized at the ribbon's top edge are zero on even sites and similar conditions hold for the bottom edge wave functions and odd sites yielding two sets of answers. In order to proceed further we need to consider the following two different conditions. If the amplitudes of wave function are equal on both layers, \(\psi_{n,A_1,k_x}=\psi_{n,A_2,k_x}\), the solution is symmetric and to that end: \[\psi^+_{(n,A_{1},k_x)}=(\frac{-\tau^{\parallel}_1(k_x)}{\tau^{\parallel}_{2}(k_x)+(\tau^{\perp}_{11}(k_x))^{\ast}})^n\psi^+_{(0,A_{1},k_x)}\] By the same token, the solution is asymmetric if the amplitudes of the wave function on different layers is opposite to each other, \(\psi_{n,A_1,k_x}=-\psi_{n,A_2,k_x}\), and therefore: \[\psi^-_{(n,A_{1},k_x)}=(\frac{-\tau^{\parallel}_1(k_x)}{\tau^{\parallel}_{2}(k_x)-(\tau^{\perp}_{11}(k_x))^{\ast}})^n \psi^-_{(0,A_{1},k_x)}.\] Now, we expect the \(\ket{\psi^{+(-)}{(k_x)}}_{(edge)}\) to be normalized. This implies that\(\langle{\psi^{+(-)}}\ket{\psi^{+(-)}}_{edge}=1\).
## The effect of other hopping parameters on wave function and dispersion of edge-bands
The resulting edge bands obtained in the previous sub-section is not dispersive. However, in this section we consider the effect of other hopping parameters which we ignored before and discuss their effects on the dispersion of our resulting edge bands. Therefore this section starts out with a quick look at Table[1](#tab.hopping){reference-type="ref" reference="tab.hopping"}, which shows that except for the main hopping parameters (represented in figure [\[ssh1\]](#ssh1){reference-type="ref" reference="ssh1"} (a)), there are some considerable hopping parameters that may have an influence on the wave function and dispersion of the edge band. Such important hopping parameters are shown in figure [\[ssh2\]](#ssh2){reference-type="ref" reference="ssh2"} (a) by dash/dot lines. We first, turn our attention to the hopping integral \(t^{\parallel}_6\) which has a relatively larget value among the intra-layer hopping parameters. This hopping integral connects inequivalent atoms at the upper and lower layers and can not change the dispersion of the edge bands. This is because as we have already discussed, the amplitudes of wave function is zero on even sites. However, it is interesting to calculate its effect on the analytic form of the wave function. In order to calculate an analytic form of the wave funcion in presence of \(t^{\parallel}_6\), one needs to obtain the effective momentum dependent parameter \((\tau^{\parallel}_6(k_x)= t^{\parallel}_6exp^{-ik_x/2})\) and then by solving the same Schrodinger equation for the edge states can obtain the following symmetric and asymmetric solutions: \[\psi^+_{(1,A_1,k_x)}=\frac{-\tau^{\parallel}_1(k_x)}{\tau^{\parallel}_2(k_x)+(\tau^{\perp}_{11}(k_x))^{\ast}}\psi^+_{(0,A_1,k_x)}\] \[\begin{split} \psi^+_{(n,A_1,k_x)}&=\frac{(-\tau^{\parallel}_6(k_x))^{\ast}}{\tau^{\parallel}_2(k_x)+(\tau^{\perp}_{11}(k_x))^{\ast}}\psi^+_{(n-2,A_1,k_x)}\\ &+\frac{-\tau^{\parallel}_1(k_x)}{\tau^{\parallel}_2(k_x)+(\tau^{\perp}_{11}(k_x))^{\ast}}\psi^+_{(n-1,A_1,k_x)} \end{split}\] \[\psi^-_{(1,A_1,k_x)}=\frac{-\tau^{\parallel}_1(k_x)}{\tau^{\parallel}_2(k)-(\tau^{\perp}_{11}(k_x))^{\ast}}\psi^-_{(0,A_1,k_x)}\] \[\begin{split} \psi^-_{(n,A_1,k_x)}&=\frac{(-\tau^{\parallel}_6(k_x))^{\ast}}{\tau^{\parallel}_2(k)-(\tau^{\perp}_{11}(k_x))^{\ast}}\psi^-_{(n-2,A_1,k_x)}\\ &+\frac{-\tau^{\parallel}_1(k_x)}{\tau^{\parallel}_2(k_x)-(\tau^{\perp}_{11}(k_x))^{\ast}}\psi^-_{(n-1,A_1,k_x)}. \end{split}\] Another hopping parameters which seems to be significant is \(t^{\perp}_{14}\). It does not eliminate the degeneracy of edge bands and the effect that the mentioned parameter has on the deformation of the wave function is exactly the same as the previous recurrence equation considering that \(\tau^{\perp}_{14}(k_x)=t^{\perp}_{14}e^{(-ik_x/2)}\). \(t^{\parallel}_8\) is another relevant hopping parameter that we would discuss. This is a hopping parameter that connects inequivalent sub-lattices from the upper layer of one chain to an atom in lower layer at a neighboring chain. The presence of this parameter, similar to \(t^{\parallel}_6\) and \(t^{\perp}_{14}\), does not remove the degeneracy of edge bands and consequently only affects the \(t^{\parallel}_1\) parameter by a small phase change. The last hopping parameter which we consider here is \(t^{\parallel}_3\). It is an important parameter since it is responsible for the dispersion of edge bands. This parameter connects the nearest neighbor sites of a pair of zigzag chains in the lower or upper layers. To understand its effect on the edge band dispersion, we need to perform the same analysis as before which results in the hopping form \(\tau^{\parallel}_3(k_x)= 2t^{\parallel}_3\cos(k_x)\). On the other hand, the presence of \(t^{\parallel}_3\) shifts the edge states spectrum with \(2t^{\parallel}_3\cos(k_x)\) but the electron-hole symmetry around fermi level\((E=0)\) will n't be broken as well as the degeneracy of edge bands. The results which is plotted in figure [\[ssh2\]](#ssh2){reference-type="ref" reference="ssh2"}(b) indicates that the flatness of the edge bands will be completely removed by the presence of \(t_3\).
## Degeneracy removal using the perturbative analysis
In the remainder of this section we will present the derivation of removing the degeneracy of the edge bands analytically via perturbation theory. Our investigation indicates that using the above-mentioned wave functions in the standard first-order perturbation theory and taking the expectation value of the most relevant parameters like \(t^{\perp}_{12}\) and \(t^{\perp}_{13}\), we can estimate the dispersion energy of the edge bands. In order to understand its ladder analogy we have presented the parameter \(t^{\perp}_{12}\) in figure [\[ft12\]](#ft12){reference-type="ref" reference="ft12"}. Each layer of zigzag phosphorene nanoribbon contributes a band of two-fold degenerate mid-gap modes, making four mid-gap modes for ZBPNRs. In what follows we employ the degenerate perturbation theory to obtain the dispersion of edge bands for ZBPNRs. We consider the Hamiltonian \(H\) in such a way that it can be split into two parts, namely \(\hat{H}=\hat{H}_0+\hat{H^{\prime}}\), where \(\hat{H^{\prime}}\) is known as the perturbation. Thus, we can write \[\begin{aligned}
\hat{H^{\prime}}=\begin{pmatrix} E_{11}&E_{12}\\ E_{21}&E_{22} \end{pmatrix} \end{aligned}\] and the perturbative Hamiltonian matrix elements are determined by: \[\begin{aligned}
\begin{cases} E_{11}=\bra{\psi^+}\hat{H^{\prime}}\ket{\psi^+}\\ E_{12}=\bra{\psi^+}\hat{H^{\prime}}\ket{\psi^-}\\ E_{21}=\bra{\psi^-}\hat{H^{\prime}}\ket{\psi^+}\\ E_{22}=\bra{\psi^-}\hat{H^{\prime}}\ket{\psi^-}. \end{cases} \end{aligned}\] Then, the eigenvalues in the presence of the hopping \(t^{\perp}_{12}\) and \(t^{\perp}_{13}\) will be given by the following equation: \[\begin{aligned}
\begin{vmatrix} E-\bra {\psi^+} \hat{H^{\prime}}\ket{\psi^+}& \bra {\psi^+} \hat{H^{\prime}} \ket{\psi^-}\\ \bra {\psi^-} \hat{H^{\prime}} \ket{\psi^+}& E-\bra {\psi^-} \hat{H^{\prime}} \ket{\psi^-} \end{vmatrix}=0 \quad \end{aligned}\] Applying the same procedure we used in the previous section and considering the gauge transformation, we try to write the perturbative Hamiltonian as: \[\begin{aligned} \hat{H^{\prime}}_{12}(k_x)={}& \sum_{k_x}\sum_{j} (t^{\perp}_{12})\{\hat{c}^{\dag A_1}_{(k_x,j+1)}\hat{c}^{A_2}_{(k_x,j+2)}\\ &+\hat{c}^{\dag A_2}_{(k_x,j)}\hat{c}^{A_1}_{(k_x,j+1)}+\hat{c}^{\dag B_1}_{(k_x,j)}\hat{c}^{B_2}_{(k_x,j+1)}e^{(i\frac{k_x}{2})}\\ &+\hat{c}^{\dag B_2}_{(k_x,j-1)}\hat{c}^{B_1}_{(k_x,j)}e^{(-i\frac{k_x}{2})}\} \end{aligned}\] \[\begin{aligned} \hat{H^{\prime}}_{13}(k_x)={}& \sum_{k_x}\sum_{j} (t^{\perp}_{13})2\cos(k)\{\hat{c}^{\dag A_2}_{(k_x,j)}\hat{c}^{A_1}_{(k_x,j+1)}\\ &+\hat{c}^{\dag A_1}_{(k_x,j+1)}\hat{c}^{A_2}_{(k_x,j)}+\hat{c}^{\dag B_2}_{(k_x,j-1)}\hat{c}^{B_1}_{(k_x,j-2)}\\ &+\hat{c}^{\dag B_1}_{(k_x,j-2)}\hat{c}^{B_2}_{(k_x,j-1)}\} \end{aligned}\] This equation can be solved and the resulting eigen energies are the following: \[E^{(1)}_{12}=t^{\perp}_{12}(\sum_j \psi^\ast_{(j+1,A_1)}\psi_{(j+2,A_2)})+t^{\perp}_{12}(\sum_j\psi^\ast_{(j,A_2)}\psi_{(j+1,A_1)})\] and so considering the \(\frac{-\tau^{\parallel}_{1}(k_x)}{\tau^{\parallel}_{2}(k_x)+(\tau^{\perp}_{11}(k_x))^{\ast}}=x\) and \(\frac{-\tau^{\parallel}_{1}(k_x)}{\tau^{\parallel}_{2}(k_x)-(\tau^{\perp}_{11}(k_x))^{\ast}}=y\), matrix elements of perturbation hamiltonian consist of: \[E_{11}=t^{\perp}_{12}(x)(\frac{1+x^2}{1-x^2})\] \[E_{22}=t^{\perp}_{12}(y)(\frac{1+y^2}{1-y^2})\] \[E_{12}=t^{\perp}_{12}(\frac{y}{(1-x^{\ast})(1-y)})(x^{\ast}y+1)\] and about another hopping parameter: \[\begin{split} E^{(1)}_{13}&=t^{\perp}_{13}(2\cos(k_x))(\sum_j \psi^\ast_{(j,A_2)}\psi_{(j+1,A_1)})\\ &+t^{\perp}_{13}(2\cos(k_x))(\sum_j\psi^\ast_{(j+1,A_1)}\psi_{(j,A_2)}) \end{split}\] \[E_{11}=2t^{\perp}_{13}\cos(k_x)(\frac{x+(x^\ast)}{1-x^2})\] \[E_{22}=2t^{\perp}_{13}\cos(k_x)(\frac{y+(y^\ast)}{1-y^2})\] \[E_{12}=2t^{\perp}_{13}\cos(k_x)(\frac{1}{(1-x^{\ast})(1-y)})(x^{\ast}+y)\] This result is consistent with our numerical calculations and our conclusion is that hopping parameters \(t^{\perp}_{12}\) and \(t^{\perp}_{13}\) are the responsible parameters for the removal of the degeneracy between energy bands associated with each layer.
