ANALYSIS OF A NUMERICAL SOLVER FOR, RADIATIVE TRANSPORT EQUATION HAO GAO AND HONGKAL ZHAO Austnact. We analyze a numerical algorithm for solving radiative transport ‘equation with vacuum of reflection boundary condition that was proposed by ‘the authors in 2000 with angular discretization by the finite element method and spatial discretization by the discontinuous Galerkin method or the finite difference method. 1. INrRopUCTION Radiative transport equation (RTE) (1.1) models a wide range of applications, such as neutron transport, atmospheric radiative transfer, heat transfer and photon migration in tissues. In this paper, we study a numerical algorithi for the following RTE on a bounded domain 9 with smooth boundary 02, (11) 8. VOCE, 8) + ue(@, 8) = if M680, S)ds! + (F,8), (2,8) Ex S, where #(Z, 8) is the particle density at position # along direction 4, q is the source term, $ denotes the angular space, ic., unit circle in two dimensions (2D) S! or unit sphere in three dimensions (3D) $2. ‘The parameters in (1.1) are absorption coefficient jig, scattering coefficient j1., transport coefficient je = Ha + js and seattering kernel f. We assume all the parameters in (1.1) are non-negative and bounded. In particular, 1, is bounded below by a positive constant Cy, Le, 0< Cy < fe Very often f is rotationally invariant, ie., (3,4) = f(3- 3"), and is normalized with f f(5-3")d9’ = 1. Although we assume this rotational invariance in this paper for simplicity, our algorithm and analysis can be extended for the general case. As ‘an example, the commonly used f in optical imaging to model photon transport in tissues is the following Henyey-Greenstein (H-G) fanetion (1.2) SG-8) = { where the parameter 0 < g < 1 indicates how forward peaking the scattering is, ‘To specify the boundary condition, we define A(#) as the boundary normal for #EAN,T+(E-) = ((#,8) € ANXS, with §-A > 0(< 0)} as the outgoing (incoming) ‘Received by the editor September 13, 2011 2010 Mathematics Subject Classification. Primary 65N12. Key words and phrases, Raviative transfer equation, discrete ordinate method, source itera: tion, Henvey-Greenstein function, discontinuotss Galerkin method. "The work was partially supported by the NIH/NIBIB grant EBOLS3S7 and NSF grant DMS0s11254. @2012 American Mathematica Society Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct se14 HAO GAO AND HONGKAI ZA boundary, and n,(ng) as the refraction index of the med the reflection boundary condition for (1.1) is (1.3) (F, 4) = RUF, 4,8')(7,8"), (¥,8) ET, (Fe) ET, where & refers to angular direction at 7 € 00 that is reflected into § at # when nz # no, and R corresponds to the reflection energy ratio computed through the Fresnel formula (5). When there is no refraction index mismateh, i.e., n; = no, (1.3) is the so-called vacuum boundary condition with R= 0. With the aforementioned conditions on jig, jt, and f, there exists a unique positive solution to RTE (1.1), when Q(Z, 8) = ff qld ~ ra, a)e~ Se mer’ dy is bounded and positive [6]. Here d is the distance from Z to the domain boundary AM along the diteetion —8. In this work we will analyze the numerical algorithm proposed in [4] for RTE that works for general geometry and general reflection boundary condition. In this algorithm we use angular discretization by finite element method and spatial discretization by the discontinuous Galerkin method or the finite difference method. In Section 2 we give a brief description of the numerical method and then present various error estimates and convergence analysis in Section 3. One of the main difficulties in our analysis comes from the general reflection boundary condition. (environment). ‘Then 2, NUMERICAL ALGORITHM The analytical solution for RTE is rarely available except in a very special setup. Efficient mmerical algorithms are needed for many applications. However, the efficient numerical computation of RTE is challenging mainly due to the following difficulties: ‘© high dimensionality «different solution behavior, e.g., from the transport regime near the source to the diffusion regime after a few mean free path; + forward peaking due to anisotropic scattering kernel function; ‘* general boundary conditions. Moreover, a large and not so sparse (due