Fragmented exceptional points and their bulk and edge realizations in lattice models Subhajyoti Bid, Henning Schomerus Abstract Exceptional points (EPs) are spectral defects displayed by non-Hermitian systems in which multiple degenerate eigenvalues share a single eigenvector. This distinctive feature makes systems exhibiting EPs more sensitive to external perturbations than their Hermitian counterparts, where degeneracies are nondefective diabolic points. In contrast to these widely studied cases, more complex non-Hermitian degeneracies in which the eigenvectors are only partially degenerate are poorly understood. Here, we characterize these fragmented exceptional points (FEPs) systematically from a physical perspective, and demonstrate how they can be induced into the bulk and edge spectrum of two-dimensional and three-dimensional lattice models, exemplified by non-Hermitian versions of a Lieb lattice and a higher-order topological Dirac semimetal. The design of the systems is facilitated by an efficient algebraic approach within which we provide precise conditions for FEPs that can be evaluated directly from a given model Hamiltonian. The free design of FEPs significantly opens up a new frontier for non-Hermitian physics and expands the scope for designing systems with unconventional response characteristics. I Introduction In the past decade, topological phases in Hermitian systems, including topological insulators and superconductors Qi and Zhang 2011 ; Hasan and Kane 2010 ; Bernevig and Hughes 2013 ; Beenakker 2013 ; Asbóth et al. 2016 ; Sato and Ando 2017 as well as Dirac and Weyl semimetals Murakami 2007 ; Burkov 2016 ; Armitage et al. 2018 ; Yan and Felser 2017 have drawn substantial interest due to a wide range of possible applications. One of the key features of these topological materials is the existence of gapless edge and surface states that are protected by the underlying symmetries of the bulk Hamiltonian. Due to this protection, these states are immune to symmetry-preserving disorder present in the system. The investigation of such topological phases significantly expanded with the identification of higher-order topological insulators with quantized quadruple moments Benalcazar et al. 2017a ; Benalcazar et al. 2017b ; Schindler et al. 2018 ; Ghorashi et al. 2020a , including 3D higher-order Dirac semimetals (HODSM) Lin and Hughes 2018 ; Wieder et al. 2020 and 3D higher-order Weyl semimetals Ghorashi et al. 2020b ; Wang et al. 2020 ; Wei et al. 2021 . Unlike conventional topological insulators that display edge or surface states at their boundaries, these materials feature topological corner or hinge states when they are confined in two directions. With the recent resurgence and expansion of non-Hermitian physics Hatano and Nelson 1996 ; Moiseyev 2011 ; Bender and Boettcher 1998 ; Cao and Wiersig 2015 ; El-Ganainy et al. 2018 ; Ashida et al. 2020 , the study of topological materials has rapidly extended to their non-Hermitian equivalents, see for example Refs. Schomerus 2013a ; Malzard et al. 2015 ; Lieu 2018 ; Kawabata et al. 2019 ; Liu et al. 2019 ; Xue et al. 2020 ; Denner et al. 2021 ; Bid et al. 2023 ; Okuma and Sato 2023 . In the non-Hermitian setting, the band structure and energies of topologically protected states generally become complex, and the associated eigenstates are no longer constrained to be mutually orthogonal. One of the main resulting features is the emergence of distinct branch points in the complex energy eigenspectrum, known as exceptional points (EPs), where both the eigenenergies and eigenvectors coalesce Kato 1966 ; Dembowski et al. 2001 ; Heiss 2004 ; Heiss 2012 ; Berry 2004 . This behavior is distinct from that at diabolic points (DPs), the conventional degeneracies in Hermitian systems, where only the eigenenergies merge. As topological transitions of the band structure and the emergence of topologically protected states are linked to the closing of band gaps, and the bands and protected states are themselves now located in the complex energy plane, EPs significantly expand the scope of topological phases, and this is further enriched by an expanded set of possible symmetries Kawabata et al. 2019 . Exceptional points are a universal phenomenon in any non-Hermitian system. In the parameter space of the system, the locations of the EPs form manifolds, known as exceptional surfaces Xu et al. 2017 ; Carlström and Bergholtz 2018 ; Liu et al. 2021 ; Cerjan et al. 2019 , whose structure depends on the complexity of the EP, as well as constraints imposed by symmetries. These features result in directly observable phenomena, such as a nonlinear power-law response to parametric perturbations Wiersig 2014 ; Chen et al. 2017 ; Lai et al. 2019 ; Wiersig 2020a and a super-Lorentzian response to quantum noise and external driving Yoo et al. 2011 ; Schomerus 2020 ; Hashemi et al. 2022 . In both response settings, the qualitative and quantitative signatures Wiersig 2023 ; Wiersig 2022 ; Schomerus 2024 ; Bid and Schomerus 2025 depend on the order n of the EP, where an n th-order EP (short, an EP n ) is obtained by merging n eigenvalues, with the complete collapse of the eigenstates into a single one Demange and Graefe 2011 ; Ryu et al. 2012 ; Lin et al. 2016 ; Mandal and Bergholtz 2021 . EP n s have already been realized in various experiments demonstrating a significant increase in the system’s sensitivity Hodaei et al. 2017 ; Wu et al. 2021 ; Yang et al. 2023 ; Wiersig 2020b . Non-Hermitian degeneracies with only partially degenerate eigenstates have received much less attention even though they are as ubiquitous as higher-order EPs. Mathematically, these most general degeneracies can be classified by partial multiplicities that describe the fragmentation of the generalized eigenspace into the subspaces associated with the surviving eigenstates Hörmander and Melin 1994 . Just as higher-order EPs, these fragmented exceptional points (FEPs) naturally appear in specific positions on the exceptional surface Bergholtz et al. 2021 ; Sayyad et al. 2023 ; König et al. 2024 ; Ryu et al. 2024 . However, their appearance poses a significant challenge for the description and classification of non-Hermitian systems, as the standard perturbative approach, Arnold-Jordan normal form theory Arnold 1971 , breaks down for these scenarios. This state of affairs also hampers the systematic description and design of systems and the identification of the physical manifestations of FEPs. Therefore, obtaining a systematic handle on FEPs is a critical step to complete the description and understanding of non-Hermitian systems both conceptually as well as physically, and realize the full scope of their phenomenology. Motivated to overcome this roadblock, we establish in this work a systematic and efficient framework to design, classify, and characterize systems exhibiting FEPs. Furthermore, we apply this framework to address the interplay of non-Hermiticity and symmetry in specific two-dimensional and three-dimensional lattice models, and formulate exact analytical conditions for the formation of FEPs in bulk and boundary spectra. The framework directly links to the qualitatively different response of FEPs to perturbations and driving, and freely applies to any model based on an effective non-Hermitian Hamiltonian matrix. This delivers systematic algebraic conditions in