# Thermoelectric properties {#sec:thermoelectric}
In the thermoelectric power generation process, electrical power (output) is generated in cost of losing heat from hot lead (input). Therefore, it is desirable to produce the maximum amount of electrical power with the least heat flow. The efficiency of TEGs is then defined as: \[\label{eq:cop} \eta^{TEG}(\%) = \frac{P^{el}}{Q_{hot}}=\frac{VI}{Q_{hot}} \times100 \%,\] where \(I\) and \(Q_{hot}\) are electric and heat currents and can be obtained using the Landauer-B\(\mathrm{\ddot{u}}\)ttiker formalism and \(V\) denotes the bias voltage. The electric current in the ballistic transport regime can be obtained from the following equation: \[\label{eq:current} I = \frac{2 q_e}{h} \int dE \; T_{HC}(E) \left( f_H-f_C\right),\] where \(q_e\), \(h\) and, \(T_{HC}(E)\) denote the elementary charge, the Planck constant and, total transmission function, respectively. Also, \(f_{H}\) and \(f_{C}\) are the Fermi-Dirac distribution functions at hot and cold leads. The heat current has two electron and lattice parts. In order to calculate the electron part of the heat current, \(e\) should be replaced with the energy which each electron carries. \[\label{eq:heatcurrent} Q_{hot} = \frac{2}{h} \int dE \; T_{LR}(E) (E-\mu_h) \left( f_L-f_R\right).\] Many studies have shown that the lattice part of thermal conductivity of low-diemensional materials is much less than their corresponding bulk structures which is mainly due to increased scattering of phonons at the boundaries. Since the reduction of thermal conductivity leads to an increase in thermoelectric efficiency, this is the basis for the idea of using low-dimensional materials for thermoelectric applications. To further increase the efficiency, several methods have been proposed that increase the phonon scattering while having no significant destructive effects on the electronic properties. The use of superlattices and heterojunctions and introducing strain and defects are the most relevant approaches to thermal conductivity reduction. Optimizing the electronic properties of the system is another approach to enhance electric power and thermoelectric efficiency. As mentioned earlier, the energy filtration method by means of a boxcar transmission function leads to optimal output power and efficiency. A boxcar transmission function can eliminate the destructive current of electrons that flow from cold to hot lead and cause the enhancement of the output power. We select zigzag bilayer phosphorene nanoribbons for this study for two reasons. The first is that among two-dimensional structures it has a small thermal conductivity, and the second is the presence of mid-gap edge modes, which provide a desired boxcar transmission function. The width of edge states transmission function in pristine bilayer zigzag phosphorene nanoribbons is very large and not suitable for room temperature thermoelectric applications. To regulate the width of this function, on-site potentials have been applied periodically at the two edges of the nanoribbon. The existence of on-site potentials also increases the phonon scattering and significantly reduces the lattice thermal conductivity of nanoribbon. In this study and for the cases of dense on-site potentials, we neglected the lattice part of thermal conductivity. The transmission function can be calculated using the Landauer-B\(\mathrm{\ddot{u}}\)ttiker formalism: \[T_{LR}(\epsilon)=Tr(\Gamma_{L}G^{ret}\Gamma_{R}{G^{ret}}^\dag).\] Here, \(\Gamma_{L(R)}(\epsilon)=i[\Sigma_{L(R)}(\epsilon)-\Sigma_{L(R)}^\dagger(\epsilon)]\) is the coupling between device and left(right) lead and \(G^{ret}\) is the retarded Green's functions: \[G^{ret}(\epsilon)=\frac{1}{\epsilon-H_{center}-\Sigma_L{(\epsilon)}-\Sigma_R{(\epsilon)}}\] where, \(H_{center}\) is the device Hamiltonian and \(\Sigma_{R}\) ad \(\Sigma_{L}\) are the coupling to the leads retarded self energies \[\Sigma_{L}(\epsilon)=\tau_{LD}^\dagger~G^{ret}_{L}~\tau_{LD},\\ \Sigma_{R}(\epsilon)=\tau_{DR}~G^{ret}_{R}~\tau_{DR}^\dagger,\] in which \(\tau_{DR}\) and \(\tau_{LD}\) are the coupling Matrices between device and right and left leads. Also, \(G^{ret}_{R(L)}=[\epsilon-H_{L(R)}]^{-1}\) is the right(left) lead retarded Green's function and, \(H_{L}\) and \(H_{R}\) are the left and right leads Hamiltonian. The second law of thermodynamics sets a maximum to the heat engine's efficiency operating at two temperatures. This ideal limit which is called Carnot efficiency, depends on the temperatures of the cold and hot leads and is given by: \[\label{eq:Carnot} \eta_{Carnot}=1-\left(\frac{T_c}{T_h} \right)\] This ideal efficiency is obtained when electrons of only one certain energy are transported, which means \(T=\delta(E-E_0)\). Although very narrow transmission functions lead to high efficiencies, they provide negligible electrical power output. The use of the boxcar transmission function is the solution to optimize both electric power output and efficiency. Such a boxcar transmission function blocks all electrons except low-energy electrons close to Fermi energy, and since electron-phonon scattering is significant for high-energy electrons, we omit this electron-phonon interaction in this study. Also, due to the presence of many impurities in the system, the phonon thermal conductivity will be significantly reduced. In this study, we have omitted the heat flow due to phonon thermal conductivity, but it can later be added to the calculations as a disturbance
# Results [\[sec:results\]]{#sec:results label="sec:results"}
In the nonlinear calculations, the hot and cold lead temperatures are assumed to be \(T_h = 400~\mathrm{K}\) and \(T_c = 300~\mathrm{K}\), so the ideal Carnot efficiency at these temperatures is \(25 \%\). As shown in figure [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}, periodic on-site potentials with three periodicities are applied to both edges of the zigzag bilayer phosphorene nanoribbons. In the first case labeled with m = 2, the on-site potentials are applied to two adjacent atoms, one atom from the top layer and another atom from the bottom layer of both edges of ZBPNRs while the next two atoms have no on-site potentials. In the second and third cases with m=3 and m=4, after two atoms with on-site potentials, the next 4 and 6 atoms remain intact, respectively. Figures [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(a), [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(d), and [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(g) show the energy bandstructure of ZBPNRs for the cases of m=2, 3 and 4, respectively. It can be seen that there are a number of energy bands corresponding to the edge states in the energy bandgap. Figures [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(b), [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(e), and [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(h) illustrate a zoom of energy bands and figures [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(c), [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(f), and [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(i) show their corresponding transmission function close to the edge states. We see that changing the spacing between the on-site potentials results in a shift in the position of the transmission peaks. At larger distances between the on-site potentials, the transmission peaks come closer to each other. This is similar to the situation that occurs in continuous systems such as rectangular potential barriers and makes sense because changing the distance between the on-site potentials leads to a change in resonance between their wave function. So, more transmission peaks can be observed in the bandgap for cases with larger spacing between on-site potentials. Also, increasing the spacing between the on-site potentials reduces the energy bands dispersion and ultimately leads to narrower boxcar transmission functions. Narrowing of the transmission windows by increasing the spacing between on-site potentials is caused by decreasing the overlap of their localized states. Figures [\[fig:Pandeff\]](#fig:Pandeff){reference-type="ref" reference="fig:Pandeff"}(a), [\[fig:Pandeff\]](#fig:Pandeff){reference-type="ref" reference="fig:Pandeff"}(c), and [\[fig:Pandeff\]](#fig:Pandeff){reference-type="ref" reference="fig:Pandeff"}(e) show the electric power and figures [\[fig:Pandeff\]](#fig:Pandeff){reference-type="ref" reference="fig:Pandeff"}(b), [\[fig:Pandeff\]](#fig:Pandeff){reference-type="ref" reference="fig:Pandeff"}(d), and [\[fig:Pandeff\]](#fig:Pandeff){reference-type="ref" reference="fig:Pandeff"}(f) demonstrate the thermoelectric efficiency, using the transmission function of figures [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(c), [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(f), and [\[fig:bands\]](#fig:bands){reference-type="ref" reference="fig:bands"}(j), plotted on color scale as functions of bias voltage \(V\) and the average chemical potential \(\mu\). We can attain optimal electric power and thermoelectric efficiency by properly regulating bias voltage and average chemical potential. As indicated in figure [\[fig:Pandeff\]](#fig:Pandeff){reference-type="ref" reference="fig:Pandeff"}, the maximum thermoelectric efficiency is larger for cases with larger spacing between on-site potentials which have a narrower transmission functions. These results show that the output power at the maximum efficiency point is much smaller than the maximum power, so the maximum power point is probably a better choice than the maximum efficiency point. The highest maximum electric power is obtained for the case of \(m=2\), which has the widest transmission function, \(P^{m=2}_{max}=4550~\mathrm{pW}\) and for the other cases \(P^{m=3}_{max}=4450~\mathrm{pW}\) and \(P^{m=4}_{max}=2930~\mathrm{pW}\). In contrast to maximum electric power, the highest efficiency at maximum electric power is is achieved for the case of m=4 which has the narrowest transmission function, \(\eta^{m=4}_{maxP}=12.7\%\) and for the other cases \(\eta^{m=3}_{maxP}=11.0\%\) and \(\eta^{m=2}_{maxP}=8.6\%\). Considering the values obtained for maximum electric power and its corresponding efficiency, it seems that the optimal electric power and thermoelectric efficiency for the considered temperatures are achieved for the case of m=3, whose maximum electric power is very close to the case m=2, but with higher efficiency.\
# Conclusion [\[sec:Conclusion\]]{#sec:Conclusion label="sec:Conclusion"}
In conclusion, we have investigated the thermoelectric power generation properties of zigzag bilayer phosphorene nanoribbons in the presence of periodic on-site potentials at the edges and have calculated the generated electric power and thermoelectric efficiency. Since the width of the edge state transmission function in pristine ZBPNRs is very large, we don't expect good thermoelectric efficiency at room temperature for this structure. By periodically applying on-site potentials, we have obtained narrow transmission functions that are suitable for room temperature thermoelectric applications. Our results show that excellent electric power can be achieved at high efficiencies by edge states engineering. | {'timestamp': '2022-03-03T02:28:21', 'yymm': '2203', 'arxiv_id': '2203.01094', 'language': 'en', 'url': 'https://arxiv.org/abs/2203.01094'} |
# Introduction
recent years, the global demand for video streaming services has increased rapidly. Nowadays, video data constitutes more than \(50\%\) of the mobile data traffic. In order to optimally exploit the available wireless channel capacity, all video coding standards such as Advanced Video Coding (H.264/AVC) or the most recent High Efficiency Video Coding (HEVC) standard primarily aim at reducing the bitrate while maintaining the objective visual quality of the sequence. Simultaneously, a dramatic increase in encoding and decoding complexity due to more sophisticated and efficient compression algorithms has been observed. An analysis of the power consumption of the decoding process for a H.264 software decoder on a smartphone was performed in and showed that the major part of the device's power consumption is used for processing on the central processing unit (CPU). Hence, research aiming at reducing the decoder's processing power is of great interest. To this end, we propose modifying the classic rate-distortion optimization (RDO) approach used in many encoder solutions such as the HEVC Test Model (HM) reference software for the HEVC codec or the JM reference software for the H.264/AVC codec. The goal of the state-of-the-art RDO is to minimize the distortion \(D\) assuming the constraint that the rate \(R\) does not exceed the maximum value \(R_\mathrm{max}\). Mathematically, this optimization problem can be formulated as \[\begin{aligned}
\min D \quad \mathrm{s.t.} \quad R \le R_\mathrm{max}. \label{eq:RD_base} \end{aligned}\] As this representation is inconvenient for practical use, an equivalent formula was derived in using a Lagrange multiplier \(\lambda\). The formula then reads \[\min J = D+\lambda R, \label{eq:RD_lagr}\] where \(J\) represents the costs to be minimized. Here, \(\lambda\) can be interpreted as the trade-off parameter between distortion and rate. In a typical video codec implementation such as JM or HM, the cost \(J\) is calculated for each coding mode considered and the mode exhibiting the lowest cost is chosen. In this paper, we propose extending [\[eq:RD_lagr\]](#eq:RD_lagr){reference-type="eqref" reference="eq:RD_lagr"} to decoding-energy-rate-distortion optimization (DERDO) and propose including the decoding energy \(E\) into the formula as \[\min J = D+\lambda_R\cdot R + \lambda_E \cdot E, \label{eq:RD_lagrE}\] where we incorporate a second Lagrange multiplier \(\lambda_E\) that, in conjunction with \(\lambda_R\), determines the trade-off between distortion, rate, and decoding energy. Using this optimization function, the encoder is able to construct bit streams that
- comply with the standard and
- require less processing energy during the decoding process.
