# Notes on Fermi-Dirac Integrals 3rd Edition

Raseong Kim and Mark Lundstrom Network for Computational Nanotechnology Purdue University December 10, 2008 (Last revised on August 4, 2011)

1. Introduction Fermi-Dirac integrals appear frequently in semiconductor problems, so a basic understanding of their properties is essential. The purpose of these notes is to collect in one place, some basic information about Fermi-Dirac integrals and their properties. We also present Matlab functions (see Appendix and [1]) that calculate Fermi-Dirac integrals (the “script F” defined by Dingle [2] and reviewed by Blakemore [3]) in three different ways. To see how they arise, consider computing the equilibrium electron concentration per unit volume in a three-dimensional (3D) semiconductor with a parabolic conduction band from the expression,

n=

EC

∫ g(E) f (E)dE = ∫ 1 + e(
0 EC

g(E)dE
E − EF ) k B T

,

(1)

where g(E) is the density of states, f0 (E) is the Fermi function, and EC is the conduction band edge. For 3D electrons with a parabolic band structure,

g3D

( 2m ) (E) =

* 3/ 2

2π 2 ℏ3

E − EC ,

(2)

which can be used in Eq. (1) to write

(2m ) n=
2π ℏ

* 3/ 2 ∞ 2 3

EC

∫ 1 + e(

E − EC dE

E − EF ) k B T

.

(3)

By making the substitution,

ε = (E − EC ) kBT ,
Eq. (3) becomes

(4)

1

It is more convenient to define a related integral. Note that its value depends on ηF . F 1/ 2 (η F ) ≡ 2 ε 1/ 2 d ε . (11) It is important to recognize whether you are dealing with the “Roman” Fermi-Dirac integral or the “script” Fermi-Dirac integral. (7) can be written as n = N 3DF 1/2 (η F ) . we can express the electron concentration as n = N 3D where 2 F1/2 (η F ) . which measures the location of the Fermi level with respect to the conduction band edge.(2m k T ) n= * B 3/ 2 ∞ 2π 2 ℏ3 ∫ 1 + eε η − 0 ε 1/ 2 d ε F . This integral can only be evaluated numerically. π ∫ 1 + exp(ε − η F ) 0 ∞ (10) so that Eq. By collecting up parameters. (6) π (7) N 3D  2π m*k BT  = 2  h2   3/ 2 (8) is the so-called effective density-of-states and F1/ 2 (η F ) ≡ ∫ ∞ ε 1/ 2 d ε 1 + exp(ε − η F ) 0 (9) is the Fermi-Dirac integral of order 1/2. For example. in two dimensional (2D) semiconductors with a single parabolic band. There are many kinds of Fermi-Dirac integrals. the density-of-states is m* g 2D ( E ) = 2 . πℏ (12) 2 . (5) where we have defined η F ≡ (EF − EC ) k BT .

General Definition In the previous section. More generally. one can show that the electron density per unit area is nS = N 2DF 0 (η F ) . (13) where N 2D = and m * k BT . which must be integrated numerically. ℏ π (18) F −1/ 2 (η F ) = 1 π ∫ 1 + eε −η 0 ∞ ε −1/ 2 d ε F (19) is the Fermi-Dirac integral of order −1/ 2 . π ℏ2 (14) F 0 (η F ) = ∫ 0 ∞ ε 0dε 1+ e ε −η F = ln (1 + eηF ) (15) is the Fermi-Dirac integral of order 0. which can be integrated analytically. 2. E − EC (16) and the equilibrium electron density per unit length is nL = N1DF −1/2 (η F ) . the density-of-states is 2 m* g1D ( E ) = πℏ 1 . Finally. we define 3 .and by following a procedure like that one we used in three dimensions. (17) where N1D = and 1 2 m* k B T . we saw three examples of Fermi-Dirac integrals. in one-dimensional (1D) semiconductors with a parabolic band.