to the scattering term) linear system has to be solved after discretization that requires an efficient matrix solver. In [4], we developed an efficient multigrid numerical algorithm for RTE that works for general geometry and general reflection boundary condition. The solver is available at http://sites.google.com/site/rtefastsolver/. ‘The key features of our method include: ‘© angular and spatial discretization that (1) preserves the correct behavior of transport and seattering operator, (2) captures both short and long range interactions in spatial and angular domain effectively when using the multi- arid strategy: ‘© improved source iteration (ISI) based on Gauss-Seidel iterations with angle dependent ordering of spatial nodes which (1) captures transport efficiently, (2) converges faster than source iteration (SI); ‘* multigrid method in both angular and spatial domain with ISI as the relax- ation which further accelerates the solution algorithm, particularly in the diffusion regime; ‘flexibility for both structured and unstructured grids in 2D and 3D with general boundary conditions. Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct seANALYSIS OF A NUMERICAL SOLVER FOR RTE, 155, In the following, we will briefly describe our solution algorithms for RTE and the details can be found in [4] 2.1, Angular discretization. Let {Zm(8), 1 Sm < M} be the piecewise linear angular basis on an angular mesh {3, 1 0, which preserves the qualitative behavior of localized seattering even if the angular mesh resolution not be enough in an extremely forward peaking situation. We will estimate the er- ror between ¢ after the angular discretization (2.3) and RTE solution & in section 3. Remark 2.1. Although an octahedron with eight equilateral triangular faces are used for angular discretization, for better uniformity, an icosahedron, a regular polyhedron with 20 identical equilateral triangular faces, can be used. 22. Spatial discretization. Let (0m(i), 1 Win’ where 20, b= "22 > 0, and i,j are indices for x and y coordinates, respectively. We will analyze FDM for the system (2.3) on structured mesh in section 3.3. 2.3. Improved source iteration (ISI). SI (source iteration) is @ popular method for solving the discretized linear system of RTE. In SI, the scattering entirely lags one step behind transport, ie., BV HOMES) = on f 1 ) ‘This simple iterative procedure has a clear physical interpretation, e.g," captures all particles that have been scattered no more than n times. However, SI has a slow convergence when the scattering is dominant, e.g., in the optical thick regime where the domain size is large in terms of mean free paths. ISI was developed in [4). ISI is essentially a Gauss-Seidel method in both angular and spatial variables, which utilizes the most updated information during iterations. We also use angle-dependent ordering in ISI, ic., the ordering that aligns with § to follow the characteristic direction of the transport operator. For simplicity, we illustrate SI and ISI using the first-order upwind FDM (2.4). Here since Oy € 0. $), the ordering in ISI is i= 1,2,+*+ ,j = 1,2,+++. Then ISI is + bo, NO", Ads + az, GHD + He atm tttay (Dirt Wen’ Ome + Di mgr Wm Gt G+ DE pe Hatmm while the SI is 468-1 jm + O02; 1m + Has Comin Wo dm athe It can be easily shown that upwind spatial discretization combined with angular discretization for general boundary condition (2.8) produces a linear system with the M-matrix, which guarantees that (1) ISI converges and converges faster than SI, (2) global error in maximum norm satisfies maximum principle, (3) ISI is a smoothing operation. In seetion 3.3, we will analyze this FDM scheme (2.4) and compare ISI (2.6) with SI (2.7). With correct ordering ISI can capture the solution in transport regime efficiently. On the other hand, multiple seattering, which cat diffusive behavior in both angular and spatial domain, can be handled by our multigrid (in both angle and space) approach effectively. (2.7) Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct se158 HAO GAO AND HONGKAI ZA 24. Boundary condition. Our direct FEM-based angular discretization deals with general boundary conditions at ease. For example, if there is a mismatch of refraction index at the boundary, local