Hence, using this approach, the decoding energy can be minimized without having to reconfigure the software or the hardware of the decoding device. The only prerequisite is that the processing energy depends on the use of coding tools, which is given for any software decoder. Please note that in this work, evaluations are only performed for software decoders showing that DERDO is applicable in practice. For hardware decoders, this approach is potentially applicable if, e.g., various components like an in-loop filtering stage can be switched on and off during decoder runtime. Such a decoder is, e.g., the hardware chip proposed by Engelhardt et al.. For the HEVC intraframe coding case using a software decoder, it was shown that around \(15\%\) of decoding energy could be saved when accepting a bitrate increase of around \(15\%\). The drawback of a higher bitrate can be accepted when the available bandwidth is sufficient, which is true in many real applications. For example, our tests showed that a class C sequence (\(832\times 480\) pixels) encoded with HEVC at a medium quality (quantization parameter (QP) equal to \(30\)) requires a bandwidth of about \(2.4\) Mbit/s which is much lower than the typical bandwidth provided by modern WiFi standards such as 802.11n, which increases to nearly \(100\) Mbit/s. A major challenge imposed by this approach is determining the decoding energy \(E\) during runtime in the encoder. As a solution, we propose making use of a feature based energy model as presented in where it was shown that such an energy model could be constructed for any hybrid video codec. In this model, each feature can be described by a certain standard process in the decoder, e.g., a feature can correspond to the residual transformation of a certain block size. As the corresponding reconstruction process of this block has a defined and homogeneous processing flow with a rather constant amount of processing energy for each execution, the corresponding processing energy can be derived using a dedicated measurement and training method. The obtained information can then be used inside the encoder to decide on the coding mode that requires a small amount of energy. The paper is organized as follows: First, Section [2](#sec:lit){reference-type="ref" reference="sec:lit"} reviews relevant literature and puts our approach into the scientific context. Afterwards, Section [3](#sec:model){reference-type="ref" reference="sec:model"} introduces the various aspects of energy consumption during video decoding and presents respective energy models. Subsequently, Section [4](#sec:DERDO){reference-type="ref" reference="sec:DERDO"} discusses the theory behind DERDO and explains how DERDO can be included in a given encoder. Afterwards, Section [5](#sec:eval){reference-type="ref" reference="sec:eval"} applies DERDO to the HEVC encoder and discusses its theoretical and actual performance, as obtained by real measurements. Finally, conclusions are given in Section [6](#sec:concl){reference-type="ref" reference="sec:concl"}.
# Literature Review {#sec:lit}
A great deal of scientific contributions can be found regarding energy efficient video coding. A major subset of these works aims at optimizing implementations in hardware or software. Examples for specialized hardware can be found for intraframe coding, transformation, arithmetic decoding, HEVC's sample adaptive offset (SAO) filter, and many more. Other works, e.g., focus on the power consumption of dynamic random access memory (DRAM) where smart data manipulation techniques can reduce the power consumption by more than \(90\%\). Furthermore, several software encoders and decoders are implemented and continuously refined for real-time streaming purposes. Another branch focuses on the optimization of the encoder. In the commonly used rate-distortion (RD) approach, nearly all possible coding modes are tested for rate and distortion and the best mode is chosen for coding. This brute-force search shows a very high computational complexity such that current research tries to enable real-time encoding while losing little performance in the RD sense. For H.264/AVC, such an approach is proposed in where a corresponding complexity metric is derived for the testing of each coding mode. Then, depending on the available computational power, only a part of the possible coding modes is tested for RD costs such that the encoding time can be reduced significantly. A similar approach was presented in for a power-saving H.263 encoder. Furthermore, subsequent work included other energy reduction techniques such as Dynamic Voltage and Frequency Scaling (DVFS). On the decoder side, mainly four approaches can be found in the literature. In the first approach, similar to the approach used for the encoder, DVFS can be used for saving energy. An early work for H.263 was proposed by Pouwelse et al. in which the decoding device switches between different voltage and frequency states depending on the frame type (I, P, or PB-frame). A more sophisticated solution was presented in for a 3D Wavelet based Scalable Codec (SVC). It was proposed to transmit further side information from the encoder side using the MPEG-21 standard. In the second approach, Pakdeepaiboonpo et al. proposed optimizing the compiler for an MPEG-4 decoder and achieved energy savings of up to \(13\%\). The third approach focuses on the type of the bit stream. In that regard, Ren et al. proposed switching between different coding standards. When the battery is low, the decoder can switch from HEVC decoding to H.264/AVC decoding and extend the operating time by more than \(5\%\). In the last approach, which is similar to our proposal, the decoding power or complexity is estimated on the encoder side such that energy saving bit streams are constructed. Van der Schaar et al. profiled a decoder for the spatial-domain motion-compensated temporal filtering (SDMCTF) decoder and built a generic complexity model based on processor operations. Lee et al. constructed a similar model for H.264/AVC but focused on motion compensation. Finally, He et al. modeled motion compensation and the in-loop filter in HEVC and proposed switching to energy saving bit streams when the battery is low. In contrast to the aforementioned work, we do not base our model on complexity metrics but on actual energy measurements. Furthermore, our modeling approach is more general as we aim at modeling the complete decoding process (see Section [3](#sec:model){reference-type="ref" reference="sec:model"}). In contrast, other work only focused on selected parts of the decoder. Basically, we are capable of estimating the decoding energy for each possible coding mode with a high precision such that significant energy savings can be achieved.
# Energy Consumption of Video Decoders {#sec:model}
Determining and modeling the processing energy of a modern device is a highly complex task. To this end, Carrol et al. performed a general analysis of the power and energy consumed by a portable smartphone. The analysis was conducted for a high-end smartphone that was released in 2012. As the major setup of a portable device did not change significantly, the results are still valid for modern devices. In their analysis, a power breakdown for software decoding of a high and a low-quality H.264-coded bit stream was determined. A relevant workload was found on the following components of the smartphone: RAM, graphics processing unit (GPU), internal hardware systems, core, memory interface (MIF), display, and remaining miscellany of the system on chip (SoC). Except for the display, all of the aforementioned components are also considered in our model (including the CPU and the RAM). The display is not considered as its power consumption is mainly determined by the backlight brightness which cannot be controlled by our approach. A more detailed explanation on the measurement setup is given in Section [5.3](#secsec:meas){reference-type="ref" reference="secsec:meas"}. Note that in this work, we do not consider the transmission energy when using WiFi or a GSM network, for reasons of brevity. In, the influence of the transmission energy in a WiFi-streaming scenario for HEVC-intra coding was tested. The energy savings were found to be smaller, but only by less than \(5\%\). Since the rates and, hence, transmission energies are much lower for the general inter-coding case we address here, we can say that this simplification is valid. Further work should examine this problem more closely. We estimate the energy consumption during video decoding \(\hat E_\mathrm{dec}\) that is mainly composed of the CPU and RAM energy. Various models have been proposed to estimate this energy. For the HEVC decoder, a helpful overview is given in, where a high amount of different models was introduced and compared in terms of estimation accuracy. For our work, we concentrate on two types of models: A *high-level model* (HL) and a *bit stream feature based model* (BF). The former model provides insights into theoretical DERD considerations (see Section [4.2](#secsec:lambdaQP){reference-type="ref" reference="secsec:lambdaQP"}) while the latter is used for explicit DERDO (Section [4.4](#secsec:impl){reference-type="ref" reference="secsec:impl"}) in the encoder. In we showed that both models can be employed for various codecs such that the considerations given in this section apply generally. For the *high-level model*, the decoding energy is estimated using high-level parameters, namely the resolution in pixels per frame \(S\), the number of frames \(N\), the bit stream file size \(B\), and the fraction of intra coded frames \(p_\mathrm{I}\). \(p_\mathrm{I}\) is the share of intra coded frames of all frames (\(p_\mathrm{I}=\frac{N_\mathrm{I}}{N}\) with the number of I-frames \(N_\mathrm{I}\)). We consider a simple linear model that was originally proposed in and that targets the frame-based decoding energy. In, it was generalized to the decoding energy of a complete sequence. It is given by \[\hat E_\mathrm{dec,HL} = \left( c_1 \cdot p_\mathrm{I} \cdot \frac{B}{N\cdot S} + c_2 \cdot p_\mathrm{I} + c_3 \cdot \frac{B}{N\cdot S} + c_4\right) \cdot N \cdot S. \label{eq:Raoufi}\] In essence, this model estimates the mean per-pixel decoding energy based on the mean number of coded bytes per pixel. Using the parameter \(p_\mathrm{I}\), the difference between the decoding energy for inter and intra coded frames can be considered. The parameters can be interpreted as follows: The constant \(c_4\) considers a minimum energy required to decode one pixel. \(c_2\) can be interpreted as an additional energy required when the pixel is located inside an I-frame. \(c_3\) describes the additional energy depending on the bitrate: The assumption is that the energy rises linearly with the mean number of coded bits per pixel. Finally, \(c_1\) is used in a mixed term referring to both the mean bits per pixel and the fraction of I-frames. Please note that the latter two constants are used to develop the DERD theory in Section [4.2](#secsec:lambdaQP){reference-type="ref" reference="secsec:lambdaQP"}. Due to the empirical relation between the bit stream file size and the energy, incorporating the energetic properties into the analytic rate-distortion function is possible. These constants \(c_1...,c_4\) describe the energetic properties of the decoding system. For sufficiently long sequences and when the encoder settings and the type of input sequence are fixed, the estimation error is reported to be smaller than \(22\%\) for all tested codecs and decoder implementations. The *bit stream feature based model* can be described by a linear equation: \[\hat E_\mathrm{dec,BF} = \sum_{ f \in F} n_f \cdot e_f. \label{eq:BFM}\] In this model, a constant set \(F\) of bit stream features \(f\) is defined. The number of occurrences of each of these features in a coded bit stream is represented by the feature number \(n_f\). Each of these features is described by a corresponding specific energy \(e_f\) that can be interpreted as the processing energy consumed each time this feature occurs. To provide a vivid example, one feature can be the transformation of a residual block at a fixed block size. Interpreting the bit stream we can count how often this process must be executed for reconstructing the complete sequence. As the transformation itself has a rather constant processing flow for a given block size, the corresponding processing energy on the CPU is nearly constant and can be determined using a suitable training method. Generally, the same holds for the RAM energy because the amount of memory accesses (read and write) is also nearly constant for each execution. Please note that considering the RAM energy, at the beginning of the decoding process, the data is not yet available in the cache such that the RAM is accessed more often resulting in a significantly higher energy. Hence, for single frame energy estimation, the estimation errors are relatively high (more than \(15\%\) ). Furthermore, other processes that are running concurrently and that replace data in the cache can lead to energy variations and bad estimations. However, in it was shown that for different processors (ARM and Intel), the RAM energy can be estimated accurately (mean estimation error below \(5\%\)) for sufficiently long sequences (more than \(8\) frames which is given for most practical applications). Hence, the model is generally applicable for both the CPU and the RAM energy consumption of video decoders. The set of features used in this work corresponds to the simple model presented in detail in. A list summarizing these features is given in Table [1](#tab:featList){reference-type="ref" reference="tab:featList"}.