terms of the so-called modes of the adjugate matrix, which can be directly evaluated for any given Hamiltonian matrix and yield explicit conditions in terms of microscopic model data. Given the physical starting point, the very same quantities directly determine the physical response of the system to external and parametric perturbations. For the models, we first consider the two-dimensional Lieb lattice and apply the framework to construct and classify non-Hermitian versions that host diverse FEPs with different degeneracy structures in the bulk dispersion. As a second model, we turn to a three-dimensional HODSM that supports topologically protected hinge states, and design variants that display FEPs in the bulk and boundary spectra. The FEP formation for hinge states reveals a highly intricate interplay of nonreciprocal non-Hermitian couplings and symmetry constraints, illuminating the richness of mechanisms determining FEPs. The general framework and specific applications presented here provide a template to construct and classify models with unconventional spectral and response features across a wide range of platforms, significantly expanding the scope to understand and design non-Hermitian physical systems and devices. The outline of the paper is as follows. After briefly recapitulating the role of DPs and EPs in the spectral analysis of Hermitian and non-Hermitian systems in Sec. II , we provide in Sec. III our efficient and systematic framework for a mathematical and physical description of FEPs. In Sec. 1 we apply this formalism to investigate the two-dimensional Lieb lattice, and use it to construct and classify non-Hermitian versions displaying FEPs of all admissible types in the bulk dispersion relation. In Sec. V we extend these consideration to a three-dimensional HODSM and demonstrate the existence of diverse FEPs in the bulk eigenspectra of appropriately designed non-Hermitian variants. In Sec. VI we discuss the hinge states in these HODSM variants, resulting in the emergence of fragmented exceptional lines with different degeneracy structures. We give our conclusions in Sec. VII . A key technical step in the derivation of the employed formalism is provided in Appendix A . II Exceptional and diabolic points To prepare our discussion of FEPs, we briefly review the notion of an EP of order (EP n ), which is the most commonly studied spectral degeneracy scenario of non-Hermitian physics, and contrast it to the notion of an n -bolic point, such as a diabolic point, which constitutes the spectral degeneracies of Hermitian physics. We place our results in the context of systems with an effective non-Hermitian Hamiltonian, represented by an N\times N dimensional square matrix matrix H . The eigenvalues are determined from the condition p(E)\equiv\mathrm{det}\,(E\openone-H)=0, (1) hence, are given by the roots of a polynomial p(E) of order N . The algebraic multiplicity \alpha_{i} of an eigenvalue E_{i} is the order of the root, and these multiplicities add up to \sum_{i}\alpha_{i}=N . Given an eigenvalue E_{i} , the geometric multiplicity \gamma_{i} determines the number of linearly independent right eigenvectors \mathbf{u}_{i,j} , j=1,2,\ldots,\gamma_{i} , fulfilling (E_{i}\openone-H)\mathbf{u}_{i,j}=0, (2) as well as the number of linearly independent left eigenvectors \mathbf{v}_{i,j} , j=1,2,\ldots,\gamma_{i} , fulfilling \mathbf{v}_{i,j}(E_{i}\openone-H)=0. (3) The geometric multiplicity is therefore determined by the rank of these homogeneous systems of equations, \gamma_{i}=N-\mathrm{rnk}\,(E_{i}\openone-H), (4) and is constrained according to 1\leq\gamma_{i}\leq\alpha_{i}. (5) For a nondefective system, the multiplicities \alpha_{i}=\gamma_{i} coincide for all eigenvalues. Hence, the number of linearly independent right (or left) eigenvectors \sum_{i}\gamma_{i}=N , so that each set of eigenvectors forms a basis. The system can then be diagonalized by a suitable similarity transformation \Lambda=VHU, (6) where \Lambda is a diagonal matrix containing the eigenvalues E_{i} of H according to their multiplicity, U contains the corresponding right eigenvectors \mathbf{u}_{i,j} as its columns, and V contains the corresponding left eigenvectors \mathbf{v}_{i,j} as its rows. Given that for a similarity transformation V=U^{-1} , the eigenvectors constructed in this way fulfill the biorthogonality condition \mathbf{v}_{i,j}\mathbf{u}_{i^{\prime},j^{\prime}}=\delta_{ii^{\prime}}\delta_{jj^{\prime}}. (7) Degenerate eigenvalues with \alpha_{i}=\gamma_{i}=2 are known as diabolic points (DPs), and higher order degeneracies where \alpha_{i}=\gamma_{i}=n with some n>2 can be referred to as n -bolic points. Analogously, maximally defective eigenvalues with \gamma_{i}=1 but \alpha_{i}=n with some n>1 are known as an n^{th} -order exceptional points (EP n s). At these EPs, n algebraically degenerate eigenvalues share a single, unique, eigenvector. The existence of EPs therefore renders a system nondiagonalizable, leading to the distinct physical signatures in the response to driving and perturbations surveyed in the introduction. III Fragmented exceptional points In between the n -bolic point and the EP n , a degenerate eigenvalue E_{i} can have several linearly independent eigenvectors, corresponding to FEPs with a geometric multiplicity \alpha_{i}>\gamma_{i}>1 . We first detail the mathematical and physical signatures of these most general spectral scenarios, and then provide an efficient algebraic framework by which we can determine these signatures directly from model data. III.1 Partial multiplicities For degenerate eigenvalues with \alpha_{i}>\gamma_{i}>1 one can define partial multiplicities l_{i,1}\geq l_{i,2}\geq l_{i,3}\dots\geq l_{i,\gamma_{i}}\geq 1 that designate how many algebraically degenerate eigenvalues are associated with a given eigenvector. These partial multiplicities therefore partition the algebraic multiplicity according to l_{i,1}+l_{i,2}+l_{i,3}+\dots+l_{i,\gamma_{i}}=\alpha_{i}, (8) where each partition of \alpha_{i} describes a different spectral scenario. Furthermore, by including the cases of a single partial multiplicity l_{i,1}=\alpha_{i} (EP n with n=\alpha_{i} ) and the case of all l_{i,1}=l_{i,2}=\dots=l_{i,\gamma_{i}}=1 ( n -bolic point with n=\gamma_{i} ), the specification of the partial multiplicities covers all possible degeneracy scenarios. The frequency with which a given partial multiplicity l repeats in the sequence l_{i,j} defines the partial multiplicity function \beta_{i}(l) . While the partial multiplicities l_{i,j} partition \alpha_{i} according to Eq. ( 8 ), and hence are of an algebraic nature, the values of \beta_{i}(l) partition the geometric multiplicity according to \sum_{l=1}^{N}\beta_{i}(l)=\gamma_{i}. (9) These values can hence be interpreted as geometric partial multiplicities, which count how many independent eigenvectors can be associated with the partial multiplicity l . Furthermore, Eq. ( 8 ) itself can be cast into the form \sum_{l=1}^{N}l\beta_{i}(l)=\alpha_{i}. (10) Thereby, the partial multiplicity function \beta_{i}(l) completely characterizes the algebraic and geometric structure of a given degeneracy in a convenient way. For an EP n , only \beta_{i}(n)=1 is finite, while for an n -bolic point, only \beta_{i}(1)=n is finite. For the illustrative example of an FEP with algebraic partial multiplicities (4,4,3,2,2,1), the