mesh-based photon flux in one direc- tion can be reflected and refracted into other direetions according to Fresnel for- mula [5]. ‘To discretize (1.3) at a boundary point # along an incoming direction Sm, Which belongs to the angular mesh, there is a outgoing direction 3” such that OLE, Bm) = RC, Sn, 8"), 3"), where (Z, 4m) € E>, (Z,8!) €T*. However, & may not belong to the angular mesh, so we use linear interpolation of fluxes at neigh- Doring angular mesh points 3, to approximate the flux in direction #. So after angular discretization, the reflection boundary condition becomes (28) Om = Lot mm (Fm (B, (BS) €L-,(F am’) EL. ‘Two neighboring directions whose interval encloses 4’ in 2D and three neighboring directions whose triangle encloses # in 3D are used for the linear interpolation. In particular, Dy tram’ = Rs, 5!) <1 tissue air a Ficure 3. Reflection at the bow are on the angular grid. When the incoming direction 8, is almost tangential to the boundary at Z, the corresponding outgoing direction 3" is also almost tangential to the boundary. ‘Thus some 8, used in the interpolation may be an incoming direction. See Figure 3 for 2D example: Sy, is the reflection angle for 3,, which is not on the angular grid and is interpolated from sy, and yy. In 2D it is easy to show that the general form (2.8) is still valid. That is ¢m(#) = r[aem (#) + (1 ~ @)om(#)], where 1 = RZ, 5m,8") < Land the weight a > 0 in the linear interpolation. ‘So r(l-a), (29) en(@) = 8 Om (E) < rm (2). In 3D the situation is more complicated since three neighboring directions are in- volved in the linear interpolation. Incoming directions that are almost tangential to Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct seANALYSIS OF A NUMERICAL SOLVER FOR RTE, 19 the boundary may not be explicitly represented in terms of outgoing directions, es. when two of the three directions used in the interpolation are incoming directions In summary, the general reflection boundary condition (2.8) can be regarded just as some kind of scattering relation among different directions at boundary. So our angular discretization plus linear interpolation for general boundary conditions inherits those nice properties for approximating the scattering term in RTE, ic., a true averaging operation among directions, When coupled with the finite difference scheme for spatial discretization, it can be easily shown that the fully discretized Iinear system has an M-matrix even with general reflection boundary condition (see the proof for Theorem 6). 3, ERROR ESTIMATES AND CONVERGENCE ANALYSIS 3.1. Angular discretization. The main result in this section is to show that the solution @ after angular discretization of (2.3) and RTE solution ® to (1.1) satisfies (3.1) lle - | m(#) ~ ®(, 8)]? de fk From now on we define the following weighted L? norm em i o a (32) lle? In order to prove Theorems 1-3, we will first prove the following estimate in 2D (Lemma 1) and 3D (Lemma 2) (3) Uh — Too = OC). Here [lx is with respoet to the angular vatiable. To shnplify the notation, hi used to represent the angular mesh size hg in this section from now on. Lemma 1. Assuming f, € C*(S!), the aforementioned 2D angular dis- cretization with I"(3m) = Spy tam’ mv has the second order convergence, ie., (Gm) = P*Gm)| = O(H2), forall m: Proof. {1 — I*| < |f ~ I| +|f—1*|, where I is defined by (2.1). We estimate these two terms separately. First, lari Due to the piecewise linear approximation, ||f fll = O(0?), || thus (J — I] = O(h?). fuu- HB + FH — Odd) < 2x (Ulf ~ foc @lloo + lI fllocl!® ~ Pll) lo = O(N), Next, observe that 1— wy = fou(F — f)dd = O(f) and wy — Wom’ has the same sign for all m’ (because w,,,'s are all non-negat normalized Why, 15 Wm! = Mm {Sm Wan) [2 =| S(e Sane — Winn) re [lool Swe — 1 = OC?) Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct se and Wm is the a160 HAO GAO AND HONGKAI ZA We have proved (3:3) similar except that the gulation of the sphere wi 8 normalized from wy Lemma 2. Assuming f, € C2(S?), the 3D angular discretization as described above with I"(m) = Cy Wmm’Pm’ has the second order convergence, i.e, |J(3m)— T*(5y,)| = O(h2), for all m. 2D. In 3D, the angular discretization (Figure 2) is ht u2,,,- is computed on the flat element of the trian- piecewise linear approximation. Then the weight Wyam? Proof. Consider a triangle element of the triangulation of S* as depicted in Figure 2. Let be a point on the triangle denoted by A and let # = #/\z| be the corresponding projection onto the sphere. ‘The map is one-to-one, onto and smooth on each triangular element. In particular, @” —# = 2"(1 —[2}) with 1-7] = O(?). ‘The change of variable #”(#) is smooth and wt %,7 Ve = BU? It can be ensly seen that I ~ , @ ji almost the projection to the triangle plane up to a O(h) perturbation. So the corresponding surface area on the sphere and the surface arca on each triangle clement satisfies ds(2) = ds(#)(1-+O(h#)). Also the piecewise linear approximation f(#) of any function J(" (#)) € C2(A) on each triangle satisfies || = f\|s = O(02). Therefore, = f Jdds(@) = vf [7 + OW) + O(2)][1 + OU#)] as(z) = 7 + 0142), where LiLo, fds(@). In particular, [(3,) = Dyy Umer Use the same argument as in Lemma 1, |f(3n) ~1"(3n)| = O(h?). Qo Remark 3.1. H-G function (1.2) belongs to C? for g <1. When g = 1, f isa Delta function in angle. So there is no scattering among, different directions and 1(3) = (3). As a result, the system (2.3) is reduced to a set of uncoupled first- order transport equations, which can be solved easily and the analysis is the same as that for a single transport equation, ‘Next we will use Lemmas 1 and 2 to prove the second-order convergence (3.3) for the angular discretization in Theorem 1. However, due to no rotation invariance for the triangulation on the angular sphere, we need the following lemma in order to proceed in 3D. For comparison, the grid is uniform and rotationally invariant on the circle in 2D. As aresult, wy’'s are equal in 2D, while they are not in 3D. For the same reason, the angular Weight tmq"s are symmetric in 2D, i.e. Wyn’ = Wim Which is not true in 3D either. However, this broken symmetry in 3D can be remedied in the proof of Theorem 3 using the following lemma between wyam! and Wyn Notice that although the lemma is for the normalized weight tums, it also holds for wy SINCE Wn’ has the same dilation from w2,,,, and “p= = 14+ 0(h2) for all m’ Lemma 3%! ~ “at — O(h) for angular discretization in 3D, Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct seANALISIS OFA NUMERICAL SouvER FOR RE vs Proof. We have ms fh Flim Men G)aa = fm dn) + O18) = Gm -Bye) + 00) Similarly, Wy = (F(8mr-8m)+O(h))ty. Therefore, Sam’ — "hem = O(h). From the fact. "z=" = 1+ O(h?) for each m’, the claim follows. o ‘Theorem 1. With the angular discretization (2:3), the approximate solution ¢ satisfies (34) max |m(#) ~ #(2, 8m) = O(0?), if the solution (7, §) to the RTE is C*(5). Proof. For the si plicity of notation, let y, denote ®(F, 8m). (5) Poe f Flin 2} @7)d¥ = YO tome ~ Be) + Dn wa with (35) Dou = Dtnn ne = f Sms (et Let ¢, Om — Bm, and then €,, satisfies ™ (37) Bm = Vem + Hem = +Dy. Since 0 is a bounded closed domain, assume that maxy,p [Om (2) ~ ®,8m)| = eae is obtained for for some rn at some point Z. (1) If#isin the interior of 2, Ve yi(#) = 0 and because tym’ 20, Dm Wm! = 1, from (3.7) we have Heese Sa (2) IF # belongs to the boundary 0, - (a) If (8m) EL, ke, 8m “A > 0, we have 8 - Vem(#) > 0, Again from (8.7) we have ex = O(h?). (b) Otherwise if (7, 4.) € P>, for the general reflection boundary condition with discretization (2.8) we have enl®) = Tram ew @)+D, nm’€xe + |Din| = Halse S ||DIloc = eae = OCH) aa) eT, im) T+, where D is the truncation error due to linear interpolation of the real reflec- tion condition for at the boundary and is of order O(h2). Also ram’ 2 O and 1’ Tayn’ <1. Combined with the argument for (a) we have ex. = O(h2).. Hence we complete the proof for (3.4). a Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct se1 HAO GAO AND HONGKALZHA0 ‘Theorem 2 (Vacuum Boundary Condition). Assume the solution «(z, 8) to the RTE is C?