It is performed for one evaluation sequence of each class. The resulting values for QP\(^*\) are saved for the validation of [\[eq:QP_lambda_final\]](#eq:QP_lambda_final){reference-type="eqref" reference="eq:QP_lambda_final"} and for single QP optimization. Hist\((\)QP\()\) is the histogram of all QPs that were selected by the encoder. As an example, we plot Hist\((\)QP\()\) for the BQTerrace sequence in Figure [\[fig:chosenQPs\]](#fig:chosenQPs){reference-type="ref" reference="fig:chosenQPs"} for three values of \(\tau\). We can see that for all QPs and for all values of \(\tau\) distinct peaks occur. To show that [\[eq:QP_lambda_final\]](#eq:QP_lambda_final){reference-type="eqref" reference="eq:QP_lambda_final"} is a valid approximation, we collect all \(\{\lambda_\mathrm{R}, \lambda_\mathrm{E}, \mathrm{QP}^*\}\) triples from Algorithm [\[alg:multQP\]](#alg:multQP){reference-type="ref" reference="alg:multQP"} for the BQTerrace sequence and perform a least-squares fitting approach to calculate optimal values for \(\rho\) and \(\varepsilon\). The fitting algorithm is the trust region approach presented in. The experimentally derived surface as well as the fitted surface are shown in Fig. [\[fig:QP_lam_surf\]](#fig:QP_lam_surf){reference-type="ref" reference="fig:QP_lam_surf"}. First, the experimental result (blue) is considered. Please note that the surface is slightly tilted backwards showing that high \(\lambda\)s result in higher QPs. The left wing and the right wing of the surface tend to be parallel to the \(\lambda_\mathrm{R}\)-axis and the \(\lambda_\mathrm{E}\)-axis, respectively. This indicates that on these wings, the parallel axes' Lagrange multipliers only have a small influence on the optimal QP value. The fit using [\[eq:QP_lambda_final\]](#eq:QP_lambda_final){reference-type="eqref" reference="eq:QP_lambda_final"} (red) is very close to the experimental points. To evaluate the goodness of the fit, the absolute error, which is the absolute difference between QP\(^*\) and the QP calculated by [\[eq:QP_lambda_final\]](#eq:QP_lambda_final){reference-type="eqref" reference="eq:QP_lambda_final"}, is assessed. In the case of the BQTerrace sequence, the mean absolute error of the fit is \(0.73\) (maximum error \(2.11\)). Please note that the mean error is smaller than the minimum QP step of \(1\). Collecting all \(\{\lambda_\mathrm{R}, \lambda_\mathrm{E}, \mathrm{QP}^*\}\) triples for all sequences and calculating the least-squares fit, a mean absolute error of \(0.81\) and a maximum error of \(2.28\) is achieved which is only slightly higher than for the single BQTerrace sequence. This indicates that the fit is rather independent from the used input sequence. Hence, [\[eq:QP_lambda_final\]](#eq:QP_lambda_final){reference-type="eqref" reference="eq:QP_lambda_final"} is a very accurate approximation.
## Modeled Energy Savings {#secsec:theor}
In this section, the theoretical DERD-performance is analyzed. For evaluating the decoding energy, the estimated decoding energy values \(\hat E_\mathrm{dec,BF}\) as returned by model BF [\[eq:BFM\]](#eq:BFM){reference-type="eqref" reference="eq:BFM"} are used. To calculate the distortion we use the YUV-PSNR calculated by \[D=\mathrm{PSNR}_\mathrm{YUV} = \frac{1}{8}\left( 6\cdot \mathrm{PSNR}_\mathrm{Y} + \mathrm{PSNR}_\mathrm{U} + \mathrm{PSNR}_\mathrm{V} \right) \label{eq:YUV_PSNR}\] and the rate \(R\) given by the bit stream file size. We base our evaluations on the rate-distortion as well as the decoding-energy-distortion curves. An example for such curves is given in Fig. [\[fig:allRD_cactus\]](#fig:allRD_cactus){reference-type="ref" reference="fig:allRD_cactus"} for the BQTerrace sequence and multi QP optimization. The plots show the theoretical performance of our proposed method. In the plots, each curve corresponds to a constant value of \(\tau\). The left plot shows the RD performance that is commonly used in the literature. The blue line that corresponds to classic RDO (\(\lambda_\mathrm{E}=0\)) is located at the top left of all curves. We can see that increasing \(\tau\) results in RD losses, which is caused by the increased importance of the decoding energy. The right plot shows the estimated decoding energies on the x-axis that correspond to the energy values that are calculated during encoding. The y-axis is the same as in the left plot. The basic RDO (the blue line) is located the furthest to the right. We can see that increasing \(\tau\) now introduces energy savings as the curves tend to the left for higher values of \(\tau\). This general behavior can be observed for any input sequence. However, the amount of energy savings varies depending on the input sequence and the QP. As interpreting these rate-distortion and energy-distortion curves is rather complicated, we calculate mean values using the well-known Bjøntegaard Delta method with piece-wise cubic interpolation. For rate-distortion values, we use the Bjøntegaard Delta Rate (BDR) and for mean decoding-energy-distortion values, we replace the rate with the decoding energy in the Bjøntegaard-Delta calculus and obtain the Bjøntegaard Delta Decoding Energy (BDDE). This metric can be interpreted as the mean energy savings in percent at a constant objective visual quality. Hence, for each value of \(\tau\), the two metrics BDR and BDDE describe the performance of the encoder in terms of rate and decoding energy. Please note that \(\tau=0\), which is the classic RDO, is used as a reference. The BDR-BDDE-tuples can be plot as a curve as shown in Fig. [\[fig:single_multi_BD\]](#fig:single_multi_BD){reference-type="ref" reference="fig:single_multi_BD"} (BQTerrace sequence). In the following, this plot is called rate-energy diagram. At first, we only consider the blue multi QP curve. Each marker on the curve corresponds to one value of \(\tau\). We can see that increasing \(\tau\) exhibits two effects: First, the rate increases and second, the decoding energy decreases, which was expected. Interestingly, for \(\tau\) values close to \(0\) and \(1\), the curve shows a saturation: Close to these points, rate increases and decoding energy savings, respectively, are nearly constant. Furthermore, highest energy savings of almost \(50\%\) are obtained for \(\tau \approx 0.5\). Considering the red curve for single QP optimization (algorithm explained below) we can see that the performance is very similar and even better for small values of \(\tau\), i.e. the curve is slightly shifted to the left. This shows that the single QP approach can perform even better than the multi QP approach. This can be explained by the fact that no delta QPs must be coded, thus resulting in rate savings. Note that for a fair comparison, we used the multi QP sequences coded with pure RDO (\(\Delta QP=5\) and \(\tau=0\)) as a reference for both single QP and multi QP sequences to obtain the curves in Fig. [\[fig:single_multi_BD\]](#fig:single_multi_BD){reference-type="ref" reference="fig:single_multi_BD"}. Additionally, note that single QP optimization, in contrast to multi QP optimization, reflects the major advantage of a strongly reduced encoding time (on average, in our use-case, single QP optimization is ten times faster than multi QP optimization). Algorithm [\[alg:singleQP\]](#alg:singleQP){reference-type="ref" reference="alg:singleQP"} shows the single QP optimization process.
We collect all QP\(^*\) and their corresponding \(\lambda_\mathrm{R}\), \(\lambda_\mathrm{E}\) values from multi QP optimization. Then, we use the \(\{\)QP\(^*, \lambda_\mathrm{R}, \lambda_\mathrm{E}\}\) triples to train \(\rho\) and \(\varepsilon\) such that ([\[eq:QP_lambda_final\]](#eq:QP_lambda_final){reference-type="ref" reference="eq:QP_lambda_final"}) fits optimally in a least-squares sense (cf. training method in Section [5.1](#secsec:evalLam){reference-type="ref" reference="secsec:evalLam"}). To prevent training on the target sequence, its triples are discarded from the training set. Afterwards, the trained values for \(\rho\) and \(\varepsilon\) are used for encoding with a \(\Delta\)QP of \(0\). The resulting bit streams are used for measurements in the following.
## Measured Energy Savings {#secsec:meas}
In order to evaluate the measured performance of the DERDO approach, we choose the Pandaboard as a decoding device. It provides a smartphone-like architecture using an ARMv7 processor for its hardware. As the main decoding software we select the FFmpeg framework version 2.8 since it was developed for practical applications. Please note that the FFmpeg decoder makes use of dual core processing, showing that the approach is not restricted to the single core case. For comparison purposes, we also conduct tests using the libde265 (version 0.7) and the HM-13.0 decoding software. We base our evaluation of the DERD performance on real-world energy measurements since estimated energies using the models presented in Section [3](#sec:model){reference-type="ref" reference="sec:model"} may be inaccurate. To ensure that background processes do not interfere with our measurements, they are performed on runlevel 1 (where the operating system (OS) is booted with a reduced set of background processes) with disabled light emitting diodes and, as far as possible, with further supporting background services disabled. The measurement setup is explained in detail in. Generally, we measure the power consumed through the main supply jack of the board such that all modules of the board are taken into account (including CPU, RAM, and periphery). A display is not attached to the device and no wireless network connection is established. As a power meter, we choose the high precision ZES Zimmer LMG95 power meter. The power consumption during the decoding process of a showcase bit stream is displayed in Fig. [\[fig:decPower\]](#fig:decPower){reference-type="ref" reference="fig:decPower"}. The red curve shows the power consumed in idle mode (static power) which is used as a reference. The blue curve shows the power during the decoding process (static power plus dynamic power). High variations can be observed. To obtain the pure decoding energy, both curves are integrated over time and subtracted from each other. As a result, the area in between both curves represents the dynamic decoding energy which is the target of this work. To ensure that the measured energies are correct, multiple measurements are performed for each sequence with a subsequent significance test to ensure validity as explained in. Figure [\[fig:multiQP_BD\]](#fig:multiQP_BD){reference-type="ref" reference="fig:multiQP_BD"} shows the rate-energy diagram with decoding energies measured for six sequences that were taken from each class of the HEVC common test conditions. First of all, we can see that the energy savings are lower than the modeled savings shown in Fig. [\[fig:single_multi_BD\]](#fig:single_multi_BD){reference-type="ref" reference="fig:single_multi_BD"}. Furthermore, the right part of the curves tend towards smaller energy savings although we had expected monotonically increasing savings. Both observations can be attributed to the energy model applied: since it is imperfect and was trained for conventionally coded bit streams, it is consequently less accurate for energy optimized bit streams. Additionally, we can see that highest energy savings are obtained for the BQTerrace sequence, which attains savings of more than \(30\%\). Even the lowest resolution (BlowingBubbles) exhibits energy savings of nearly \(20\%\). Note that similar results can be observed for the other sequences of the corresponding class. A more detailed analysis on the energy consumption and the use of features for various rate optimized and energy optimized bit streams can be performed online at. Finally, the proposed optimization process need not be restricted to a single software solution. To illustrate this, we measure the energy consumption of the optimized bit streams (note that optimizations were explicitly performed for the FFmpeg software) for varying software decoders, namely HM-13.0 and libde265. The hardware is the same as before. The rate-energy diagram for all three decoders is displayed in Figure [\[fig:BD_decoders\]](#fig:BD_decoders){reference-type="ref" reference="fig:BD_decoders"} (results are shown for the BQTerrace sequence). We can see that energy savings are obtained for all solutions, which comes as no surprise since all decoders implement the same functionality. Surprisingly, the greatest savings are obtained for the libde265 decoder, which attains a BDDE of more than \(40\%\). To explain this observation, we compare the differences between the feature numbers \(n_f\) for the sequence coded with \(\tau=0\) to the sequence coded with \(\tau=0.5\). The feature numbers for the fractional pel interpolation and SAO show the highest differences such that, presumably, the implementation of the subpel filters and/or the SAO loop filter is relatively inefficient in the libde265 decoder software, thus resulting in higher energy savings. Note that we expect even greater savings when using a model that is explicitly trained for the libde265 decoder. Table [2](#tab:BDvals){reference-type="ref" reference="tab:BDvals"} summarizes the mean BDDE and BDR values for a selected subset of evaluation cases. We chose \(\tau=0.2\) for intermediate energy savings with a small rate trade-off and \(\tau=0.5\) for high energy savings. We also suggest to choose one of these values in practical applications, depending on the requirements of the user.