finite geometric partial multiplicities are \beta_{i}(1)=1 , \beta_{i}(2)=2 , \beta_{i}(3)=1 , and \beta_{i}(4)=2 , which corresponds further to an algebraic multiplicity \alpha_{i}=16 and a geometric multiplicity \gamma_{i}=6 . III.2 Geometric interpretation In concrete geometric terms, the value \beta_{i}(l) of the partial multiplicity function designates the dimensionality of the eigenspace associated with a partial multiplicity l . In the assignment of these spaces themselves, there exists a gauge freedom. Uniquely defined is only the space spanned by the eigenvectors of maximal partial multiplicity Trefethen 1999 ; Bid and Schomerus 2025 l_{i,1}\equiv\ell_{i}. (11) The eigenvectors with maximal partial multiplicity are called the leading eigenvectors, and we will denote their geometric multiplicity \beta_{i}(\ell_{i}) more concisely by \beta_{i} , so that \ell_{i}=l_{i,1}=l_{i,2}=\ldots=l_{i,\beta_{i}} . For sectors with a smaller partial multiplicity, the eigenspace is defined modulo addition of eigenvectors from sectors with a larger partial multiplicity. This freedom can be exploited to orthogonalize the subspaces of different partial multiplicity. The underlying hierarchical structure of these subspaces is also directly reflected in the physical signatures of FEPs, to which we turn next. III.3 Physical signatures The intricate features of FEPs are directly reflected in their physical signatures. Examining these signatures will provide the key to their identification and classification in a systematic framework. In particular, the response to external driving and parametric changes are both dominated by the sector of maximal partial multiplicity \ell_{i} Trefethen 1999 ; Graefe et al. 2008 ; Bid and Schomerus 2025 : Physical response to driving. At an FEP, the Green’s function G(E)=(E\openone-H)^{-1}, (12) which describes the response of the system to external driving, diverges according to P(E)\equiv\mathrm{tr}\,[G^{\dagger}(E)G(E)]\propto\frac{|\eta_{i}|^{2}}{|E-E_{i}|^{2\ell_{i}}}, (13) where \eta_{i} is a characteristic physical response strength. Spectral response to parameter changes. Upon a perturbation of strength \varepsilon , the degenerate eigenvalues split up into a multiplet, and become displaced by a maximal amount \varepsilon E_{i}\sim(\varepsilon\xi_{i})^{1/\ell_{i}}, (14) where \xi_{i} is a characteristic spectral response strength. Both response strengths are natural extensions of the behaviour at conventional EP n s Yoo et al. 2011 ; Hashemi et al. 2022 ; Wiersig 2023 ; Wiersig 2022 ; Schomerus 2024 ; Bid and Schomerus 2025 . The explicit expressions of these response strengths can be obtained in two approaches. With exact prior knowledge of all eigenvalues and their partial multiplicity structure, the system can be brought into a Jordan normal form, H=TJT^{-1} , where T is a similarity transformation. Expressed in this way, the system can then be analyzed in Arnold-Jordan perturbation theory Arnold 1971 . However, this perturbation theory breaks down for FEPs, while more generally the Jordan normal form construction is ill conditioned and changes singularly under infinitesimal parameter changes. This route can therefore only be carried out in systems for which the spectral properties are exactly known. The second approach Bid and Schomerus 2025 , which we adopt here, starts with a systematic expansion of the Green’s function around an arbitrary references energy \Omega , G(E)=\frac{\sum_{k=0}^{N-1}(E-\Omega)^{k}\mathcal{B}_{k}}{\sum_{k=0}^{N}(E-\Omega)^{k}c_{k}}. (15) The quantities \mathcal{B}_{k} in the numerator are N\times N matrices known as the modes of the adjugate matrix . These modes follow directly from the Hamiltonian matrix by the Faddeev-LeVierre recursion relation \displaystyle\mathcal{B}_{N-1}=\openone,\quad \displaystyle\mathcal{B}_{k-1}=A\mathcal{B}_{k}-\frac{\mathrm{tr}\,(A\mathcal{B}_{k})}{N-k}\openone, (16) where A=H-\Omega\openone . The modes also determine the coefficients c_{k}=-\frac{\mathrm{tr}\,(A\mathcal{B}_{k})}{N-k} (17) in the denominator, which are the expansion coefficients of the shifted characteristic polynomial p(E)=\mathrm{det}\,(E\openone-H)=\mathrm{det}\,(\lambda\openone-A)=\sum_{k=0}^{N}\lambda^{k}c_{k}\equiv q(\lambda), (18) where \lambda=E-\Omega . In terms of this directly calculable data, the response strengths are then given by Bid and Schomerus 2025 \eta_{i}^{2}=\frac{\tr(\mathcal{B}_{\alpha_{i}-\ell_{i}}^{\dagger}\mathcal{B}_{\alpha_{i}-\ell_{i}})}{|c_{\alpha_{i}}|^{2}} (19) and \xi_{i}^{2}=\frac{||\mathcal{B}_{\alpha_{i}-\ell_{i}}||_{2}^{2}}{|c_{\alpha_{i}}|^{2}}, (20) where ||\cdot||_{2} denotes the spectral norm, and the modes are calculated with \Omega=E_{i} , hence A=H-E_{i}\openone . This algebraic approach based on the modes \mathcal{B}_{k} therefore enables us to calculate the response strengths reliably from the Hamiltonian, without having to carry out the Jordan decomposition. As we will show next, the data entering the response strengths directly links to the partial multiplicity structure of FEPs (as a consequence of this connection, the two response strengths \eta_{i} and \xi_{i} happen to coincide for \beta_{i}=1 ). III.4 Framework for efficient analysis and design As we describe now, the algebraic approach to determine the response strengths can be extended into an efficient mathematical framework that identifies and characterizes FEPs systematically and completely, and delivers explicit conditions for their formation that can be evaluated in specific models. To formulate this framework, we again set A=H-E_{i}\openone , and then consider the modes \mathcal{B}_{k} as determined by the Faddeev-LeVierre recursion relation ( 16 ). The main vehicle of this characterization are two conditions. The eigenvalues and their algebraic multiplicity are obtained from the condition C_{k}\equiv\mathrm{tr}\,(A\mathcal{B}_{k})=0,\quad k=0,1,2,\ldots,\alpha_{i}-1, (21) originally derived in Ref. Bid and Schomerus 2025 and reviewed in Appendix A . The second condition resolves the partial multiplicities according to \beta_{i}(l)=\mathrm{rnk}\,(\mathcal{B}_{\alpha_{i}-l-2})-2\times\mathrm{rnk}\,(\mathcal{B}_{\alpha_{i}-l-1})+\mathrm{rnk}\,(\mathcal{B}_{\alpha_{i}-l}), (22) with l\geq 1 . We derive this additional condition, which is central to the present work, in the same Appendix A . Together, these conditions deliver the partial multiplicity function of an eigenvalue directly in terms of the shifted Hamiltonian A . From this information, we can reconstruct the other multiplicities as already described in subsection III.1 . Furthermore, the maximal partial multiplicity can be determined directly by the condition \mathcal{B}_{k}=0,\quad k=0,1,2,\ldots,\alpha_{i}-\ell_{i}-1, (23) so that the first finite mode is given by \mathcal{B}_{\alpha_{i}-\ell_{i}} , i.e., exactly by the mode that enters the algebraic expressions ( 19 ) and ( 20 ) for the response strengths \eta_{i} and \xi_{i} , respectively. The geometric multiplicity \beta_{i} of the leading eigenvectors is then given by the rank of this matrix, \beta_{i}=\mathrm{rnk}\,(\mathcal{B}_{\alpha_{i}-\ell_{i}}), (24) which results in the above-mentioned identity of the spectral strengths ( 19 ) and ( 20 ) if \beta_{i}=1 . All the algebraic conditions stated in this subsection are basis invariant, and hence can be evaluated directly from model Hamiltonians, where