(S). With vacuum toundery condition and sufficiently fine angular mesh yu [omar ® (38) lle 4) on). ‘m))? Proof. Multiplying both sides of (3.7) with wycm, integrating over the spatial domain and summing over all angular directions, we have aM fied, +m Le me, (3.9) MoM ue With vacuum boundary condition, ¢m = 0 for 3m -A <0, so the boundary integral bocomes iv 1 102) 1 we [ ae ff Wm dm “fie?, > 0 ie non on Next we analyze the frst term on the right-hand side of (3.9), a) f tatrmtinmeniem <4 SS f tanta be +e) On one hand, Dy, yy Wonton’ mney Since Nyy Wan? = 1 from our angular normalization; on the other hand, using Lemma 3, YL wma dy = TD lle tina + wit O(h) 2 = Swe (1+O(h) ed. ‘Therefore, when h is small enough, we have DX wmnwmm ee S (1+ 26) Yo wee with 6 ~ O(h). Notice that 6 = 0 in 2D due to rotational invariance. ‘Therefore, from (3.9) and (3.11) we have a ww (3.12) Yun f (a Sua)eh, < 2 tm ff Dat Since jg is bounded below with a positive constant Cy, we have pia — Bjte > Cy in 2D with 5 = 0, while jig ~Sjtg > C in 3D for some positive constant C when h is small enough. ‘Then we apply Holder's inequality tothe right-hand side of (3.12) such that 3, tm fo) Diném < ||| -|lelh ‘As a result, we have (313) ile < CDI for some positive constant C. Moreover, from Lemmas 1 and 2, we have || Dll. (42). Thus, Jo ~ #)| = (02) a Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct seANALYSIS OF A NUMERICAL SOLVER FOR RTE, 163 ‘The boundary term Twn Jon Sm *fe% becomes more complicated for re flection boundary condition (1.3). For simplicity, let us first assume that for each incoming direction §,, on the angular mesh there is a corresponding outgoing di- rection S, on the angular mesh with the following reflection condition (31M) Sy f= Sy fh, Om = Fam Omry my’ OS Tam <1 Then (3.15) Se 9 J. in 2, = Jo, ee = wee Which is analogous to the continuous case. In 2D, each term on the right-hand side is non-negative since tim = Wm and 1—r2,,» 2 0. In general, the above assumption (3.14) is not true. However, if we assume a non-perfect reflection condition, i.c., R(Z,5,5") <1 in (1.3), we can prove the following lemma which will be enough for the error estimate for the general reflection boundary condition in 2D. ‘Lemma 4. If there is no perfect reflection at the boundary, ic., R(Z, Lou Tmm? <1 is true in 2D. Nl, Proof. In 2D, since the angular mesh is uniform, the corresponding reflected angles are also uniform. Moreover, we use linear interpolation between two neighboring angles to interpolate the reflection condition (Figure 3), so at the boundary a2, we have (8.16) orm where Sy is an outgoing direction that contributes to both incoming directions én, and 8m, in the linear interpolation and oy/y9', my’ are the linear weights. From the non-perfect reflection condition, we have Sy, Timm’ <0 <1 Oo nn RB Sing) + Ang RB 8 Omi + Ont Next we prove the similar L2 error estimate for general reflection boundary condition in 2D. ‘Theorem 3 (Reflection Boundary Condition). In 2D, assume the solution (#3) to the RTE is2(S!). With no perfect reflection at the boundary, i.e, RZ, §,5") < <1 in (1.3), and if the angular mesh is fine enough, (3.17) \l6- || = 00) Proof. The key is to control the boundary term 7M _, wim f 8m -fe2,d0 in (3.9). For the general reflection boundary condition (1.3) approximated by linear interpolation (28), we have 3m) EL, (F, be) Lt, Em(E) = D> rmm (Hem (B+ Dn(Es ( Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct se104 HAO GAO AND HONGKAI ZA where Dy, (2) is the truncation error due to linear interpolation of the real reflection condition for at the boundary and is of order O(%?). ~~ ne flee = Down fm Ales Now let us estimate the second term in the above expression: Layeago tm fog lee“ F(Caqyez0 Fm em’ + Dra)? $Lsn-aco tm Jon lim Al [ECs azo tmm'tm’)? + 715 D3] aco tm Jog lim ° il [Ts azo Fmm'eRye + 503) (1 0") x mf, lim “fle2, Fil + Cage Romar mil) From (3.12) we have a Ste Ps ng the above real we pet (-o') = = Io iowile2, +Cllell? < eIPP. <6 Combi 1 d um Io [itl GIDIF Since both truncation errors in angular discretization for boundary condition and scattering operator are of order O(h?), we have proved the result. a ‘The complication for reflection boundary condition in 3D mainly comes from: (1) the difficulty for obtaining the similar inequality as (2.9) in some pathological case when is almost tangential to the boundary at #, and (2) the non-uniformity of the angular mesh, i., the failure for establishing the similar claim in Lemma 4, Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct seANALYSIS OF A NUMERICAL SOLVER FOR RTE, 16s, 3.2, Spatial discretization by