:::
We can see that approximately \(15\%\) decoding energy can be saved when accepting a bitrate increase of less than \(5\%\) (\(\tau=0.2\)). If higher rates can be tolerated, energy savings of up to \(30\%\) can be achieved (\(\tau=0.5\)). Finally, the additional complexity of the encoder in terms of encoding time is assessed. As a processor, an AMD Opteron 2356 at \(2.3\) GHz is used. The encoding of the BQTerrace sequence is analyzed for four QP values (QP \(\in\{22, 27, 32, 37\}\)) and three values of \(\tau\) (\(\tau\in\{0, 0.5, 1\}\)). As a reference, the unchanged encoder version HM-14.0 is used with the same four QP values. The results indicate that the complexity increase is highly variable. In addition to the complexity needed for energy estimation and cost calculations, more coding modes are tested because the fast mode decisions, which are explicitly designed for regular RDO, are not optimal for the new cost function. The tests revealed that the complexity increase ranges from \(17\%\) to \(70\%\) and shows an average value of \(42\%\).
# Conclusions {#sec:concl}
In this paper, we introduced the concept of decoding-energy-rate-distortion optimization for video coding. We gave an elaborate introduction into the DERD theory and derived a helpful relation between the two Lagrange multipliers and the QP. Extensive tests using various HEVC software decoders as a use case proved the performance of the proposed approach. The experiments showed that up to \(30\%\) of the energy can be saved at the same visual objective quality when accepting bitrate increases in the order of \(20-50\%\), depending on the sequence. The HEVC encoder software, including the specific energies used for the energy model and trained values for \(\rho\) and \(\varepsilon\), can be downloaded and freely used for personal use. The proposed method of DERDO creates a vast amount of possible further research topics: Research can focus on improving the decoding energy model or the training process of the specific energies. Furthermore, in the DERDO process, the transmission energy can be considered explicitly and new encoder speed-up methods targeting the decoder energy consumption could be developed. Additionally, DERDO could be applied to other codecs such as VP9 or H.264 and even newly developed codecs could benefit from incorporating energy saving coding modes. Furthermore, potential savings with a specified hardware decoder would be of particular interest, in which the RAM modeling will have a much higher influence on the potential savings. Finally, the DERDO theory could be translated to other functionalities such as a decoding-time-rate-distortion theory that would aim at decoding time constraints for real time capability. | {'timestamp': '2022-03-03T02:28:26', 'yymm': '2203', 'arxiv_id': '2203.01099', 'language': 'en', 'url': 'https://arxiv.org/abs/2203.01099'} |
null | null |
null | null |
# Introduction {#sec:intro}
In this note, we describe some new examples of holomorphic modular forms on the real Lie group \(\SU(2,1)\). The forms have weight \(\frac{1}{3}\), and their levels are certain congruence subgroups of \(\SU(2,1)\). We recall that various half-integral weight forms on \(\SU(2,1)\) have been studied in the past, but our examples seem to be the first whose weight is not half-integral. Indeed, they are not metaplectic forms, in the sense that they do not arise from automorphic forms on the metaplectic cover of an adèle group. This is because the adèlic metaplectic covers of forms of \(\SU(2,1)\) are at most two-fold covers.
#### Notation.
To state our results a little more precisely, we introduce some notation. Consider the Hermitian form of \(\mathbb{C}^3\) defined by \(\langle v,w\rangle = \bar v^t J w\), where \(J=\begin{pmatrix} &&1\\&1\\1\end{pmatrix}\). Here and throughout the paper, we write \(\bar v^t\) for the conjugate transpose of \(v\). The group \(\SU(2,1)\) consists of the matrices in \(\SL_3(\mathbb{C})\) which preserve this Hermitian form. Equivalently, the elements of \(\SU(2,1)\) are the matrices \(g\) in \(\SL_3(\mathbb{C})\) which satisfy \(\bar g^t J g = J\).[^1] The symmetric space associated to this Lie group may be identified with the following complex manifold: \[\mathcal{H} = \left\{ \utau = \begin{pmatrix} \tau_1\\ \tau_2 \end{pmatrix} \in \mathbb{C}^2: \left\langle \begin{pmatrix} \tau_1 \\ \tau_2 \\ 1\end{pmatrix},\begin{pmatrix} \tau_1 \\ \tau_2 \\ 1\end{pmatrix}\right\rangle < 0 \right\}.\] Let \(g=\begin{pmatrix} A&B\\mathbb{C}&D \end{pmatrix}\) be an element of \(\SU(2,1)\), where \(A\) is a \(2\times 2\) matrix, \(D\) is a complex number, etc. The action of \(g\) on \(\mathcal{H}\) is given by \[g*\utau =\frac{1}{C\utau+D} \cdot (A\utau+B).\] We also define \[\label{eqn: j definition} j(g,\utau) = C \utau + D.\] The function \(j(g,\utau)\) satisfies the usual condition of a multiplier system: \[\label{eqn: j multiplier system} j(gh,\utau) = j(g,h * \utau)\cdot j(h,\utau), \qquad g,h \in \SU(2,1),\quad \utau \in \mathcal{H}.\]
Half-integral weight multiplier systems have been known about for some time, and have been studied (for example) in and. The half-integral weight multiplier systems arise from the theory automorphic forms on the metaplectic double cover of a unitary adèle group. However, no other fractional weight multiplier systems on \(\SU(2,1)\) have been shown to exist until recently. By a *fractional weight modular form* of level \(\Gamma\) with weight \(\frac{a}{b}\), we shall mean a holomorphic function \(f: \mathcal{H} \to \mathbb{C}\), such that for all \(\utau\in \mathcal{H}\) and all \(\gamma\in \Gamma\), we have \[f(\gamma*\utau) = \ell(\gamma,\utau) \cdot f(\utau),\] where \(\ell\) is a multiplier system with weight \(\frac{a}{b}\). The usual growth condition in cusps is redundant for holomorphic modular forms on \(\SU(2,1)\).
#### Statement of results.
The purpose of this paper is to describe certain modular forms on \(\SU(2,1)\) of weight \(\frac{1}{3}\). The levels of these forms are congruence subgroups of the following group \[\label{eqn:Gamma(1) defn} \Gamma(1) = \SU(2,1) \cap \SL_3(\mathbb{Z}[\zeta]),\qquad \text{where } \zeta = e^{\frac{2 \pi i}{3}}.\] (The reason for focusing on subgroups of \(\Gamma(1)\) is because we make use a presentation for \(\Gamma(1)\) found in and the description of integral weight forms in ). For a non-zero \(\beta\in \mathbb{Z}[\zeta]\), we shall write \(\Gamma(\beta)\) for the principal congruence subgroup of level \(\beta\) in \(\Gamma(1)\). The levels of our forms are intermediate groups between \(\Gamma(\sqrt{-3})\) and \(\Gamma(3)\), all of which have index \(3\) in \(\Gamma(\sqrt{-3})\). One way in which one might expect to construct fractional weight forms is to use the theory of Borcherds products associated to the lattice \(\mathbb{Z}[\zeta]^3\). Recall that such a Borcherds product is a meromorphic modular form \(\Psi:\mathcal{H} \to \mathbb{C}\cup\{\infty\}\), whose divisor is a certain integer linear combination of Heegner divisors. Given a positive integer \(n\) and a congruency class in \(v \in \frac{1}{\sqrt{-3}}\mathbb{Z}[\zeta]^3 / \mathbb{Z}[\zeta]^3\), the Heegner divisor \(D_{n,v}\) is defined by \[\label{eqn:Heegner} D_{n,v} = \begin{cases} \displaystyle{\sum_{w \in \mathbb{Z}[\zeta]^3 + v \;:\; \langle w, w \rangle = n} w^\perp} & \text{if \(2v \in \mathbb{Z}[\zeta]^3\)}\\[8mm] \displaystyle{\frac{1}{2}\sum_{w \in \mathbb{Z}[\zeta]^3 + v \;:\; \langle w, w \rangle = n} w^\perp} & \text{if \(2v \not\in \mathbb{Z}[\zeta]^3\),} \end{cases}\] where \(w^\perp\) is the set of \(\utau \in \mathcal{H}\) such that \(\langle \begin{pmatrix}\utau\\ 1\end{pmatrix},w\rangle = 0\). Notice that if \(w\) is a vector occurring in the sum ([\[eqn:Heegner\]](#eqn:Heegner){reference-type="ref" reference="eqn:Heegner"}), then \(\zeta w\) and \(\zeta^2 w\) also arise in the sum, and we have \(w^\perp = (\zeta w)^\perp = (\zeta^2 w)^\perp\). Hence each Heegner divisor \(D_{n,v}\) is a multiple of \(3\). This implies that each Borcherds lift to \(\SU(2,1)\) must be the cube of a meromorphic function on \(\mathcal{H}\). If a Borcherds lift has weight \(k\), then its cube root will be a fractional weight form of weight \(\frac{k}{3}\), implying that there is a multiplier system with weight \(\frac{k}{3}\). The level of a Borcherds product associated with the lattice \(\mathbb{Z}[\zeta]^3\) is the principal congruence subgroup \(\Gamma(\sqrt{-3})\). In view of this, it makes sense to begin searching for third-integral weight forms at level \(\Gamma(\sqrt{-3})\). In spite of this optimism, we obtain the following result:
This result is proved in ; it is a proof by contradiction, using a carefully chosen relation in the group \(\Gamma(\sqrt{-3})\). Note that the theorem implies that the weight of each of the Borcherds product associated to the lattice \(\mathbb{Z}[\zeta]^3\) must be a multiple of \(3\). Searching a little further, we do indeed find some fractional weight multiplier systems:
The result is proved computationally, by finding a presentation for each of the subgroups. The thirteen subgroups mentioned in the theorem are listed in . gives us a list of levels where we might conceivably find a modular form of weight \(\frac{1}{3}\). Using results of, we obtain a description of the ring of integral weight modular forms of level \(\Gamma(3)\), as well as the action of \(\Gamma(\sqrt{-3})/\Gamma(3)\) on this ring. We find a list of forms of level \(\Gamma(3)\) and weight \(1\), all of whose divisors are multiples of \(3\). By considering the action of \(\Gamma(\sqrt{-3})/\Gamma(3)\) on these forms, we discover that each of them is a modular form (with character) of some larger level. Each of these larger levels is one of the groups found in Theorem 2. The cube roots of these forms are forms of weight \(\frac{1}{3}\). In this way, we prove the following:
#### Further comments and questions.
leaves an obvious open question: why is there no form of weight \(\frac{1}{3}\) on the remaining group \(\Gamma\) in our list. We have no clear answer to this question. We have investigated the multiplier systems on many other congruence subgroups of \(\Gamma(1)\). For each congruence subgroup that we have investigated, the denominator of the weight of every multiplier system is a factor of \(6\). It is tempting to think that this might be true for all congruence subgroups of \(\Gamma(1)\). One might even guess that if we replace the field \(\mathbb{Q}(\zeta)\) involved in the construction of \(\Gamma(1)\) by another CM field \(k\), then a similar bound might be the number of roots of unity in \(k\). However, we have not proved any such result; neither do we have any numerical evidence beyond the case \(k=\mathbb{Q}(\zeta)\). The situation for non-congruence subgroups is very different. For every positive integer \(n\), there exists a (non-congruence) subgroup \(\Gamma_n\) of finite index in \(\Gamma(1)\), such that \(\Gamma_n\) has a multiplier system of weight \(\frac{1}{n}\). This fact follows from a more general result proved in, and independently in. For \(d > 2\), it is not known whether there exist any modular forms whose level is a congruence subgroup \(\SU(d,1)\), and whose weight has denominator greater than \(2\). However, such forms certainly do exist at non-congruence levels by results in.
#### Organization of the paper.
The paper is organized as follows. Suppose \(\Gamma\) is an arithmetic subgroup of \(\SU(2,1)\). We prove (see ) that there is an upper bound on the denominators of weights of multiplier systems on \(\Gamma\). We define in a natural number called the *weight denominator* of \(\Gamma\). The weight denominator is the lowest common denominator of the weights of all multiplier systems on \(\Gamma\). Let \(\tilde\Gamma\) be the pre-image of \(\Gamma\) in the universal cover of \(\SU(2,1)\). One may easily calculate the weight denominator of \(\Gamma\) if one has a presentation of \(\tilde\Gamma\). In , we recall some methods for obtaining such presentations. These methods give us a computational approach to determining the weight denominator of a given arithmetic group. In we describe in detail how the methods from have been implemented for subgroups of finite index in the group \(\Gamma(1)\). We calculate the weight denominators of various arithmetic groups, and we list in thirteen congruence subgroups with weight denominator \(3\), all with index \(3\) in \(\Gamma(\sqrt{-3})\). In we construct, for twelve of the thirteen groups found in , a modular form of weight \(\frac{1}{3}\).
# Background on SU(2,1) {#section:background}
In this section we shall define and investigate a positive integer, which we call the *weight denominator* of an arithmetic subgroup of \(\SU(2,1)\). Let \(\widetilde{\SU(2,1)}\) be the universal cover of \(\SU(2,1)\). For an arithmetic group \(\Gamma\) we write \(\tilde\Gamma\) for the pre-image of \(\Gamma\) in \(\widetilde{\SU(2,1)}\). The *weight denominator* of \(\Gamma\) is defined to be the order of the kernel of the projection map \(\tilde\Gamma / [\tilde\Gamma,\tilde\Gamma] \to \Gamma / [\Gamma,\Gamma]\). We prove in that the weight denominator is finite. We prove in that \(\Gamma\) has a multiplier system of rational weight \(w\) if and only if the denominator of \(w\) is a factor of the weight denominator of \(\Gamma\). Some other simple properties of the weight denominator are proved in . Methods for computing the weight denominator of an arithmetic group are discussed in and .
## The universal cover
The Lie group \(\SU(2,1)\) has fundamental group \(\mathbb{Z}\). We shall write \(\widetilde{\SU(2,1)}\) for its universal cover. We therefore have a central extension of groups \[1 \to \mathbb{Z} \to \widetilde{\SU(2,1)} \to \SU(2,1) \to 1.\] Such extensions are classified by elements of the measurable cohomology group \(H^2_{\meas}(\SU(2,1),\mathbb{Z})\) (see ). In order to perform calculations, it will be helpful to have a specific measurable 2-cocycle \(\sigma\) representing this group extension. We describe such a cocycle in below. As a first step, we define a function \(X: \SU(2,1)\to \mathbb{C}\) by \[X\begin{pmatrix} *&*&*\\ *&*&*\\ a&b&c \end{pmatrix} = \begin{cases} -a & \text{if \(a\ne 0\),}\\ c & \text{if \(a = 0\).} \end{cases}\]
allows us to define, for each \(g\in \SU(2,1)\), a branch \(\tilde j(g,-)\) of the logarithm of \(j(g,-)\) as follows: \[\label{eqn:j-tilde} \tilde j(g,\utau) = \log\left( \frac{j(g,\utau)}{X(g)} \right) + \log(X(g)),\] where each logarithm in the right hand side of ([\[eqn:j-tilde\]](#eqn:j-tilde){reference-type="ref" reference="eqn:j-tilde"}) is defined to be continuous away from the negative real axis and satisfy \(- \pi < \Im ( \log z) \le \pi\). For each fixed \(g\in\SU(2,1)\) the function \(\tilde j(g,\utau)\) is continuous in \(\utau\) by . The multiplier system condition ([\[eqn: j multiplier system\]](#eqn: j multiplier system){reference-type="ref" reference="eqn: j multiplier system"}) on \(j(g,\utau)\) implies the following congruence for \(\tilde j\): \[\tilde j(gh,\utau) \equiv \tilde j(g,h*\utau) + \tilde j(h,\utau) \mod 2\pi i\cdot \mathbb{Z}.\] In particular, we may define for \(g,h \in\SU(2,1)\) an integer \(\sigma(g,h)\) by \[\label{eqn:sigma definition} \sigma(g,h) = \frac{1}{2\pi i} \Big(\tilde j(gh,\utau)-\tilde j(g,h*\utau)-\tilde j(h,\utau)\Big).\] The right hand side of ([\[eqn:sigma definition\]](#eqn:sigma definition){reference-type="ref" reference="eqn:sigma definition"}) is independent of \(\utau\), since it is a continuous \(\mathbb{Z}\)-valued function of \(\utau\).
From now on, we shall refer to elements of \(\widetilde{\SU(2,1)}\) as pairs \((g,n) \in \SU(2,1)\times \mathbb{Z}\), with the group operation given by ([\[eqn:sigma multiplication\]](#eqn:sigma multiplication){reference-type="ref" reference="eqn:sigma multiplication"}). To find out whether a fractional weight multiplier system exists on a group \(\Gamma\), we shall use the following result.
## The weight denominator {#sec:wt denominator}
The following proposition is useful to know, although we do not use it in the rest of this paper.
In the proof we'll use the following facts about the cohomology of a connected Lie group \(G\):
- There is an isomorphism \(H^2_{\meas}(G,\mathbb{Z}) \cong \Hom(\pi_1(G),\mathbb{Z})\) (see ). In particular \(H^2_{\meas}(G,\mathbb{Z})\) is torsion-free, so in injects into \(H^2_{\meas}(G,\mathbb{Z})\otimes\mathbb{C}\).
- By, there is an isomorphism \(H^\bullet_{\meas}(G,\mathbb{Z}) \otimes \mathbb{C} \cong H^\bullet_{\cts}(G,\mathbb{C})\), where \(H^\bullet_{\cts}\) denotes continuous cohomology.
- If \(\Gamma\) is a cocompact arithmetic subgroup of \(G\) then by results in, the restriction maps \(H^\bullet_{\cts}(G,\mathbb{C}) \to H^\bullet(\Gamma,\mathbb{C})\) are all in injective.
Combining these results, we see that if \(\Gamma\) is cocompact in \(G\), then the map \(H^2_{\meas}(G,\mathbb{Z}) \to H^2(\Gamma,\mathbb{C})\) is injective.
# Calculus of presentations {#sec:presentations}
To calculate the weight denominators of arithmetic groups \(\Gamma\), we need some methods for producing presentations of the groups \(\tilde\Gamma\). In this section, we shall describe some methods for doing this. We have used as our main source for this section. The methods described here apply to arbitrary finitely presented groups.
## Covering groups {#sec:covering groups}
Suppose \(\tilde G\) is a central extension of a group \(G\) by a cyclic group \(\langle z \rangle\) of the form \[1 \to \langle z \rangle \to \tilde G \to G \to 1.\] Assume that we have a presentation of \(G\): \[G = \langle g_1,\ldots,g_n | r_1, \ldots, r_m \rangle,\] where each relation \(r_i\) is a word in the generators \(g_i\). We shall describe a presentation of \(\tilde G\). For each generator \(g_i\) in \(G\), we choose a pre-image \(\hat g_i\) of \(g_i\) in \(\tilde G\). The group \(\tilde G\) is generated by the elements \[\hat g_1, \ldots, \hat g_n, z.\] There are three obvious kinds of relation in \(\tilde G\):
1. For each relation \(r_i\) in our presentation of \(G\), we let \(\hat r_i\) be the word in the generators of \(\tilde G\), obtained by replacing each \(g_i\) with the corresponding element \(\hat g_i\) in \(\tilde G\). Since \(r_i=1\) in \(G\), it follows that \(\hat r_i \in \langle z\rangle\), so we have a relation \(\hat r_i = z^{c_i}\) in \(\tilde G\).
2. Since the extension is assumed to be central, the element \(z\) is in the centre of \(\tilde G\), so we have a relation \([z,\hat g_i]\) for each generator \(g_i\) in \(G\).
3. If \(z\) has finite order \(n\), then we have the relation \(z^n=1\).
The generators and relations listed above give a presentation of \(\tilde G\).
## Subgroups {#sec:Reidemeister Schreier}
Let \(H\) be a subgroup of finite index in \(G\). Assume that we have a finite presentation of \(G\) with generators \(g_i\). The method for constructing a finite presentation of \(H\) is called the Reidemeister--Schreier algorithm. This algorithm constructs a certain directed graph, whose vertex set \(R\) is a set of coset representatives for \(H\) in \(G\). The Reidemeister--Schreier graph is constructed as follows:
1. We begin by setting \(R_0=R=\{1_G\}\), where \(1_G\) is the identity element in the group \(G\).
2. Choose an element \(r \in R_0\) and remove it from the set \(R_0\).
3. For each element \(g\) which is either one of the generators \(g_i\) of \(G\) or its inverse \(g_i^{-1}\), consider the coset \(Hrg\). Check whether \(Hrg\) is one of the cosets \(Hr'\) for some element \(r'\) in \(R\).
1. If \(Hrg=Hr'\), then we must have \(rg=hr'\) for some \(h\in H\). We then add an edge from \(r\) to \(r'\), and label that edge \((g, h)\).
2. If \(Hrg\) is not one of these cosets, then we add \(rg\) to our set \(R\) and also to our set \(R_0\). We also add an edge from \(r\) to \(rg\) in our graph, labelled \((g,1)\).