one can make use of the efficient recursion relation ( 16 ). It follows that the conditions not only serve to systematically characterize a given degenerate eigenvalue, but also can be employed to design systems that exhibit a certain desired degeneracy scenario. With this, we are now ready to turn to specific lattice models and determine the conditions under which they exhibit FEPs in the bulk and boundary spectrum. IV Non-Hermitian Lieb lattices Our main focus for the remainder of this work is to demonstrate the existence of FEPs in two- and three-dimensional lattice models and to systematically incorporate these degeneracy scenarios into the body of widely studied cases of generic EPs with geometric multiplicity \gamma_{i}=1 . We start with the Lieb lattice, a tight-binding model based on a unit cell of three sites, corresponding to three sublattices that we denote as A, B, and C (see Fig. 1 ). This provides the minimal setting of a system with three bands, which, hence, can display bulk degeneracies of algebraic multiplicity \alpha=3 . These third-order degeneracy scenarios serve as our motivation for considering this model, as n=3 is the first nontrivial order that can harbor an FEP, with partial multiplicities (l_{i,1},l_{i,2})=(2,1) . In contrast, for second-order degeneracies, there are only two scenarios: a DP within the Hermitian framework and a generic EP2 within the non-Hermitian framework. The original Hermitian version of the Lieb lattice displays a flat band as well as two symmetric dispersive bands that meet in an n -bolic point of order n=3 (a tribolic point) Lieb 1989 ; Weeks and Franz 2010 ; Goldman et al. 2011 , while non-Hermitian variants have been designed to display EP n ’s of the same order in their bulk band structure Demange and Graefe 2011 ; Mandal and Bergholtz 2021 . We first review these common degeneracy scenarios in the standard approach and then design a minimal model exhibiting an FEP. Subsequently, we apply our formalism from the previous section to systematically classify these degeneracy scenarios and identify the general conditions for Lieb lattices that support FEPs in their bulk band structure. Finally, we utilize these conditions to design a system that displays exceptional rings and lines hosting FEPs at special locations. Figure 1: (a) Standard Hermitian version of the Lieb lattice, based on a three-site unit cell with uniform couplings t . (b) Non-Hermitian version of the model, with nonreciprocal couplings p,q,r,s inside the unit cell, while the couplings between the cells remain fixed to t . This modification allows us to realize fragmented exceptional points in the bulk dispersion of the model. IV.1 Hermitian model and tribolic points In its original Hermitian formulation, the nearest neighbors in the Lieb lattice are coupled by a uniform, real, coupling constant t [see Fig. 1 (a)]. Setting t=a=1 , where a is the lattice constant, the infinite periodic system is described by the Bloch Hamiltonian \mathcal{H}(\mathbf{k})=\left(\begin{array}[]{ccc}0&1+e^{ik_{y}}&0\\ 1+e^{-ik_{y}}&0&1+e^{-ik_{x}}\\ 0&1+e^{ik_{x}}&0\\ \end{array}\right), (25) where \mathbf{k}=(k_{x},k_{y}) . The Bloch Hamiltonian reflects the lattice topology of the system, in which the A and C sites are only coupled to B sites, and vice versa. This constraint on the couplings induces a chiral symmetry \mathcal{X}\mathcal{H}(k)\mathcal{X}=-\mathcal{H}(k), (26) where \mathcal{X}=\left(\begin{array}[]{ccc}1&0&0\\ 0&-1&0\\ 0&0&1\\ \end{array}\right). (27) The chiral symmetry maps any eigenstate \mathbf{u}_{i}(\mathbf{k}) with eigenvalue E_{i}(\mathbf{k}) to an eigenstate \mathcal{X}\mathbf{u}_{i}(\mathbf{k}) with eigenvalue -E_{i}(\mathbf{k}) . Therefore, the band structure is comprised of two symmetric dispersive bands E_{+}(\mathbf{k})=-E_{-}(\mathbf{k}) , given by E_{\pm}(\mathbf{k})=\pm\sqrt{2\left(\cos k_{x}+\cos k_{y}+2\right)}. (28) By the same argument, the remaining third eigenvalue is forced to vanish throughout the Brillouin zone, E_{0}(\mathbf{k})=0\mbox{ for all $\mathbf{k}$}, (29) hence, forms a dispersionless flat band. A key characteristic of this flat band is its localization on the AC sublattice, while the dispersive bands exhibit equal weights on the B and AC parts of the system. Figure 2: Contour plots of the dispersion relation |E(\mathbf{k})| for (a) the conventional Hermitian Lieb lattice with Bloch Hamiltonian ( 25 ), depicting a tribolic point at the momenta values (k_{x},k_{y})=(\pi,\pi) around which the dispersion relation is linear, and (b) the non-Hermitian variant with Bloch Hamiltonian ( 30 ), proposed in Ref. Mandal and Bergholtz 2021 , displaying four conventional EP 3 s with geometric multiplicity \gamma=1 arranged symmetrically around the point where the tribolic point existed for the Hermitian case. The parameter \varepsilon determining the strength of the non-Hermitian perturbation is set to be equal to 1. We design model systems that exhibit the more complex FEP degeneracy scenarios in their energy dispersion. At k_{x}=k_{y}=\pi , the radicand in Eq. ( 28 ) vanishes, and the three bands merge at energy E_{0}=E_{\pm}=0 , i.e., form a degeneracy with algebraic multiplicity \alpha=3 . Furthermore, since the underlying system is Hermitian, it maintains three linearly independent eigenvectors, which then span the complete Hilbert space. Therefore, the geometric multiplicity is \gamma=3 , and the degeneracy is a tribolic point. This scenario is illustrated in Fig. 2 (a) in terms of a contour plot of the absolute value |E_{\pm}(\mathbf{k})| of the dispersive bands. IV.2 Exceptional points of order three As pointed out in Refs. Demange and Graefe 2011 and Mandal and Bergholtz 2021 , introducing a suitable non-Hermitian perturbation into the Hamiltonian while preserving the chiral symmetry, the tribolic point can be split into four 3^{rd} -order exceptional points (EP3s), each characterized by an algebraic multiplicity of \alpha=3 and a geometric multiplicity of \gamma=1 . The simplest variant is given by Mandal and Bergholtz 2021 \mathcal{H}(\textbf{k})=\left(\begin{array}[]{c@{\,\,}c@{\,\,}c}0&1+e^{ik_{y}}+i\varepsilon&0\\ 1+e^{-ik_{y}}+i\varepsilon&0&1+e^{-ik_{x}}-i\varepsilon\\ 0&1+e^{ik_{x}}-i\varepsilon&0\\ \end{array}\right), (30) which preserves the chiral symmetry ( 26 ) as well as reciprocity \mathcal{R} , defined by the condition \mathcal{H}(\textbf{k})=\mathcal{H}^{T}(-\textbf{k}) . The flat band is maintained, while the dispersive bands become E_{\pm}(\mathbf{k})=\pm\sqrt{2\left[(1-i\varepsilon)\cos k_{x}+(1+i\varepsilon)\cos k_{y}-\varepsilon^{2}+2\right]}. (31) As shown in Fig. 2 (b), the four EP3s are symmetrically located in the Brillouin zone at momentum values k_{x}=\mu k_{0} , k_{y}=\nu k_{0} , where k_{0}=\arccos\left(\frac{\varepsilon^{2}}{2}-1\right) and \mu,\nu=\pm 1 . At each of these EP3s, there is only a single pair of right and left eigenvectors, whose nonnormalized form is given by \displaystyle\mathbf{u}_{\mu,\nu}=\left(\begin{array}[]{c}i+e^{i\nu\varphi_{\varepsilon}}\\ 0\\ i-e^{-i\mu\varphi_{\varepsilon}}\end{array}\right),\quad\mathbf{v}_{\mu,\nu}=\mathbf{u}_{-\mu,-\nu}^{T}, where \varphi_{\varepsilon}=\arccos\frac{\varepsilon}{2} . As a result, the geometric multiplicity is \gamma=1 , which is a signature of an EP. Indeed, each eigenvector pair is self-orthogonal, \mathbf{v}\mathbf{u}=0 , which is a