DG. In this section, we give an error estimate for the spatial DG discretization of coupled transport equations (2.3). First we give a brief summary on DG method and its analysis on the following single transport equation on a bounded domain @ with boundary condition on inflow part of the boundary a = {#€ AD: A(z)-s <0}, 18) u=g on OD, { 8. Vu) + e@)u(a) = f(#) in Q, where (2) is the outward unit normal to 9 at #. People have been studying the DG method for the convection equation (3.18) since 1974 with the first error estimate on general meshes [7], ie. (19) Ju = ula < CA ules, followed by the improved convergence result [2| in 1986 with (3.20) [lu unlla SCH ul as, where k is the degree of polynomial approximation in each element. Actually, the result was proved with a stronger norm than L? norm. Besides, LP stability and error estimates for p # 2 were analyzed for piecewise linear DG uniform or piecewise uniform triangulation [2]. However, since the sign of o() is undetermined, a special auxiliary function has to be introduced to control the possible exponential growth and get the error estimate. The construction of such an ausiliary function requires convexity of the domain 2. Two years later, assuming the conforming triangulations that are obtained with slabs of parallelograms divided into two triangles, [3] showed an optimal convergence estimate (2) llu=ualle < Cn ful esa Recently, [1] proved a similar result (3.22) [lu —uall < CAE for meshes that are not necessarily conforming or subject to any uniformity condi- tion, however, made of simplexes of a unique outflow face Although we need to deal with a system of coupled transport equations (2.8), the coupling is not very strong. In particular, due to the absorption, jy — ts = fla > 0, which is also a crucial condition for the stability of the original RTE (1.1), our proof is relatively simple. For notation convenience, we use h(h,) to represent the spatial (angular) mesh size hereafter. Let us start with the DG formulation of RTE (23) with vacuum boundary condition. Given a triangulation 7), with h = supaer,diam(IX) and the finite-dimensional space Vj' which is composed of functions that are polynomials of at most & degrees on each K€ Ty, let (u,v)x = fc wv and (u,v)5 = fy, wvls- Ail with [lull = (uu) and (u))% = (usu), where BCT, Ua (Uxex, K)\OQ. For each angular direction m, let 0", = {F € ON: A(z) (5)0}- Then DG approximation 6! € (Vjt)"" of the solution of (2.8) satisfies (3.23) Am( ms) (4,0) + (a, ¥)one, for all v € Vif for each m, Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct se166 HAO GAO AND HONGKAI ZA where Arn Gas 8) = Sm TOM, + Hel, = ta J Wm Ghar Bac (324) F = +n” — oh dnp + (Oh Yon with &,* = lim, or OA (2 4 6S) and (uyu)rp = Jp, well «fil. Here q and gy represent the internal and boundary source respectively. Note that we can replace "by the exact solution @ in (3.23). That is we have the following consistency relation (3.25) Am (Om — 0%,.0) =0, for all v € Vik, for each m, Notice that (3.23) can be regarded as a system of (3.18) coupled by the scattering term. Therefore, (3.23) can be analyzed similarly as the vector form of (3.18) with non-negative c, once we can control the scattering term. This is exactly how it will be carried out in the proof for Theorem 4, Here we assume the shape-regularity condition on the mesh, i., there is constant-¢ > O sch that lig /px. > 0 for each simplex K € T),, where hy is the diameter of the simplex K and py is the diameter of the largest ball included in K. ‘Theorem 4 (Vacium Boundary Condition). With vacuum boundary condition, when the triangulation 7), satisfies the shape regularity condition and hi, fia is small enough, the error between DG solution «/" via (3.23) and the exact solution & of (2.3) with inflow toundary condition satisfies (3.26) lle oh] SCH ong where the norm || «|| is defined by (3.2) and jol2.. 