(10). (20): (21b) (21c) F 1/ 2 (η F ) ≡ 1 ε 1/ 2 d ε . 4 (23) . use the general definition. (20) and Eqs. 1 π Γ(3 / 2) = Γ(1/ 2 + 1) = Γ(1/ 2) = . (21a) Also Γ(1 / 2) = π . (15) and (19). 3. (21b-c). The Γ function is just the factorial when its argument is a positive integer. Γ(1/ 2 + 1) ∫ 1 + eε −η F 0 ∞ (22a) so we need to evaluate Γ(3 / 2) . Γ( j + 1) ∫ 1 + exp(ε − η F ) 0 ∞ (20) where Γ is the gamma function. Derivatives of Fermi-Dirac Integrals Fermi-Dirac integrals have the property that dF j dη F = F j −1 . As an example. we find. (21ac) to show that the results for F 0 (η F ) and F −1/ 2 (η F ) agree with Eqs. Using Eqs. Γ( n) = (n − 1)! (for n a positive integer) . For more practice. let’s evaluate F 1/ 2 (η F ) from Eq. Eq.1 ε j dε F j (η F ) ≡ . 2 2 so F 1/ 2 (η F ) is evaluated as (22b) F 1/ 2 (η F ) ≡ 2 π ∫ 1 + eε −η 0 ∞ ε 1/ 2 d ε F . and Γ( p + 1) = pΓ( p) . (22c) which agrees with Eq.

1 = 2 p1.… . to evaluate F −4 (η F ) = eηF generated from Eqs. For example. For the non-degenerate limit. we can show that there is an analytic expression for any Fermi-Dirac integral of integer order. 222 of [4]) pk . F j (η F ) = eη F ( 1+ e ηF ) −j P− j − 2 eη F . we have an analytical expression for F 0 (η F ) . j. ( ) (25) where Pk is a polynomial of degree k. and we find p1.0 = 1 . k = − pk −1.2 = − p1. (26c) is missing in p. F −1 = dF 0 1 = .0 = 1. polynomial coefficients are p0.i = (1 + i ) pk −1.1 = 1 .0 = 1. 5 .i −1 i = 1. p1. p2. p2.0 = −1 .0 = 1 . (26a-c) as [4] (1 + e ) ηF 4 × P2 ( eηF ) . which means that we have an analytical expression for F −1 (η F ) .1 − 2 p1.i are generated from a recurrence relation [4] (note that the relation in Eq. (27) F −4 (η F ) = eη F (1 + e ) ηF 4 ∑ p2.0 = −4. the result is particularly simple.i − ( k + 1 − i ) pk −1. for j ≤ −2 . pk . ) (28) 4. Asymptotic Expansions for Fermi-Dirac Integrals It is useful to examine Fermi-Dirac integrals in the non-degenerate ( ηF << 0 ) and degenerate ( ηF >> 0 ) limits.i eiηF = i=0 2 eηF (1 + e ) ηF 4 (1 − 4e ηF + e 2ηF .k −1 . (26a) (26b) (26c) pk . k − 1 . dη F 1 + e −ηF (24) Similarly. and the coefficients pk .1 = − p0.which often comes in useful. p2. For example.

F (29) which means that for all orders. 5 8η F / 2 . tn = ∑ µ =1 (−1) µ −1 µ n = ( − 21− n )ζ (n ) . j. F −1/ 2 (η F ) → F 1/ 2 (η F ) → 1 2 2η 1/ 2 F π . note that the integrals to calculate carrier densities in Eqs. we consider the complete expansion for the Fermi-Dirac integral for j > −1 and η F > 0 [2. (30) can be related to the well-known Sommerfeld expansion [8. Specific results for several Fermi-Dirac integrals are shown below. (32) can be written as (see [9] for a detailed derivation) 6 . First.F j (η F ) → eη . 9]. (31a) (31b) (31c) (31d) (31e) 3/ 4η F 2 . (30) as F j (η F ) → η Fj +1 Γ ( j + 2 ) [7]. then we can write the ( E − EF ) . 6] F j (η F ) = 2η ∞ j +1 F ∞ ( −1) e− nηF t2 n ∑ Γ ( j + 2 − 2n )η 2n + cos (π j ) ∑ n j +1 . 3 π 2 F 1 (η F ) → η F . To examine Fermi-Dirac integrals in the degenerate limit. the Fermi-Dirac integral approaches the exponential in the non-degenerate limit. 0 (32) Taylor expansion of H ( E ) about EF as [9] ∞ If H ( E ) does not vary rapidly in the range of a few k BT about EF . n =0 n =1 F ∞ n −1 (30) where t0 = 1 2 . F 3/ 2 (η F ) → 15 π 1 3 F 2 (η F ) → η F . 6 The complete expansion in Eq. the integral in Eq. and ζ (n ) is the Riemann zeta function. 1 The expressions for the Fermi-Dirac integrals in the degenerate limit ( ηF >> 0 ) come from Eq. dn H ( E ) = ∑ n H ( E ) E =E F n! n = 0 dE n (33) Using this Taylor series expansion. (1) and (3) are all of the form ∞ -∞ ∫ H ( E ) f ( E ) dE . 5.