4. If \(R_0\) is non-empty, then go back to step 2. Otherwise stop.
The algorithm will terminate with a graph whose vertices are labelled by representatives \(r\) for cosets \(Hr\). For each element \(g\) which is either a generator \(g_i\) or its inverse \(g_i^{-1}\), there is an edge from \(r\) to a vertex \(r'\), where \(Hr g = Hr'\). This edge is labelled \((g,h)\), where the element \(h\in H\) satisfies \(rg=hr'\). We shall refer to \(g\) and \(h\) as the \(G\)-label and \(H\)-label of the edge respectively. We'll now explain how one uses the Reidemeister--Schreier graph to construct a finite presentation of \(H\) from one of \(G\). Suppose we have a word \(w = w_1 \cdots w_s\) in the generators \(g_i^{\pm 1}\) of \(G\). For any vertex \(r\) in the graph, we may form a path starting at \(r\), and moving along the vertices with \(G\)-labels \(w_1, w_2, \ldots, w_s\) in turn. The end of the path will be a vertex \(r'\) such that \(Hr w_1 \cdots w_s = Hr'\). If \(h_1, \cdots, h_s\) are the \(H\)-labels of the edges in the path then we have \[r\cdot w_1 \cdots w_s = h_s \cdots h_1\cdot r'.\] If \(w_1 \cdots w_s=1\) is one of our relations in \(G\), then this path will be a closed loop, i.e. \(r'=r\). We will therefore have \[r = r \cdot w_1 \cdots w_s = h_s \cdots h_1 \cdot r,\] This gives us a relation \(h_s \cdots h_1 = 1\) in the subgroup \(H\). It is known that the subgroup \(H\) is generated by the elements \(h\) which are the \(H\)-labels of edges in the Reidemeister--Schreier graph. Furthermore, the set of all relations arising as described above gives a presentation of the subgroup \(H\).
## Rewriting Presentations {#sec:rewriting}
Typically, the Reidemeister--Schreier algorithm produces a presentation with many generators, when far fewer generators are actually needed. The *rewriting* process allows us to reduce the number of generators. Suppose that we have a group presentation: \[G = \langle g_1,\ldots,g_r | r_1,\ldots,r_t \rangle,\] and we also have a second set of generators \(\{h_1,\ldots,h_s\}\) for \(G\). We'll explain now how to find a presentation using \(h_1,\ldots, h_s\) as generators. This process is called "rewriting" the presentation.
1. Find an expression for each element \(h_i\) as a word in the generators \(g_j\): \[h_i = w_i(g_1,\ldots,g_r).\] Then we can the add the generators \(h_i\) to the presentation of \(G\), together with the relations \(h_i=w_i(g_1,\ldots,g_r)\).
2. Find an expression for each \(g_i\) as a word in the generators \(h_j\): \[g_i = x_i(h_1,\ldots,h_s).\] We can add the relations \(g_i = x_i(h_1,\ldots,h_s)\) to the presentation of \(G\).
3. In each relation apart from the relation \(g_i = x_i(h_1,\ldots,h_s)\), we replace each occurrence of \(g_i\) by the word \(x_i(h_1,\ldots,h_s)\).
4. Finally, we remove the generators \(g_i\) and the relations \(g_i=x_i(h_1,\ldots,h_s)\).
# Weight denominators of certain arithmetic groups {#sec:results}
This section concerns computer calculations. The code for these calculations runs on sage version 9.0, and is available at. As before (see ([\[eqn:Gamma(1) defn\]](#eqn:Gamma(1) defn){reference-type="ref" reference="eqn:Gamma(1) defn"})), we shall write \(\Gamma(1)\) for the subgroup of \(\SU(2,1)\) consisting of matrices whose entries are in the ring \(\mathbb{Z}[\zeta]\) of Eisenstein integers, where \(\zeta=e^{2\pi i /3}\). Let \(\Gamma(\sqrt{-3})\) denote the principal congruence subgroup in \(\Gamma(1)\) of level \(\sqrt{-3}\). As mentioned in the introduction, the group \(\Gamma(\sqrt{-3})\) is a level where one might expect to find forms of third-integral weight. In spite of this, we show in that \(\Gamma(\sqrt{-3})\) has weight denominator \(1\). Looking just a little further up the tower of congruence subgroups, we do indeed find groups with weight denominator \(3\). The group \(\Gamma(3)\) has weight denominator \(3\). Furthermore, there are precisely 13 subgroups of index \(3\) in \(\Gamma(\sqrt{-3})\) containing \(\Gamma(3)\), which have weight denominator \(3\) (see ).
## The subgroup \(\Upsilon\)
Rather than dealing with the group \(\Gamma(\sqrt{-3})\) directly, it will be slightly more convenient to consider a certain subgroup \(\Upsilon\) of index \(3\) in \(\Gamma(\sqrt{-3})\). This is because \(\Upsilon\) has a slightly smaller presentation (see below).
Note that by the groups \(\Upsilon\) and \(\Gamma(\sqrt{-3})\) have the same weight denominator, since they are the same modulo the centre.
We shall find a presentation for \(\Upsilon\). We start by finding a small generating set. It will be useful to have the following notation for upper-triangular elements of \(\Upsilon\): \[n(z,x) = \begin{pmatrix} 1 & \sqrt{-3} \cdot z & \frac{-3\Norm(z) + x\sqrt{-3}}{2}\\[2mm] 0 & 1 & \sqrt{-3}\cdot\bar z\\[2mm] 0 & 0 & 1 \end{pmatrix}, \qquad z\in \mathbb{Z}[\zeta],\quad x\in\mathbb{Z},\quad x\equiv \Norm(z) \bmod 2.\] Here and later we write \(\Norm\) and \(\Tr\) for the norm and trace maps from \(\mathbb{Q}(\zeta)\) to \(\mathbb{Q}\).
## The weight denominators of \(\Upsilon\) and \(\Gamma(\sqrt{-3})\)
To calculate the weight denominators of \(\Upsilon\) and \(\Gamma(\sqrt{-3})\) we shall use a presentation of \(\Upsilon\).
Although was found using a computer, it would be possible to verify the result by hand (given a few days). One needs only to calculate the element \(\hat R\) in \(\tilde{\Upsilon}\), and note that this element is in the commutator subgroup. It would however be difficult to find the relation \(R\) without the presentation of .
## Subgroups of \(\Upsilon\)
There is code (see ) which calculates the number \(\denom (\Gamma)\) for certain arithmetic subgroups \(\Gamma\) of \(\Upsilon\). The method of calculation is as follows.
1. We already have a presentation of \(\Upsilon\) (). Using the Reidemeister--Schreier algorithm, we find a presentation of \(\Gamma\): \[\Gamma = \langle \gamma_1, \ldots, \gamma_r | \delta_1 =\cdots = \delta_s = 1 \rangle.\]
2. For each generator \(\gamma_i\) we let \(\hat\gamma_i = (\gamma_i,0) \in \tilde\Gamma\). For each relation \(\delta_j\), we let \(\hat \delta_j\) be the element of \(\tilde\Gamma\) obtained by replacing each generator \(\gamma_i\) in \(\delta_j\) by its lift \(\hat\gamma_i\). In \(\tilde\Gamma\) we have relations \[\label{eqn:delta hat} \hat \delta_j \cdot (I_3,n_j) = 1, \qquad n_j \in \mathbb{Z}.\] The integers \(n_j\) are calculated by multiplying in \(\tilde{\Gamma}\) using ([\[eqn:sigma multiplication\]](#eqn:sigma multiplication){reference-type="ref" reference="eqn:sigma multiplication"}).
3. To each of the relations ([\[eqn:delta hat\]](#eqn:delta hat){reference-type="ref" reference="eqn:delta hat"}) we form a row vector in \(\mathbb{Z}^{r+1}\), where the first \(r\) entries are the multiplicities of the generators \(\hat\gamma_i\) and the last entry is \(n_j\). If we let \(M\) be the matrix formed of these rows, then we have a presentation \[\tilde\Gamma^{\ab} \cong \mathbb{Z}^{r+1} / (\mathbb{Z}^s \cdot M).\]
4. The subgroup \(\mathbb{Z}^s \cdot M\) is unchanged by integral row operations on \(M\). We may transform \(M\) by a sequence of such operations to a matrix \(M'\) in Hermite normal form. The last non-zero row of \(M'\) must have the form \[\begin{pmatrix} 0 & \cdots & 0 & n \end{pmatrix}, \qquad n > 0.\] Indeed if this were not the case, then the generator \((I_3,1)\) would have infinite order in \(\tilde\Gamma^{\ab}\), contradicting .
5. The positive integer \(n\) is the order of \((I_3,1)\) in \(\tilde{\Gamma}^{\ab}\), which is by definition the weight denominator of \(\Gamma\).
The speed of the computation depends on the index of the subgroup \(\Gamma\) in \(\Upsilon\). The current version of the code is able to handle subgroups of index up to around 500 on a home computer. The most time consuming step is currently the row reduction in step 4. There is an opportunity (using ) to speed up this step by row-reducing modulo the prime powers which divide the index of \(\Gamma\) in \(\Upsilon\), rather than row-reducing over \(\mathbb{Z}\). There is also scope for simplifying the presentation of \(\Gamma\) at the end of step 1, using a smaller set of generators (as we did for \(\Upsilon\) in ). At the moment, this is not implemented.
## Congruence subgroups with weight denominator \(3\) {#sec:index 3 subgroups}
The principal congruence subgroup \(\Gamma(3)\) is a normal subgroup of \(\Upsilon\) with index \(81\). One easily checks using the code at that \(\Gamma(3)\) has weight denominator \(3\). This means that \(\Gamma(3)\) has a multiplier system with weight \(\frac{1}{3}\). If any reader would like to run this code and experiment further, then there are some instructions online. The sage code for this particular calculation is
sage: load("SU21.sage") sage: G = Gamma(3) sage: print(G.weight_denominator())
However, one does not need to go as far as \(\Gamma(3)\) to find such a multiplier system. We shall discuss the intermediate groups between \(\Upsilon\) and \(\Gamma(3)\)
In view of the lemma, the intermediate group between \(\Upsilon\) and \(\Gamma(3)\) correspond to subgroups of \(\mathbb{F}_3^4\). In particular, for \(a,b,c,d\in \mathbb{F}_3\) we define a subgroup of \(\Upsilon\) by \[\Upsilon_{\mathrm{index}\ 3}(a, b,c,d) = \{ g \in \Upsilon: a\cdot g_{1,2}+b\cdot g_{1,3}+c\cdot g_{2,1}+d\cdot g_{3,1} \equiv 0 \bmod 3\}.\] If \(v\) is a non-zero vector in \(\mathbb{F}_3^4\) then \(\Upsilon_{\mathrm{index}\ 3}(v)\) has index \(3\) in \(\Upsilon\). Every subgroup of index \(3\) in \(\Upsilon\) containing \(\Gamma(3)\) has this form. Two of these subgroups \(\Upsilon_{\mathrm{index}\ 3}(v)\) and \(\Upsilon_{\mathrm{index}\ 3}(v')\) are equal if and only if \(v \equiv \pm v'\bmod 3\). We therefore have 40 subgroups of index \(3\) containing \(\Gamma(3)\). By , each of these subgroups has weight denominator either \(1\) or \(3\). We have calculated all of these weight denominators. For example, to calculate the weight denominator of \(\Upsilon_{\mathrm{index}\ 3}(0,0,1,0)\), the command is
sage: load("SU21.sage") sage: G = Index3congruence(0,0,1,0) sage: print(G.weight_denominator())
The results of our calculations are as follows.