1 Wal bmalfress Proof. Let dm be the Le projection of Gm into Vi, Le, (27) (Om - v € Vi. ‘Then let us consider (3.23) for e wwe have the following with (3.28) Ayn(€ms€m) = We first estimate Ay (¢mas€ya)- Since —o, Using the consistency relation (3.2 rn (ms Em): (Gm Venema = —(emrbn: Wendie + fd fey we have Db: Vemsemx = — Heh + ea rehs — ele — ZllemlVine + 5 (lem) )Bneg T Asa rent, DG Vem ema + (Om — en end + (Om) Vinee (eh — ay + 3 lem) Bae Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct seANALYSIS OF A NUMERICAL SOLVER FOR RTE, 1st “The seattering term ean be estimated as follows = Eg tin Een — fs Cg Won’ Cu > En walla ies Ee om fly +) On one hand, (331) Dem Dein fae = Dam f med (Qn the other hans, using Lemme 3, T= Ling tim Lom Wn’ So ts (tt + oy Jn Moen in! fy Mat + Ola) For any hg, there exists a constant 5 = O(hg) such that (3.82) (333) Ts 429 Dem f we ™ in ‘Combining (3.30) to (3.33), we have (3.34) J>Cllel!? where C = C,, ~ 6j|jslao with lower bound C,, of jg, and is non-negative when hig is simall enough in 3D. Notice that C=C, in 2D due to rotation invariance of the angular mesh. Therefore, 838) Yumdn(emen) ® D [iar + Deal (eh — calli + (endin| : where D = min(C, 3), Next, let us estimate Am (Cm, ém), and it goes as follows. For the boundary term, we have is — Cm Endege + (Ems Em)on™ S 2(Uem — em) Bg + Cem) Bnm) + BEB + (Em)Bnm)- For the convection term, using the property (3.27) for the 2 projection, (3.36) (337) (Bm + Vem; Em) = (Bm * Vem Em) = A Aa, Similarly, as before, the scattering term can be estimated as follows: Dem Dole = ts time's Em) mae m < Dual P ben + ele eR) es (338) : +d EG lew + Alt) wy m@ nl] D << Fllell? + elle? Last, from the standard estimates for the L2 projection, we have 33 lemllas $ CHR Hylan (K), (3.39) an Uemllox Chi? aesn (BK). Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct se168 HAO GAO AND HONGKAI ZA With the shape-regularity assumption, from (3.36) to (3.39), we have (40) > mA (Emsn) 2 |e? o> inl (ea ~ Cm) Bg + Ue) + OHO Combining (3.28), (3.85) and (3.40), we have shown that (3a) llell < Ch Iba. In the proof C always denotes some constant. a Remark 3.2. Combining estimate (3.1) for angular discretization and (3.26) with ke = 1 for our piecewise linear DG spatial discretization, we have the following estimate for our scheme: (3.42) [I — 6"|] = OC) + O(n”), However, second order aceuracy is observed in both angle and space in the numerical tests in [i] Remark 3.3. The proofs with O(h#*1) convergence in [1, 3} do not apply here since the required property on the spatial mesh in general is not satisfied for RTE with all possible angular directions. For general reflection boundary condition (2.8), the DG formulation (3.24) be- sO) Am (Gh +6 — where Spy Pmm! = R(Sm,4’) <1 and Syv's are the neighboring directions used in the linear approximation to approximate the outgoing direction # that is reflected into the incoming direction ém. Now let us prove the same result as in Theo for general reflection boundary condition in 2D. ‘Theorem 5 (Reflection Boundary Condition). In 2D, assume that the triangula- tion 7), satisfies the shape-regularity condition and h, h, is small enough, the error between DG solution @” via (3.23) and the exact solution @ of (2.3) with reflection boundary condition (2.8) satisfies (34a) Ilo ol] SCH na if R(¥, 4,8") 0. ‘We have £0 Sy, Wine Boa Therefore, Lam tm {Lan Fm’ Orn’ mn Dose 30” Dm tm (Cem) )Bprm Hence we get an estimate similar to (3.35) (3.46) Ealenitn)® fll? + Leal as > ea) Now we estimate for Aj(¢msém)- Again, the only complication cones from the somes For the first term, we have (emémoas < POEM enn + Be Em) ‘To estimate the second term, we first estimate pointwisely with respect to the angular atible Lmtwacotm (Cs the max norm with respect to both angular and spatial . Without loss of generality, assume Ar = Ay = h. Then we have the following result on the error estimate of FDM (2.4) for (1.1) ‘Theorem 6. Assume the solution to (1.1) (#, 8) isC?(x S) anddenotee, jn = by ~ (2s Sree Where 6, 1S the numerical solution from (2.4), then (3.49) llelloo = O(h) + O(H8). Proof. For non-boundary nodes, ¢,jn satisfies (3.50) (@+b-+ pr)eiggm — (Ge. j,m + BELA) ~ Ha D> Wim €s jam? = Tam where the local truncation error T, jm ~ O(h) + O(N) due to first order upwind difference in space and piecewise linear approximation in angle (‘Theorem 1). Let €:,jm = llellac- Since a+ 5+ pe, a, B and jt are all non-negative, and (a+ b+ 1H) ~(@ +84 Hs Dpy Wnm') = Ha > 0, at (i,j,m), the linear system has an M-matrix and we have (351) (a+b + p4)]lellae S allellso + bllelloe + He > Wm llellae + IT. jm: Since Dyy tm = 1, we have (3.52) Hallelloo < |ITlloe- Let C be the positive lower bound of ja, thus we have proved the statement that 1 (3.53) llelloo S GT loc = O(h) + OCF For boundary nodes with general discretized ref © if ig -f1> 0, ie. (Fi,Uj>3m) EP, after we use relation (2.8) to substitute all €ssam With Sy “SO, We hae on boundary condition (2.8), (a+ B+ pe es. jim — (@e6—1,5 9m + be, 5—149) — Hs Loy 90 Wenn! Cs, 5 mn? + Ln sigriso Here Tn may include the error due to linear interpolation of the exact reflection boundary condition, which is of order O(N). Since rym” > 0 and SyyTm'nr <1, the M-matrix system is preserved by our angular discretization at the boundary. ‘The same argument used above shows elle = O(h) + O(3). am’ Sotgua30 Panag”) = Taso Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct seANALYSIS OF A NUMERICAL SOLVER FOR RTE, m <0, from bo indary condition (2.8) we have Again, T.,.m may include the error due to linear interpolation of the exact reflection boundary condition. Combining the above cases, we see that \eljc = O(lt) + O(h2) is also true for boundary nodes. a ince the fully discretized linear system has a M-matrix structure, the ISI (2.6), which is a Gauss-Seidel iteration, converges and converges faster than the standard SI (2.7). Here is a simple argument, Let £* == |l62jm ~ Ohjmllscr by the similar argument in Theorem 3, for IST (26), we have (354) (1= pmE", where r= [ja/Halle an fy Similarly, for SI (2.7), we have (355) 1 py) EB" Obviously, IST converges faster than SI since py > py When scattering is weak, eg., r= |jts/Hallae < 1, or in the case of forward- peaking regime, which is usually the case in optical imaging, e.g., g ~ 1 in H-G, 1— tam <1, it can be seen easily from the above argument that the improvement Of IST over SI is significant. In the extreme case with g = 1, nme = 0 for m % m’, ISI takes only one iteration when the spatial ordering is along the characteristic for each transport direction, When scattering is dominant (r >> 1) and f is isotropic, or the domain is optical thick, even IST may converge quite slow. Since IST captures transport effectively and is a smoothing operation, itis used asa natural relaxation scheme for a multigrid algorithm in both space and angle [4], which can deal with RTE in different regimes effectively. Rerenences 1. Cockburn B., Dong B., and Guzman J., Optimal convergence of the original discontinuous Galerkin method for the transport-reaction equation on special meshes, SIAM J, Numer. Anal 3 (2008), 1250-1265. MR2390002 (2009%:65209) 2, Johnson C. and Pitkiranta J., An analysis of the discontinuous galerkin method for a scalar hhyperbotic equation, Math. Comp. 46 (1986), 1-26. MRS15828 (84b:65100) 3. Richter G. R., An optimal-onder error estimate for the discontinuous galerkin method, Math. ‘Comp. 50 (1988), 75-88. MROITS19 (8865197) 4. Gao Hand Zhao H.K., A fast forward solver of radiative transfer equation, Transport Theory ‘and Statistieal Physics 38 (2000), 149-192. MR2572432 (20110-86008) 5, Jackson J.D, Classical electrodynamics, rd edition, Wiley, New York (1909). 6. Case K. M. and Zwefel P.F., Linear transport theory, Addison-Wesley, Massachusetts (1967). MR0225547 (37:1140) 7, Lesaint P. and Raviart P. A., On a fnite element method for solving the neutron transport equation, Mathematical aspects for finite elements in partial differential equations, Academie Press, New York (1974). MR06SSL42 (58:31918) Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct sem HAO GAO AND HONGKAI ZA 8. Han W., Huang J., and Bichbols J. A., Discrete-ordinate discontinuous galerkin methods for soleing the radiative transfer equation, SIAM Journal on Scientie Computing 32 (2010), 477- 497. MR2600827 (2011665198) 9, Reed W. H. and Hill T. R., Triangular mesh methods for the neutron transport equation, Los Alamos Scientific Laboratory Report LA-UR-73-479 (1973), Deparesr or Marmatanies, Usivenstry oF CAuPonstA, Los ANcrtes, CaLionsta, ‘90095-1555 E-mail adress: haoggmath.ucla.edu Derren oF Marieatanies, UNIVERSITY OF CALIFORNIA, IRVINE, CALIFORNIA 92697-3875, E-mail address: 2heogmath.uci ed Leanaato Huard Use: Peparedon Thu Ap 15184040 €0T 2013 dombaston I 2813 4882 Lean cay tition my apie rede btn: Iawm ans cyjumastomsct se