9]. 3 4  2  (35) and it is noted that an = 2t2 n . 7.7 % for 1/ 2 < j < 5 / 2 . The source code of this relatively short function is listed in the Appendix. In the degenerate limit. AymerichHumet et al. then Eq. (30) except that the second term in Eq. (36) If we scale E by k BT in Eq. The Matlab fuction. Equation (34) is known as the Sommerfeld expansion [8. d ε 2 n −1 ε =η F (37) Then the Sommerfeld expansion for the Fermi-Dirac integral of order j can be evaluated by letting H ( ε ) = ε j Γ ( j + 1) in Eq.m. [12] proposed a single analytic approximation that evaluates the Fermi-Dirac integral of order j = 1/ 2 with errors less than 0. so the Eqs. (10) with orders j ≥ −1/ 2 using these analytic approximations. (30) vanishes. (31a-e). the first term in the sum in Eq. Bednarczyk et al. (30) and (38) give the same results as Eqs. [13. and the error increases with larger j. 10. the second term in Eq.2 % for −1/ 2 < j < 1/ 2 and 0.” [1] calculates the Fermi-Dirac integral defined in Eq. 14] introduced an analytic approximation for a general j. We briefly mention some of the analytic approximations and refer the reader to a Matlab function. Typically. Approximate Expressions for Common Fermi-Dirac Integrals Fermi-Dirac integrals can be quickly evaluated by tabulation [2. 5. (34) becomes ∞ −∞ ∫ H (ε ) f0 (ε ) dε = ηF −∞ ∫ H ( ε ) d ε + ∑ an n =1 ∞ d 2 n −1 H (ε ) . 11] or analytic approximation [12-14]. (34) is all that is needed. 7 .∞ −∞ ∫ H ( E ) f 0 ( E ) dE = EF −∞ ∫ H ( E ) dE + ∑ ( k BT ) an 2n n =1 ∞ d 2 n −1 H (E) . dE 2 n −1 E = EF (34) where 1 1 1   an = 2 1 − 2 n + 2 n − 2 n + ⋯  . (30) is omitted [5]. (37). (34). and the result is F j (η F ) = 2η Fj +1 ∑ t2 n . “FD_int_approx. ε ≡ E k BT . however. and it gives an error of 1.4 % [3]. 2n n = 0 Γ ( j + 2 − 2 n )η F ∞ (38) Equation (38) is the same as Eq. and the result is ∞ −∞ ∫ H ( E ) f 0 ( E ) dE ≃ EF −∞ ∫ H ( E ) dE + π2 6 ( kBT ) 2 H ′ ( EF ) .