Note that for each of the groups \(\Gamma\) listed above, the group \(\Gamma\cdot Z\) is a subgroup of index \(3\) in \(\Gamma(\sqrt{-3})\), and also has weight denominator \(3\) by .
# Modular forms {#section: modular forms}
Multiplier systems arise when one wants to consider modular forms of non integral weight. It is a natural problem to realize a given multiplier system by a modular form. In some cases the whole rings of modular forms of integral weight have been determined and we can use these results to construct forms of non integral weight. Recall that a modular form on an arithmetic subgroup \(\Gamma\subset \U(n,1)\) and with respect to a multiplier system \(\ell(g,\tau)\) is a holomorphic function \(f:\mathcal{H}\to\mathbb{C}\) with the property \[f(g\tau)=\ell(g,\tau)f(\tau),\quad g\in\Gamma,\] where in the case \(n=1\) the usual regularity condition at the cusps has to be added. Notice that we work in this section with the group \(\U(n,1)\) instead of \(\SU(n,1)\). This is due to the fact that in the theory of modular forms reflections play an important role. Of course Definition 1 works literally in the case \(\U(n,1)\). A general result using Poincaré series or compactification theory states that for every multiplier system of weight \(\frac{a}{b}\) there exists a natural number \(k\gg0\) such that the multiplier system \(\ell(g,\tau)j(g,\tau)^k\) of weight \(\frac{a}{b}+k\) admits a non zero modular form. We will see that 12 of the above 13 multiplier systems admit a modular form of weight \(\frac{1}{3}\).
## Rings of modular forms
We consider the \((n+1)\times (n+1)\) matrix \[S=\begin{pmatrix}0&-1&&\\-1&0&&\\&&1&\\&&&\ddots\\&&&&1\end{pmatrix}.\] It defines a Hermitian form of signature \((n,1)\), \[\langle z,w\rangle =\bar z^t S w,\quad z,w\in\mathbb{C}^{n+1}\ \hbox{(columns)}.\] We denote by \(G_n\) the group of all \((n+1)\times(n+1)\)-matrices \(g\) with coefficients in the ring of Eisenstein integers with the property \(\bar g^t S g=S\). Notice that we do not assume that \(\det(g)=1\). In the case \(n=2\) it is related to the form \(J\) from section [1](#sec:intro){reference-type="ref" reference="sec:intro"} by \[J=\bar V^t SV,\quad V=\begin{pmatrix}-1&0&0\\0&0&1\\0&1&0\\ \end{pmatrix}.\] So the group \(\Gamma(1)\) of section [4](#sec:results){reference-type="ref" reference="sec:results"} is embedded into \(G_2\), \[\Gamma(1)\longrightarrow G_2,\quad g\longmapsto h=VgV^{-1}.\] For each non zero Eisenstein integer \(\beta\) we can consider the principal congruence subgroups \(\Gamma(\beta)\) and \(G_n(\beta)\). The group \(\Gamma(\beta)\) can be identified with the subgroup of \(G_2(\beta)\) of matrices with determinant one. Rings of modular forms of integral weight have been determined in the literature for \(G_4(\sqrt{-3})\),, , for \(G_3(\sqrt{-3})\),, and for the group \(G_3(3)\). The ring of modular form for \(G_3(3)\) is rather complicated. We can use its structure to derive the rings for \(G_2(2)\) (and also for \(G_1(3)\)). We formulate now these results and give some hints how to get them. First we recall the definition of the ring of modular forms. Here we use the unitary group for the Hermitian form \(S\) and denote it from now on by \(\U(1,n)\). So \(G_n(\ell)\) is a subgroup of \(\U(1,n)\). We have to replace the homogeneous space \(\mathcal{H}\) by the space \(\mathcal{H}_n\) of all columns \(z\in \mathbb{C}^n\) with the property \[{\begin{pmatrix}1\\\bar z\end{pmatrix}}^tS{\begin{pmatrix}1\\\ z\end{pmatrix}}<0.\] This means \[\mathcal{H}_n=\{z\in\mathbb{C}^n;\quad \Re z_1>\vert z_2\vert^2+\cdots+\vert z_n\vert^2\}.\] The action of \(\U(n,1)\) is given by \[g*z =\frac{1}{A+Bz} \cdot (C+Dz).\] Here \(g\) is a block matrix \(g=\begin{pmatrix} A&B\\mathbb{C}&D \end{pmatrix}\), where \(A\) is a complex number, \(D\) a \(2\times2\)-matrix etc. We also define \[\label{eqn: j definition 2} j(g,z) = A + Bz.\] The definition of a multiplier system is the same as in Definition 1 and the notion of a modular form has to be defined as in the beginning of section [5](#section: modular forms){reference-type="ref" reference="section: modular forms"}. The ring of modular forms \(A(G_n(l))\) is the direct sum of the spaces of all modular forms with respect to integral powers of \(j(g,z)\). From compactification theory one knows that this is a finitely generated algebra whose associated projective variety is a compactification (by finitely many points) of \(\mathcal{H}_n/G_n(\beta)\). The embedding \[\mathcal{H}_n\longrightarrow \mathcal{H}_{n+1},\quad z\longmapsto (z,0),\] induces a homomorphism \[A(G_{n+1}(\beta))\longrightarrow A(G_n(\beta)).\] From compactification theory follows that \(A(G_n(\beta))\) is the normalization of the image.
## Examples of rings of modular forms
We apply the discussion above to determine the ring of modular forms for \(G_2(3)\). We want to use the known ring for \(G_3(3)\). This ring is rather complicated. There are 15 basic modular forms \(B_1,\dots,B_{15}\) of weight 1. They have been constructed in as Borcherds products. The algebra \(\mathbb{C}[B_1,\dots,B_{15}]\) agrees with \(A(G_3(3))\) in weight \(\ge 7\). The zeros of the \(B_i\) are located at Heegner divisors which can be considered as \(G_3(\sqrt{-3})\)-translates of \(\mathcal{H}_2\) which can be identified with the subspace of \(\mathcal{H}_3\) defined by \(z_3=0\). A finite system of generators of the ideal of relations has been determined. Using these results it is easy to prove that the restrictions of \[B_1,B_2,B_3,B_4+B_5,B_6+B_7,B_8-B_9,B_{10}-B_{11},B_{12}+B_{13},B_{14}+B_{15}\] to \(\mathcal{H}_2\) are zero. The quotient of \(\mathbb{C}[B_1,\dots,B_{15}]\) by the ideal that is generated by these relations turns out to be normal and of Krull dimension \(3\). Hence this ideal is the kernel of the homomorphism \(\mathbb{C}[B_1,\dots,B_{15}]\to A(G_2(3))\) and this homomorphism must be surjective. In this way we can determine \(A(G_2(3))\). We denote the restrictions of \[B_5,\ B_{11},\ B_{15},\ B_7,\ B_{13},\ B_9\] to \(\mathcal{H}_2\) in the same order by \[X_1,\ X_2,\ X_3,\ X_4,\ X_5,\ X_6.\]
As we mentioned these results are consequences of the paper \[5\] where the case \(G_3\) has been treated. For the action of \(G_3\) on the \(B_i\) we refer to Lemma 8.3 in this paper. The action of \(G_2\) on the \(X_i\) is a consequence.
## Modular forms of weight \(\frac{1}{3}\)
Let \(f\in A(G_2(3))\) be a modular form of weight \(r\). Since the structure of \(A(G_2(3))\) is known and simple enough, one can determine the primary decomposition of the principal ideal \((f)\), \[(f)=fA(G_2(3))=\mathfrak{q}_1\cap\dots \cap \mathfrak{q}_m.\] If one has luck, the multiplicities of the \(\mathfrak{q}_i\) are multiples of 3. Then the multiplicities of the zero components of \(f\) in \(\mathcal{H}_2\) are also multiples of \(3\). So wa can take a holomorphic cube root of \(f\) to produce a form of weight \(\frac{r}{3}\). Hence our construction of forms of weight \(1/3\) rests on the knowledge of the structure of the ring of modular forms of *integral* weight. We give an example. We consider the form \(X_1+X_2\) of weight one. It has been found by trial and error. Its primary decomposition can be computed by means of MAGMA. It turns out that \((X_1+X_2)\) is the intersection of three primary ideals \[\begin{aligned}
\mathfrak{q}_1&=(X_1 + X_2,X_4+(\zeta+1)X_6,X_3^3,X_2^3 + X_5^3-X_6^3),\\ \mathfrak{q}_2&=(X_1 + X_2,X_4-\zeta X_6,X_3^3,X_2^3 + X_5^3-X_6^3),\\ \mathfrak{q}_3&=(X_1 + X_2,X_4-X_6,X_3^3,X_2^3 + X_5^3-X_6^3). \end{aligned}\] The associated prime ideals (which are the radicals) \(\mathfrak{p}_1,\mathfrak{p}_2,\mathfrak{p}_3\) are obtained if one replaces in each case \(X_3^3\) by \(X_3\). For example \[\mathfrak{p}_3=(X_1 + X_2,X_4-X_6,X_3,X_2^3 + X_5^3-X_6^3).\] This defines the elliptic curve \[X_2^3 + X_5^3-X_6^3=0\] in \(P^2\mathbb{C}\). Its \(j\)-invariant is \(0\). All three curves are isomorphic.
In the action of \(G_3\) on \(A_3(G_3)\) has been determined. Using this we can determine the invariance group of \(X_1+X_2\) to verify the following result.
The form \(X_1+X_2\) is a special example of a form of weight 1 which admits a holomorphic third root. To get more such forms, we tried also the forms \(X_i\pm X_j\). It turned out that for each pair \((i,j)\), \(i\ne j\), there is one of the two signs such that the form has third root of unity. The following table contains forms \(X_i\pm X_j\) which have the same property. They belong to a subgroup defined by a congruence \(L(h)\equiv0\mod 3\). The linear form \(L\) is in the first column. The second column contains the form and the third column contains the vector \(v\) such that intersection with \(\SU(2,1)\) corresponds to \(\Upsilon_{\mathrm{index}\ 3}(v)\). These forms are not uniquely determined. In fact, one can replace each form \(X_i\pm X_j\) in the table by \(X_i\pm\zeta^\nu X_j\), \(0\le\nu\le2\) and in this way we get forms with respect to the same group that admit also third roots of unity. This means that we constructed 36 forms of weight \(\frac{1}{3}\). So we have proved that 12 of the above 13 groups in subsection [4.4](#sec:index 3 subgroups){reference-type="ref" reference="sec:index 3 subgroups"} admit three modular forms of weight \(\frac{1}{3}\). One can show that this is false for the remaining group which is given through \(h_{13}\equiv h_{31}\mod 3\). Its intersection with \(\SU(2,1)\) corresponds to \(v=(1,0,2,0)\). In this case one can prove that no form of weight one which belongs to this congruence group and which admits a holomorphic third root can exist. By general arguments there must exist in \(A(G_2(3))\) a modular of weight \(3r+1\), \(r\in\mathbb{Z}\) suitable, which admits a holomorphic third root of unity. So far we do not know an example.
[^1]: One could of course replace \(J\) by any Hermitian matrix of signature \((2,1)\). Our choice of \(J\) is convenient, because the upper triangular matrices of \(\SU(2,1)\) form a Borel subgroup and the diagonal matrices form a maximal torus with this choice of \(J\). | {'timestamp': '2022-03-03T02:28:40', 'yymm': '2203', 'arxiv_id': '2203.01106', 'language': 'en', 'url': 'https://arxiv.org/abs/2203.01106'} |
null | null |