symptom of the defectiveness of the eigensystem Kato 1966 . Figure 3: (a) Energy band diagram E(k_{x},k_{y}=\pi) as a function of k_{x} for the conventional Hermitian Lieb lattice with Bloch Hamiltonian ( 25 ), displaying the tribolic point at k_{x}=\pi . (b,c) Complex energy band diagrams (real part in blue, imaginary part in red) for the non-Hermitian Lieb lattice with Bloch Hamiltonian ( 36 ) and \varepsilon=1 . In (b) k_{y}=\pi , revealing the FEP of geometric multiplicity \gamma=2 and partial multiplicities (l_{1},l_{2})=(2,1) at k_{x}=\pi . In (c), k_{y}=-2\arccot\frac{1}{2} is fixed to display the generic EP3 in this model, see Eq. ( 39 ). Panel (d) reveals the distinct dispersion around the FEP and EP3 in this model in terms of the energy contours |E_{+}(\mathbf{k})| . In the low energy limit, the generic EP3 displays an isotropic dispersion relation, while the dispersion relation around the FEP is highly anisotropic. IV.3 Fragmented exceptional points To prepare our general classification in the next section, we now identify a minimal example of a non-Hermitian Lieb lattice that displays an FEP. As the system only contains three states, the FEP will have partial multiplicities (l_{1},l_{2})=(2,1) , corresponding to an algebraic multiplicity of \alpha=3 and a geometric multiplicity of \gamma=2 . Therefore, the maximal partial multiplicity is \ell=2 . In the design of the system, we are then guided by the general observation that this degeneracy structure lies in between those of a tribolic point and an EP3. This leads us to the following choice of a minimal example, \mathcal{H}(\mathbf{k})=\left(\begin{array}[]{ccc}0&1+e^{ik_{y}}+i\varepsilon&0\\ 1+e^{-ik_{y}}&0&1+e^{-ik_{x}}\\ 0&1+e^{ik_{x}}-i\varepsilon&0\\ \end{array}\right). (36) We observe that this model preserves the chiral symmetry, but breaks reciprocity, \mathcal{H}^{T}(\mathbf{k})\neq\mathcal{H}(-\mathbf{k}) . The flat band E_{0}=0 remains intact, while the dispersive bands are given by E_{\pm}(\mathbf{k})=\pm\sqrt{4+2\cos k_{x}+2\cos k_{y}+i\varepsilon(e^{-ik_{y}}-e^{-ik_{x}})}. (37) Solving this relation for E_{\pm}(\mathbf{k})=0 gives us two band degeneracy points in the Brillouin zone, which are located at k_{x}=k_{y}=\pi (38) and k_{x}=-k_{y}=2\arccot\frac{\varepsilon}{2}. (39) Therefore, unlike for the model ( 30 ), the tribolic point now breaks into two degeneracy points, rather than four. Of these, the point at k_{x}=k_{y}=\pi possesses two pairs of linearly independent right and left eigenvectors, which we may choose as \displaystyle\mathbf{u}_{1}=\begin{pmatrix}0\\ 0\\ 1\end{pmatrix},\quad\mathbf{u}_{2}=\begin{pmatrix}1\\ 0\\ 0\end{pmatrix}, \displaystyle\mathbf{v}_{1}=\begin{pmatrix}1,0,1\end{pmatrix},\quad\mathbf{v}_{2}=\begin{pmatrix}0,1,0\end{pmatrix}. (40) This point is the desired FEP. On the other hand, the point at k_{x}=-k_{y}=2\arccot\frac{\varepsilon}{2} forms a generic EP3 with \gamma=1 . Consequently, two third-order degeneracy points with different geometric multiplicities coexist in the bulk dispersion of the Hamiltonian ( 36 ). Figure 3 illustrates the signatures of the different degeneracy scenarios in cross-sections of bulk band diagrams. Panel (a) shows the tribolic point of the Hermitian system, while panels (b) and (c) show the FEP and EP3 in the model ( 36 ), respectively. A characteristic feature of the FEP is the cusp-like shapes displayed by either the real or the imaginary parts as one approaches the degeneracy from either side. As shown in contour plots of the dispersive bands in panel (d), the dispersion around the FEP is highly anisotropic. Based on such design-by-inspection, we can come up with further examples of Lieb lattices that display FEPs. However, instead of listing further such cases here, we next apply our formalism to classify the possible degeneracy scenarios systematically and completely. Afterwards, we revisit the examples given so far, and describe how they can be generalized. IV.4 Complete classification We now utilize our framework based on the modes \mathcal{B}_{k} to embed the examples given so far into a full classification of the possible degeneracy scenarios in the bulk dispersion relation of the Lieb lattice, and determine the general conditions for FEPs. To carry out this classification in sufficient generality, we consider a Bloch Hamiltonian of the form \mathcal{H}(\textbf{k})=\left(\begin{array}[]{ccc}0&P&0\\ Q&0&R\\ 0&S&0\\ \end{array}\right), (41) where P , Q , R , and S are functions of \mathbf{k} . This maintains the chiral symmetry of the system, so that all degeneracies occur at energy E=0 . We therefore apply the modal analysis with a vanishing energy shift \Omega=0 , so that the matrix A=\mathcal{H}(\textbf{k}) . IV.4.1 Modes and their rank Applying the Faddeev-LeVierre recursion relation ( 16 ) we obtain \displaystyle\mathcal{B}_{2} \displaystyle=\openone, \displaystyle\mathcal{B}_{1} \displaystyle=\mathcal{H}(\textbf{k}) \displaystyle=\left(\begin{array}[]{ccc}0&P&0\\ Q&0&R\\ 0&S&0\\ \end{array}\right), \displaystyle\mathcal{B}_{0} \displaystyle=\mathcal{H}(\textbf{k})^{2}-\frac{\mathrm{tr}\,\mathcal{H}(\textbf{k})^{2}}{2}\openone \displaystyle=\left(\begin{array}[]{ccc}-RS&0&PR\\ 0&0&0\\ QS&0&-PQ\\ \end{array}\right). To apply our formalism, we need to determine the rank of these matrices. This analysis is facilitated by the fact that we can write \mathcal{B}_{0}=\left(\begin{array}[]{c}R\\ 0\\ -Q\end{array}\right)\left(-S,0,P\right). (48) The ranks can then be read off directly from the simple form of these matrices, where we distinguish three cases: • CASE 1 : If (|P|+|S|)(|Q|+|R|)>0 (hence, neither P and S both vanish at the same time, nor Q and R both vanish at the same time), then \displaystyle\rnk\mathcal{B}_{0}=1,\quad\rnk\mathcal{B}_{1}=2,\quad\rnk\mathcal{B}_{2}=3. (49) • CASE 2 : If (|P|+|S|)(|Q|+|R|)=0 but |P|+|Q|+|R|+|S|>0 (where the latter means that at least one of the quantities is finite) then \displaystyle\rnk\mathcal{B}_{0}=0,\quad\rnk\mathcal{B}_{1}=1,\quad\rnk\mathcal{B}_{2}=3. (50) • CASE 3 : If P=Q=R=S=0 , then \displaystyle\rnk\mathcal{B}_{0}=0,\quad\rnk\mathcal{B}_{1}=0,\quad\rnk\mathcal{B}_{2}=3. (51) IV.4.2 Algebraic multiplicity Next, we determine the algebraic multiplicity of the eigenvalue from condition ( 21 ). For this, we evaluate \displaystyle C_{0} \displaystyle=\mathrm{tr}\,(\mathcal{H}(\textbf{k})\mathcal{B}_{0})=0, (52) \displaystyle C_{1} \displaystyle=\mathrm{tr}\,(\mathcal{H}(\textbf{k})\mathcal{B}_{1})=2(PQ+RS), (53) \displaystyle C_{2} \displaystyle=\mathrm{tr}\,(\mathcal{H}(\textbf{k})\mathcal{B}_{2})=0. (54) According to condition ( 21 ), the algebraic degeneracy is \alpha=1 if PQ+RS\neq 0 , in which case the eigenvalue is the nondegenerate eigenvalue of the flat band. Furthermore, the algebraic degeneracy is \alpha=3 if PQ+RS=0. (55) We recover that degeneracies of order 2 are ruled out, as enforced by the symmetries of the system—but did not need to explicitly envoke these symmetries. IV.4.3 Partial multiplicity With this we come to the last step of our general analysis, the determination of the partial multiplicities of the degeneracy. For this, we apply the main general result of this work, Eq. ( 22 ), to the matrices given in Eq. ( IV.4.1 ), whose ranks are specified in Eqs. ( 49 ), ( 50 ), and ( 51 ). For completeness, we first discuss the nondegenerate scenario, and then turn to the