approx − F 1/ 2.m”) and the rigorous numerical integration (“FD_int_num.num represent the results from the approximate expression and the numerical integration respectively. The relative errors of the approximate expressions are calculated as (F 1/ 2. then the approximations proposed by Halen and Pulfrey [15. and the error is less than 10-5 for −1/ 2 ≤ j ≤ 7 / 2 [15]. The Fermi-Dirac integral of order j = 1 2 ( F 1/ 2 (η F ) ) is calculated for −10 ≤ η F ≤ 10 with η F spacing = 0. The source code is listed in the Appendix. as η F = F 1/ 2−1 ( n N 3 D ) .num .01 for η F max ≃ 5. and the double exponential (DE) rule is used for larger η F .m” and “FDjx. The elapsed time measured for each approach (using Matlab commands “tic/toc” for Pentium 4 CPU 3.5 and η F max ≃ 20 . There also have been discussions on the simple analytic calculation of the inverse FermiDirac integrals of order j = 1/ 2 [3]. for generalized Fermi-Dirac integrals with order j > −1 [21].If a better accuracy is required and a longer CPU time is allowed. This function includes tables of coefficients.m. but the CPU time is considerably longer.01 for η F max ≃ 5. 16] may be used.03 for η F max ≃ 5 [3]. The Matlab function. we compare the accuracy and the timing of the three approaches that calculate F j (η F ) .num ) F 1/ 2.m”) with double-precision. 8 In this approach.m.0 GB RAM) clearly shows the compromise between the accuracy and the CPU time. Nilsson later presented two empirical approximations [20] that give ∆η F ≤ 0. “FDjx.2 ×10−16 ) can be achieved after 60 to 500 iterations [21]. (11). Nilsson proposed two different full-range ( −10 ≤ η F ≤ 20 ) expressions [19] with ∆η F ≤ 0. Numerical Evaluation of Fermi-Dirac Integrals Fermi-Dirac integrals can be evaluated accurately by numerical integration. Double precision (eps.org [22].01 and ∆η F ≤ 0.” [1] evaluates the Fermi-Dirac integral numerically using the composite trapezoidal rule following the approach in [21]. The Matlab function.4 GHz and 2. Joyce and Dixon [17] examined a series approach that gives ∆η F ≤ 0. An online simulation tool that calculates the Fermi-Dirac integrals using this source code has been deployed at nanoHUB.approx and F 1/ 2. but it can be downloaded from [1].5 [3]. several approximate expressions are introduced based on the series expansion in Eq. Note that the numerical approach we consider in this note is relatively simple. This has been of particular interest because it can be used to calculate the Fermi level from the known bulk charge density in Eq. and a simpler expression from Joyce [18] gives ∆η F ≤ 0. Here we briefly review the approach by Press et al. This approach provides very high accuracy. and there are other advanced numerical integration algorithms [23] suggested to improve the calculation speed.005 [3]. so it is not simple enough to be shown in the Appendix. 1. the composite trapezoidal rule with variable transformation ε = exp (t − e− t ) is . respectively [3]. “FD_int_num. where F 1/ 2. 6.01 using approximate expressions (“FD_int_approx. (30). ~ 2. used for η F ≤ 15 .” [1] is the main function that calculates the Fermi-Dirac integrals using this model. In Fig. In this model.

All Matlab functions are available in [1].(b) The absolute values of the relative errors in the log scale.m”.Fig. 9 . (A) Relative error from “FD_int_approx.m”. The elapsed time measured for the three approaches clearly shows the trade-off between the accuracy and the CPU time. 1. (B) Relative error from “FDjx. (a) Relative errors from the approximate expressions for F 1/ 2 (η F ) with respect to the numerical integration (“FD_int_num.m”).