classification and characterization of the different degenerate scenarios. For the nondegenerate scenario, PQ\neq-RS , P and S cannot both vanish at the same time, and R and Q cannot both vanish at the same time. This corresponds to CASE 1, where the ranks are given by Eq. ( 49 ). Since in this scenario \alpha=1 , the multiplicity function ( 22 ) reduces to the single value \beta(1)=\mathrm{rnk}\,\mathcal{B}_{0}=1, (56) while formally \beta(2)=\beta(3)=0 . These values correspond to \gamma=1 and a single partial multiplicity l_{1}=1 , and thereby reproduce the only allowed values for the multiplicity function of a nondegenerate eigenvalue. In the degenerate scenarios, where PQ+RS=0 and \alpha=3 , the multiplicity function ( 22 ) follows from the expressions \displaystyle\beta(1) \displaystyle=\mathrm{rnk}\,\mathcal{B}_{0}-2\times\mathrm{rnk}\,\mathcal{B}_{1}+\mathrm{rnk}\,\mathcal{B}_{2}, \displaystyle\beta(2) \displaystyle=-2\times\mathrm{rnk}\,\mathcal{B}_{0}+\mathrm{rnk}\,\mathcal{B}_{1}, \displaystyle\beta(3) \displaystyle=\mathrm{rnk}\,\mathcal{B}_{0}. (57) In its evaluation, we distinguish the following cases. • CASE 1 : (|P|+|S|)(|Q|+|R|)>0 . The rank of the modes is the same as in the nondegenerate case, i.e., given by Eq. ( 49 ). Inserting these values into Eq. ( 57 ) we then find \displaystyle\beta(1)=\beta(2)=0,\quad\beta(3)=1. (58) These values correspond to an EP3, with a single partial multiplicity l_{i,1}=3 and geometric multiplicity \gamma=1 . • CASE 2 : (|P|+|S|)(|Q|+|R|)=0 but |P|+|Q|+|R|+|S|>0 . Inserting the ranks from Eq. ( 50 ) into Eq. ( 57 ) we find \displaystyle\beta(1)=\beta(2)=1,\quad\beta(3)=0. (59) This corresponds to an FEP with partial multiplicities (l_{1},l_{2})=(2,1) and a geometric multiplicity \gamma=2 . • CASE 3 : P=Q=R=S=0 . This is the tribolic point. The ranks are given by Eq. ( 51 ), from which we obtain \beta(1)=3,\quad\beta(2)=\beta(3)=0. (60) Consequently, the partial multiplicities are given by (l_{1},l_{2},l_{3})=(1,1,1) . Furthermore, the sum rule ( 9 ) recovers the geometric multiplicity \gamma=3 . IV.4.4 Exact general conditions for FEPs Therefore, FEPs are realized exactly when either (i) the quantities P and S both vanish and at least one of the quantities Q and R is finite, or (ii) the quantities Q and R both vanish and at least one of the quantities P and S is finite. We note that the degeneracy condition PQ+RS=0 , Eq. ( 55 ), is automatically fulfilled in both cases. IV.5 Examples revisited To connect these general findings to the earlier examples, we consider a Lieb lattice with nonreciprocal couplings inside the unit cell, as shown in Fig. 1 (b). This modifies the intracell couplings between the A and B sites to take values p (from B to A) and q (from A to B), and the intracell couplings between the B and C sites to take values r (from C to B) and s (from B to C). The intercell couplings maintain their reciprocal value t , which we again set to t\equiv 1 . The Bloch Hamiltonian then takes the form of Eq. ( 41 ) with \displaystyle P \displaystyle=p+e^{ik_{y}}, \displaystyle Q \displaystyle=q+e^{-ik_{y}}, \displaystyle R \displaystyle=r+e^{-ik_{x}}, \displaystyle S \displaystyle=s+e^{ik_{x}}. (61) Figure 4: Contour plots of the dispersion relation |E(\mathbf{k})| for the reciprocal non-Hermitian Lieb lattice model determined by the Bloch Hamiltonian ( 64 ) for different values of the parameters \Psi and \Phi . In panel (a), we set \Psi=3\pi/4 and \Phi=\pi/2 , which realizes two FEPs at (k_{x},k_{y})=(-3\pi/4,-\pi/2) and (k_{x},k_{y})=(3\pi/4,\pi/2) along with two generic EP3s at (k_{x},k_{y})=(-3\pi/4,\pi/2) and (k_{x},k_{y})=(3\pi/4,-\pi/2) . In panel (b), we set \Phi=\Psi=\pi/4 , leading to a degenerate ring determined by the momentum relation in ( 65 ). On this ring, two FEPs are located at (k_{x},k_{y})=(\pi/4,\pi/4) and (k_{x},k_{y})=(-\pi/4,-\pi/4) , while the remainder is formed from generic EP3s. Setting the parameters \Phi=\Psi=\pi/2 in panel (c), the exceptional ring morphs into four exceptional lines, speicified in Eq. ( 66 ). Similar to before, the lines host two FEPs located at (k_{x},k_{y})=(\pi/2,\pi/2) and (k_{x},k_{y})=(-\pi/2,-\pi/2) , while the remainder of the lines is formed from generic EP3s. (1) For the original Hermitian version of the Lieb lattice, with Bloch Hamiltonian ( 25 ), p=q=r=s=1 . At the tribolic point k_{x}=k_{y}=\pi , P=Q=R=S=0 , in agreement with CASE 3. We also note that these values fulfill the degeneracy condition Eq. ( 55 ) trivially. (2) For the non-Hermitian system ( 30 ), p=q=1+\varepsilon and r=s=1-\varepsilon . The EP3s are realized at \displaystyle P \displaystyle=\varepsilon(e^{i\mu\varphi_{\varepsilon}}-i) \displaystyle Q \displaystyle=\varepsilon(e^{-i\mu\varphi_{\varepsilon}}-i), \displaystyle R \displaystyle=\varepsilon(e^{-i\nu\varphi_{\varepsilon}}+i) \displaystyle S \displaystyle=\varepsilon(e^{i\nu\varphi_{\varepsilon}}+i), (62) where again \varphi_{\varepsilon}=\arccos\frac{\varepsilon}{2} . These expressions fulfill Eq. ( 55 ) nontrivially, and as they are all finite adhere to CASE 1. (iii) For our minimal example ( 36 ), p=1+\varepsilon , q=r=1 , and s=1-\varepsilon . At the EP3, with k_{x}=-k_{y}=2\arccot\frac{\varepsilon}{2} as specified in Eq. ( 39 ), the quantities P,Q,R,S are again all finite and fulfill Eq. ( 55 ) nontrivially, still in agreement with CASE 1. At k_{x}=k_{y}=\pi , Q=R=0 while P and S are finite. Accordingly, this realizes CASE 2, and hence fulfills the conditions for an FEP. It could now be anticipated that FEPs may arise for various situations in between the trivial and nontrival ways to fulfill the degeneracy condition PQ+RS=0 , Eq. ( 55 ), of which there are four natural candidates: (i) \displaystyle P=R=0, (ii) \displaystyle P=S=0, (iii) \displaystyle Q=R=0, (iv) \displaystyle Q=S=0, (63) with the other two quantities in each case assumed to be finite (the same assumption rules out the candidates P=Q=0 and R=S=0 , as the degeneracy condition then cannot be fulfilled). However, our general analysis shows that only candidates (ii) and (iii) result in FEPs according to the conditions of CASE 2, while candidates (i) and (iv) produce additional examples of EP3s obeying the conditions of CASE 1. IV.6 Reciprocal example with exceptional lines With these insights in mind, we conclude this section by designing a system that combines different scenarios, including for the case of exceptional lines. For this we consider the following model, \mathcal{H}(\mathbf{k})=\left(\begin{array}[]{ccc}0&e^{ik_{y}}-e^{i\Phi}&0\\ e^{-ik_{y}}-e^{i\Phi}&0&e^{-ik_{x}}-e^{i\Psi}\\ 0&e^{ik_{x}}-e^{i\Psi}&0\\ \end{array}\right), (64) with real angles \Phi and \Psi . Note that in contrast to the two specific non-Hermitian models ( 30 ) and ( 36 ) described above, this model is reciprocal, \mathcal{H}(\mathbf{k})=\mathcal{H}^{T}(\mathbf{-k}) . To identify the different degeneracy scenarios realized by this model, we refer to the exact general conditions for FEPs listed at the end of Sec. IV.4 . When the angle fulfill \Phi\neq\Psi and \Phi,\Psi\not\equiv 0\pmod{\pi} , the model displays an FEP with P=S=0 at (k_{x},k_{y})=(\Psi,\Phi) , an FEP with Q=R=0 at (k_{x},k_{y})=(-\Psi,-\Phi) , an EP3 with Q=S=0 at (k_{x},k_{y})=(\Psi,-\Phi) , and an EP3 with P=R=0 at (k_{x},k_{y})=(-\Psi,\Phi) . This scenario is illustrated in terms of an absolute-energy band diagram in Fig. 4 (a). For the case \Phi=\Psi\not\equiv\pi/2\pmod{\pi} illustrated in Fig. 4 (b), on the other hand, we find that the bulk eigenspectrum hosts an exceptional ring of algebraically three-fold degenerate eigenvalues located in the Brillouin zone. In accordance with the degeneracy condition PQ+RS=0 , Eq. ( 55 ), this ring follows the momenta points fulfilling the relation \cos k_{x}+\cos k_{y}=2\cos\Phi. (65) As in the previous case, the system displays FEPs at the special positions (k_{x},k_{y})=(\Phi,\Phi) , where with P=S=0 , and (k_{x},k_{y})=(-\Phi,-\Phi) , where with Q=R=0 . All the other degenerate eigenvalues on the ring satisfying the above momentum relation are generic EP3s. Finally, for the case \Phi=\Psi\equiv\pi/2\pmod{\pi} , Eq. ( 65 ) reduces to k_{y}=\pm k_{x}. (66) The exceptional rings are hence morph into straight lines on which all the eigenvalues are algebraically three-fold degenerate, where the special locations at (k_{x},k_{y})=(\pi/2,\pi/2) , and (k_{x},k_{y})=(-\pi/2,-\pi/2) again constitute FEPs, while all other locations on this line are generic EP3s. This scenario is illustrated in Fig. 4 (c). This example illustrates how our framework, based on the modes of the adjugate matrix, systematically identifies the precise and explicit conditions that result in the emergence of FEPs in the minimal case of a system with three bands. In the two following sections, we will go beyond this minimal three-band setting, and consider a three-dimensional lattice model with four bands. We will demonstrate the presence of FEPs with higher geometric multiplicity in the bulk and for the topologically protected boundary modes, where the latter manifest themselves as exceptional lines in momentum space. V Non-Hermitian higher-order Dirac semimetals In this section, we consider the emergence of FEPs in topological four-band models based on a three-dimensional higher-order Dirac semimetal (HODSM) Lin and Hughes 2018 . Its underlying Hermitian lattice realization, depicted in Fig. 5 , is obtained by stacking two-dimensional quadrupole insulators (QIs) Benalcazar et al. 2017a along the z direction. As this model can host bulk degeneracies up to order n=4 (tetrabolic points), it provides us with more freedom to construct non-Hermitian versions displaying FEPs with various partial multiplicity structures, which is the subject of the present section. In the following section, we will further exploit that the topologically nontrivial phase of this system, under open boundary conditions, gives rise to hinge states in its eigenspectrum. Figure 5: Lattice realization of the Hermitian higher-order Dirac semimetal, obtained from stacking two-dimensional quadrupole insulators along the z direction. Each unit cell contains 4 sites, labeled A, B, C, and D. In each plane, the intracell and intercell couplings indicated by solid lines are t and s , respectively. The couplings between the B and C sublattices, indicated by the dashed lines, have the opposite sign, so that each plaquette carries a \pi flux. Neighboring unit cells in adjacent planes are connected by diagonal couplings of magnitude s/4 , with the sign again reverted for couplings between B and C sites. We first briefly discuss the model and its symmetries, and then use the modal expansion of the adjugate matrix to characterize the bulk degeneracies in the Hermitian case. Next, we present four non-Hermitian extensions of this model (see Fig. 6 ) that host all possible degeneracy scenarios in this system, including various FEPs as well as conventional EPs and DPs (see Fig. 7 ). For each of these cases, we use the ranks of the modes to determine the partial multiplicity function for the encountered degeneracy scenarios. V.1 Hermitian parent model V.1.1 Construction and symmetries As the foundation for creating the HODSM, we first review the lattice model of a two-dimensional QI, given by the front sheet in Fig. 5 . The QI consists of a square lattice with corrugated nearest-neighbor couplings, resulting in a unit cell of four sites. The intracell couplings into the x and y directions are of magnitude t , while the corresponding intercell couplings are of magnitude s . In addition to this, any square plaquette (inside the unit cell, or between the unit cells) contains a \pi flux, due to which one of the couplings around the plaquette is taken to be negative. We implement this by inverting the sign of the intracell and intercell couplings connecting B and C sites, hence, along every second vertical line, as indicated by the dashed lines in the Figure. The Bloch Hamiltonian for this model is then given by \displaystyle\mathcal{H}_{QI}(\mathbf{k})=(t+s\cos{k_{x}})\Gamma_{4}+s\sin{k_{x}}\Gamma_{3} \displaystyle+(t+s\cos k_{y})\Gamma_{2}+s\sin{k_{y}}\Gamma_{1}, (67) where the 4\times 4 matrices \Gamma_{i}=-\sigma_{2}\otimes\kappa_{i}\quad\text{and}\quad\Gamma_{4}=\sigma_{1}\otimes\kappa_{0} (68) are obtained from tensor products of Pauli matrices \sigma_{i} and \kappa_{i} , i=1,2,3 along with \sigma_{0}=\kappa_{0}=\openone_{2} . The QI model described by the Hamiltonian ( 67 ) has spinless time-reversal symmetry \mathcal{T}=\mathcal{K} , where \mathcal{K} denotes complex conjugation, reciprocity \mathcal{R} , mirror symmetries M_{x} and M_{y} along x and y direction respectively, and C_{4} rotational symmetry C_{4}\,\mathcal{H}_{QI}(k_{x},k_{y})\,C_{4}^{-1}=\mathcal{H}_{QI}(k_{y},-k_{x}) with the transformation matrix C_{4}=\left(\begin{array}[]{ccc}0&\openone_{2}\\ -i\kappa_{2}&0\end{array}\right). (69) Due to the presence of \pi fluxes in all plaquettes, C_{4}^{4}=-1 . On top of the symmetries mentioned above, the Bloch Hamiltonian also possesses chiral symmetry, which again results from the division of the system into two mutually coupled sublattices. However, unlike for the Lieb lattice, the chiral operator, \mathcal{X}=\left(\begin{array}[]{ccc}\openone_{2}&0\\ 0&-\openone_{2}\end{array}\right) (70) is traceless, since the unit cell contains equal numbers of sites from these sublattices. Therefore, the system does not exhibit a flat band. Instead, the system possesses a symmetric band structure of two two-fold degenerate bands E_{\pm}(\mathbf{k})=\pm\sqrt{2(t^{2}+s^{2})+2ts(\cos k_{x}+\cos k_{y})}. (71) The three-dimensional HODSM follows when one stacks the QIs and interconnects them with intracell criss-cross couplings [see again Fig. 5 , as well as Fig. 6 (a)], set here for convenience to s/4 . In the Bloch Hamiltonian ( 67 ) this renormalizes the intracell coupling parameter t\rightarrow t+\frac{s}{2}\cos{k_{z}} . The HODSM Bloch Hamiltonian is therefore given by \mathcal{H}(\mathbf{k})=\sum_{j=1}^{4}d_{j}\Gamma_{j}, (72) with d_{1/3}=s\sin k_{y/x} , and d_{2/4}=(t+\frac{s}{2}\cos k_{z}+s\cos k_{y/x}) . This model preserves all the symmetries of the QI model mentioned earlier. We next discuss the bulk energy bands and characterize the degeneracy scenarios occurring in the Brillouin zone. V.1.2 Modal analysis of bulk degeneracies To prepare the application of our formalism to non-Hermitian extensions of this model, we now apply the modal expansion to the Hermitian case. For this, we note that the Bloch Hamiltonian ( 72 ) can be written in the block off-diagonal form \mathcal{H}(\mathbf{k})=\left(\begin{array}[]{ccc}0&P(\mathbf{k})\\ P^{\dagger}(\mathbf{k})&0\end{array}\right), (73) where the blocks are 2\times 2 matrices