773. pp. no. Chase. ∞ Scientific Research. Frankel. 6. Philadelphia: Saunders College Publishing. 1981. "Fermi-Dirac functions of integral order. "Extension of McDougall-Stoner tables of the Fermi-Dirac functions. short series approximation to Fermi-Dirac integrals of order -1/2. P. Mathematical and Physical Sciences. 54. F.. pp. 5271 (1985)]. 3/2. 204. vol. "The Fermi-Dirac integrals F p (η ) = ( p !) −1 ∫ ∞ 0 ε p ( eε −η + 1) d ε . 4. Phys. Blakemore.nanohub. M. 1982. 5/2. Asymptotic Expansions: Their Derivation and Interpretation. and P. 1-32. no. "Notes on Fermi-Dirac Integrals (3rd Edition). vol. no. 409-410. Halen and D." J. Dingle. 2. 1985." ACM Trans. Mermin. vol. especially the function F 1/ 2 (η ) used to describe electron density in a semiconductor. Phys. D. 221-232. 981-982. 2264. 2850-2851. pp. 11. vol. Glasser. 1078. no." Applied −1 [4] [5] [6] Fermi--Dirac integrals F p (η ) = 1 Γ ( p + 1) ∫ ε p (1 + eε −η ) d ε . 57. vol. W. 1957. vol. X. M." J.. Bednarczyk and J. Serra-Mestres." Appl. W. "Approximations for Fermi-Dirac integrals. N. Phys. Phys. London: Academic Press. Math. vol. 237. Aymerich-Humet.References [1] [2] [3] R. and 7/2. L. Millan. 5271-5274.. pp. N." Philosophical Transactions of the Royal Society of London. W. and M. Kim and M. C. 10." Solid-State Electron. pp." Solid-State Electron. J." J. 59. 2001. Aymerich-Humet. F. S. no. 1955. 1067-1076. B. Serra-Mestres. "Erratum: "Accurate. no." J. F. Pulfrey. pp. 28. Joyce and R. "Zur Elektronentheorie der Metalle auf Grund der Fermischen Statistik. vol. R. p. L. 1. 4. L. Mathematical and Physical Sciences. Sommerfeld. 10 . "A generalized approximation of the Fermi--Dirac integrals. pp. "An analytical approximation for the Fermi-Dirac integral F3/ 2 (η ) . N. 1983. and J. 225-239. Lett. 1. 1/2. "Complete asymptotic expansions of the 0 [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] no. pp. Pulfrey. Rhodes. 3. "Accurate. Appl. X. C. 25. no. V. short series approximations to Fermi-Dirac integrals of order -1/2. 5/2. 1995. B. P. 1950. and 7/2" [J. A. 64. "The computation of Fermi-Dirac functions. vol. no. 6." Zeitschrift für Physik A Hadrons and Nuclei. 57. 2. E. vol. 21. 24. Phys. 396-405. 529-542.. 5. and J. 1977. Ashcroft and N.. "The approximation of the Fermi-Dirac integral F 1/ 2 (η ) ." Available: https://www. R. 42. P. 1976. 354-356.. M. 1986. 3. no. 1. 47. Stoner.org/resources/5475/. no. Goano. 1973. "Algorithm 745: computation of the complete and incomplete Fermi-Dirac integral. Dixon. 5. D. vol. J. 67-104. A. pp. Series A. vol. pp. Choquard." Helvetica Physica Acta. Garoni. Bednarczyk. 1938.. 1. S. no. 12." Physics Letters A. Millán. 3/2. Phys. Math.. Softw. McDougall and E. 1928. Appl. 1978. Series A. "Analytic approximations for the Fermi energy of an ideal Fermi gas." Proceedings of the Royal Society of London. pp. 1/2. no. 3. 31. pp. vol. 1860-1868. pp. Solid State Physics. Beer. Van Halen and D. Dingle. Appl. T. Appl. vol. Lundstrom (2008).

2001. K75-K78. vol. 653-654. B. 7. no. W. 361-365. no. T. Solidi (a). Nilsson. W. "An accurate approximation of the generalized Einstein relation for degenerate semiconductors. 3rd ed. New York: Cambridge University Press. M. "Empirical approximations for the Fermi energy in a semiconductor with parabolic bands. and R. Phys. 1973." Phys. 11 . 3. Vetterling. Press. vol. pp. Lundstrom. Phys. Mohankumar. Joyce." Available: http://nanohub. Lett. Numerical Recipes: The Art of Scientific Computing. vol. 1978. Teukolsky.. Lett. N. S. Natarajan and N. pp. "FD integral calculator." Appl. X. 32. 33. pp. A. "An accurate method for the generalized Fermi-Dirac integral. G. N." Comput. Stat. Phys. vol. H. 10. Sun. 680-681. Flannery. and B. Kim (2011). P. Commun. Nilsson. 137. A. 1978.[18] [19] [20] [21] [22] [23] W. pp. "Analytic approximations for the Fermi energy in (Al.Ga)As. 19." Appl.. G.org/tools/fdical. 2007. no..

.^ ( j + 1 ) ..^ -1 ./ gamma( j + 1 ) ) . S. xi = 3 * sqrt( pi ) . b = 2.. Phys./ gamma( j + 1 ) ) . a = ( 1 + 15 / 4 * ( j + 1 ) + 1 / 40 * ( j + 1 ) ^ 2 ) ^ ( 1 / 2 ). 54. b = 1. Solid-St.^ 2 ) ).. otherwise % Model proposed in [4] % Expressions from eqs.^ ( 1 / c ) ) . 24.8 + 0.org/resources/5475 % % References % [1]D./ gamma( j + 1 ). j ) % Analytic approximations for Fermi-Dirac integrals of order j > -1/2 % Date: September 29. J. Appl. Bednarczyk and J. y = ( ( j + 1 ) * 2 ^ ( j + 1 ) . Serra-Mestres.68 * exp( -0.m” function y = FD_int_approx( eta.^ -1. + exp( -x ) . c = 2 + ( 2 ./ ( b + x + ( abs( x .9. 25. F. switch j case 0 y = log( 1 + exp( x ) ). 409 (1978) % [2]J. Electron. (22)-(24) of [2] mu = x . end end 12 . 981 (1981) % [4]X. Aymerich-Humet.6 * x .b ) . Aymerich-Humet.j ). 64.^ ( 3 / 8 ) ).64.sqrt( 2 ) ) * 2 ^ ( .61 * j./ gamma( j + 1 ). + exp( -x ) .* ( 1 .^ c + a ^ c ) . Lett.^ -1 . Millan.') else x = eta.^ ( 1 / c ) ) . (5) of [3] % The integral is divided by gamma( j + 1 ) to make it consistent with [1] and [2]. Electron.^ 4 + 50 + 33. Solid-St.^ ( j + 1 ) . see: % "Notes on Fermi-Dirac Integrals (3rd Edition)" by Raseong Kim and Mark % Lundstrom at http://nanohub.17 * ( x + 1 ) . case 3/2 % Model proposed in [3] % Expressions from eq.b ) .x ) + xi ) . a = 14. % analytic expression case 1/2 % Model proposed in [1] % Expressions from eqs.Appendix “FD_int_approx. Blakemore. (6)-(7) of [4] % The integral is divided by gamma( j + 1 ) to make it consistent with [1] and [2]. c = 9 / 4. 1067 (1982) % [3]X./ ( 4 * mu .0. A. y = ( ( j + 1 ) * 2 ^ ( j + 1 ) . and J. F.^ c + a ) . and J. 2008 % Author: Raseong Kim (Purdue University) % % Inputs % eta: eta_F % j: FD integral order % % Outputs % y: value of FD integral (the "script F" defined by Blakemore (1982)) % % For more information in Fermi-Dirac integrals./ ( b + x + ( abs( x . y = ( exp( . Serra-Mestres. 2850 (1983) if j < -1/2 error( 'The order should be equal to or larger than -1/2. Bednarczyk. Millan.. Phys.

^ j . A. Nmax ) % Numerical integration of Fermi-Dirac integrals for order j > -1.* ( 1 + exp( -t ) ) . % generate intervals x = exp( t . Press. % Author: Raseong Kim (Purdue University) % Date: September 29./ ( 1+ exp( x .“FD_int_num.exp( -t ) ). if N > 1 % test for convergence err = abs( y . % limits for t b = 5. if err < tol break.eta ) ).m” function [ y N err ] = FD_int_num( eta. b. x = exp( t .') end y = y . % Numerical recipies: The art of scientific computing.org/resources/5475 % % Reference % [1] W. tol. f = x . f ). N + 1 ).0. see: % "Notes on Fermi-Dirac Integrals (3rd Edition)" by Raseong Kim and Mark % Lundstrom at http://nanohub. and B. H. 2007. Teukolsky. y = trapz( t. P.5./ gamma( j + 1 ). % % Inputs % eta: eta_F % j: FD integral order % tol: tolerance % Nmax: number of iterations limit % % Note: When "eta" is an array. Flannery. 13 . S.. % % Outputs % y: value of FD integral (the "script F" defined by Blakemore (1982)) % N: number of iterations % err: error % % For more information in Fermi-Dirac integrals. 3rd Ed. end end y_old = y. 208 % Extended (composite) trapezoidal quadrature rule with variable % transformation.exp( t ) ) % Valid for eta ~< 15 with precision ~eps with 60~500 evaluations.* x . Vetterling. T. W. this function should be executed % repeatedly for each component. end if N == Nmax error( 'Increase the maximum number of iterations. for N = 1 : Nmax a = -4.y_old ). t = linspace( a. Cambridge % University Press. j.