P(\mathbf{k})=p_{0}(\mathbf{k})\sigma_{0}+p_{1}(\mathbf{k})\sigma_{1}+p_{2}(\mathbf{k})\sigma_{2}+p_{3}(\mathbf{k})\sigma_{3} (74) with \displaystyle p_{0}(\mathbf{k}) \displaystyle=t+s\cos{k_{x}}+\frac{s}{2}\cos k_{z}, \displaystyle p_{1}(\mathbf{k}) \displaystyle=i\,s\sin{k_{y}}, \displaystyle p_{2}(\mathbf{k}) \displaystyle=i\left(t+s\cos{k_{y}}+\frac{s}{2}\cos k_{z}\right), \displaystyle p_{3}(\mathbf{k}) \displaystyle=i\,s\sin{k_{x}}. (75) The bulk dispersion relation then takes the form E_{\pm}(\mathbf{k})=\pm\sqrt{\sum_{i=0}^{3}|p_{i}(\mathbf{k})|^{2}}, (76) still describing two doubly-degenerate bands. The form of this expression implies that four-fold degenerate band-touching points appear when all quantities p_{i}(\mathbf{k}) vanish simultaneously, i.e., when p_{0}(\mathbf{k})=p_{1}(\mathbf{k})=p_{2}(\mathbf{k})=p_{3}(\mathbf{k})=0. (77) Within the first Brillouin zone, the location of these algebraically four-fold degenerate points therefore occur at the momentum values (k_{x},k_{y},k_{z})=(0,0,\pm\arccos\left(-2-\frac{2t}{s}\right)) (78) as well as (k_{x},k_{y},k_{z})=(\pi,\pi,\pm\arccos\left(2-\frac{2t}{s}\right)), (79) where the two cases effectively can be mapped onto each other by inverting the sign of t/s . In Fig. 7 (a) we illustrate these degeneracies by plotting the dispersion relation ( 76 ) as a function of k_{z} for k_{x}=k_{y}=0 and s=-t=1 , so that the bands touch at k_{z}=\pm\pi/2 . Figure 6: Schematics of two unit cells in the Hermitian higher-order Dirac semimetal described in Sec. V.1 , along with its four non-Hermitian versions discussed in Secs. V.3 , V.4 , V.5 , and V.6 , respectively. Panel (a) depicts the different couplings entering the Hermitian HODSM model Hamiltonian ( 72 ). The other panels focus on the additional non-Hermitian couplings of strength \varepsilon , which are all intracell and specified in Eq. ( V.2 ). The arrows denote the direction of these nonreciprocal couplings. To analyze these four-fold bulk degeneracies at E=0 , we will therefore fix k_{x}=k_{y}=0 , and in the modal formalism we set the energy shift parameter \Omega=0 , so that the matrix A=\mathcal{H}(\mathbf{k}) . This analysis is further facilitated by the identities \displaystyle\mathcal{H}^{2}(\mathbf{k}) \displaystyle=E^{2}(\mathbf{k})\,\openone, \displaystyle\mathrm{tr}\,\mathcal{H}^{2}(\mathbf{k}) \displaystyle=4E^{2}(\mathbf{k}), (80) where we denote the common value of the squared dispersion E_{+}^{2}(\mathbf{k})=E_{-}^{2}(\mathbf{k})\equiv E^{2}(\mathbf{k}) (these relations are tied to the alternative interpretation of the system as a square-root topological insulator Arkinstall et al. 2017 , which we revisit in the discussion of hinge states). The modes calculated using the Faddeev-LeVierre recursion relation ( 16 ) are then given by \displaystyle\mathcal{B}_{3} \displaystyle=\openone, (81) \displaystyle\mathcal{B}_{2} \displaystyle=\mathcal{H}(\textbf{k}), \displaystyle\mathcal{B}_{1} \displaystyle=\mathcal{H}^{2}(\textbf{k})-\frac{\mathrm{tr}\,\mathcal{H}^{2}(\textbf{k})}{2}\openone \displaystyle=-\frac{\mathrm{tr}\,\mathcal{H}^{2}(\mathbf{k})}{4}\openone, \displaystyle\mathcal{B}_{0} \displaystyle=\mathcal{H}(\textbf{k})\mathcal{B}_{1}-\frac{\mathrm{tr}\,\mathcal{H}(\textbf{k})\mathcal{B}_{1}}{3}\openone \displaystyle=-\mathcal{H}(\mathbf{k})\frac{\mathrm{tr}\,\mathcal{H}^{2}(\mathbf{k})}{4}\openone. Given the condition ( 77 ) and the form ( 73 ) of the Bloch Hamiltonian, A=\mathcal{H}(\mathbf{k}) vanishes at the bulk degeneracy point. We first use Eq. ( 21 ) to verify consistency with the algebraic multiplicity of the degenerate eigenvalue. As A=0 , all coefficients C_{i}=\mathrm{tr}\,(A\mathcal{B}_{i}) , i=0,1,2,3 , appearing in the condition vanish. This implies that the algebraic multiplicity takes the value \alpha_{i}=4 , which recovers the actual algebraic multiplicity of this degeneracy. Next, we resolve the partial multiplicities by using Eq. ( 22 ). For the present model with dimension N=4 , this gives the relations \displaystyle\beta(1) \displaystyle=\mathrm{rnk}\,\mathcal{B}_{1}-2\times\mathrm{rnk}\,\mathcal{B}_{2}+\mathrm{rnk}\,\mathcal{B}_{3}, \displaystyle\beta(2) \displaystyle=\mathrm{rnk}\,\mathcal{B}_{0}-2\times\mathrm{rnk}\,\mathcal{B}_{1}+\mathrm{rnk}\,\mathcal{B}_{2}, \displaystyle\beta(3) \displaystyle=-2\times\mathrm{rnk}\,\mathcal{B}_{0}+\mathrm{rnk}\,\mathcal{B}_{1}, \displaystyle\beta(4) \displaystyle=\mathrm{rnk}\,\mathcal{B}_{0}. (82) Applied to the present degeneracy, where the Hamiltonian itself vanishes, the only mode with nonvanishing rank is \mathcal{B}_{3} . Therefore, we have \beta(1)=4,\quad\beta(2)=\beta(3)=\beta(4)=0. (83) As required, this describes a tetrabolic point, in which each eigenvalue has partial multiplicity 1. Indeed, according to condition ( 23 ), since the first finite mode carries the index \alpha_{i}-\ell_{i}=3 , the maximal partial multiplicity is \ell_{i}=1 , as required for all degeneracies of Hermitian systems. The geometric multiplicity of the corresponding leading eigenvectors then also follows from Eq. ( 24 ) as \beta_{i}=\mathrm{rnk}\,(\mathcal{B}_{3})=4 . Finally, we also recover the total geometric multiplicity from Eq. ( 9 ) as \gamma_{i}=\sum_{l=1}^{N}\beta_{i}(l)=4. (84) This concludes the analysis of the tetrabolic points in the bulk energy eigenspectrum of the Hermitian HODSM model. In the following subsections, we construct various non-Hermitian versions of this model and demonstrate the existence of FEPs in their bulk energy dispersion relations. V.2 Overview of non-Hermitian model variants Compared with the example of the Lieb lattice described in Sec. IV , where we could only construct FEPs with partial multiplicities (l_{i,1},l_{i,2})=(2,1) , the maximal algebraic multiplicity \alpha_{i}=4 displayed by degeneracies in the HODSM model gives us more freedom to construct FEPs with distinct partial multiplicities beyond the generic EP. Therefore, we will now describe non-Hermitian variants of this system leading to FEPs with the three possible sets of partial multiplicities (l_{i,1},l_{i,2})=(3,1),\,(2,2) and (l_{i,1},l_{i,2},l_{i,3})=(2,1,1) . Furthermore, we will encounter generic EP4s with (l_{i,1})=(4) , as well as two-fold degenerate diabolic points with (l_{i,1},l_{i,2})=(1,1) . This covers all possible degeneracy scenarios that are compatible with the chiral symmetry of this model. We base these variations on the non-Hermitian extension \mathcal{H}(\mathbf{k})=\left(\begin{array}[]{ccc}0&Q(\mathbf{k})\\ R(\mathbf{k})&0\end{array}\right), (85) of the Bloch Hamiltonian ( 73 ), where the matrices \displaystyle Q(\mathbf{k}) \displaystyle=q_{0}(\mathbf{k})\sigma_{0}+q_{1}(\mathbf{k})\sigma_{1}+q_{2}(\mathbf{k})\sigma_{2}+q_{3}(\mathbf{k})\sigma_{3}, (86) \displaystyle R(\mathbf{k}) \displaystyle=r_{0}(\mathbf{k})\sigma_{0}+r_{1}(\mathbf{k})\sigma_{1}+r_{2}(\mathbf{k})\sigma_{2}+r_{3}(\mathbf{k})\sigma_{3} (87) are now no longer constrained by the condition R^{\dagger}(\mathbf{k})=Q(\mathbf{k}) , hence, in general are independent of each other. The non-Hermitian extension ( 85 ) retains the block structure and resulting chiral symmetry of the Bloch Hamiltonian, leading to a generally complex dispersion relation of the form \displaystyle E_{\pm,\pm} \displaystyle=\pm\frac{1}{\sqrt{2}}\left(\tr[RQ]\pm\sqrt{\tr^{2}[RQ]-4\det R\det Q}\right)^{1/2} \displaystyle=\pm\left(\sum_{i=0}^{3}r_{i}q_{i}\pm\sqrt{\sum_{i