PHYSICAL REVIEW E 81, 036705 2010

Bulgac-Kusnezov-Nos-Hoover thermostats
Alessandro Sergi*
School of Physics, University of KwaZulu-Natal, Pietermaritzburg, Private Bag X01 Scottsville, 3209 Pietermaritzburg, South Africa

Gregory S. Ezra
Department of Chemistry and Chemical Biology, Baker Laboratory, Cornell University, Ithaca, New York 14853, USA Received 3 February 2010; published 19 March 2010 In this paper, we formulate Bulgac-Kusnezov constant temperature dynamics in phase space by means of non-Hamiltonian brackets. Two generalized versions of the dynamics are similarly dened, one where the Bulgac-Kusnezov demons are globally controlled by means of a single additional Nos variable, and another where each demon is coupled to an independent Nos-Hoover thermostat. Numerically stable and efcient measure-preserving time-reversible algorithms are derived in a systematic way for each case. The chaotic properties of the different phase space ows are numerically illustrated through the paradigmatic example of the one-dimensional harmonic oscillator. It is found that, while the simple Bulgac-Kusnezov thermostat is apparently not ergodic, both of the Nos-Hoover controlled dynamics sample the canonical distribution correctly. DOI: 10.1103/PhysRevE.81.036705 I. INTRODUCTION PACS number s : 05.10. a, 05.20.Jj, 05.45.Pq, 65.20.De

In condensed matter studies, there are many situations in which molecular dynamics simulation at constant temperature 13 is needed. For example, this occurs when magnetic systems are modeled in terms of classical spins 47 . Deterministic methods 810 , based on non-Hamiltonian dynamics 1119 , can sample the canonical distribution provided that the motion in the phase space of the relevant degrees of freedom is ergodic 1,3 . However, classical spin systems are usually formulated in terms of noncanonical variables 20,21 , without a kinetic energy expressed through momenta in phase space, so that Nos dynamics cannot be applied directly. To tackle this problem, Bulgac and Kusnezov BK introduced a deterministic constant-temperature dynamics 2224 , which can be applied to spins. A number of numerical approaches to integration of spin dynamics can be found in the literature 2528 . However, BK dynamics, as any other deterministic canonical phase space ow, is able to correctly sample the canonical distribution only if the motion in phase space is ergodic on the timescale of the simulation. In general, this condition is very difcult to check for statistical systems with many degrees of freedom, while it is known that, despite its simplicity, the one-dimensional harmonic oscillator provides a difcult and important challenge for deterministic thermostatting methods 9,2931 . In this paper, we accomplish two goals. First, by reformulating BK dynamics through non-Hamiltonian brackets 14,15 in phase space, we introduce two generalized versions of the BK time evolution which are able to sample the canonical distribution for a stiff harmonic system. Second, using a recently introduced approach based on the geometry of non-Hamiltonian phase space 19 , we are able to derive stable and efcient measure-preserving and time-reversible algorithms in a systematic way for all the phase space ows treated here.

The BK phase space ow introduces temperature control by means of ctitious coordinates and their associated momenta in an extended phase space traditionally called demons. Our generalizations of the BK dynamics are obtained by controlling the BK demons themselves by means of additional Nos-type variables 8 . In one case, the BK demons are controlled globally by means of a single additional NosHoover thermostat 8,9 . In the following this will be referred to as Bulgac-Kusnezov-Nos-Hoover BKNH dynamics. In the second case, each demon is coupled to an independent Nos-Hoover thermostat. This will be called the Bulgac-Kusnezov-Nos-Hoover chain BKNHC , and corresponds to massive NH thermostatting of the demon variables 32 . The ability to derive numerically stable measurepreserving time-reversible algorithms 19 for Nos controlled BK dynamics is very encouraging for future applications to thermostatted spin systems. This paper is organized as follows. In Sec. II, we briey sketch the unied formalism for non-Hamiltonian phase space ows and measure-preserving integration. The BK dynamics is formulated in phase space and a measurepreserving integration algorithm is derived in Sec. III. The BKNH and BKNH-chain thermostats are treated in Secs. IV and V, respectively. Numerical results for the onedimensional harmonic oscillator using these thermostats are presented and discussed in Sec. VI. Section VII reports our conclusions. In addition we include several appendices. A useful operator formula is derived in Appendix A, while invariant measures for the BK, BKNH, and BKNHC phase space ows are derived in Appendices BD, respectively.
II. NON-HAMILTONIAN BRACKETS AND MEASUREPRESERVING ALGORITHMS

*sergi@ukzn.ac.za

gse1@cornell.edu

Consider an arbitrary system admitting a timeindependent extended Hamiltonian expressed in terms of the phase space coordinates xi, i = 1 , . . . , 2N. In this case, the Hamiltonian can be interpreted as the conserved energy of the system.
2010 The American Physical Society

1539-3755/2010/81 3 /036705 14

036705-1

ALESSANDRO SERGI AND GREGORY S. EZRA

PHYSICAL REVIEW E 81, 036705 2010

2N

Upon introducing an antisymmetric tensor eld generalized Poisson tensor 21,33 in phase space, B x = BT x , one can dene non-Hamiltonian brackets 1416 as
2n

=
i,j=1

Bij H , xi x j

a,b =
i,j=1

a b Bij , xi xj

the statistical mechanics must be formulated in terms of a modied phase space measure 1218 = ew x where = dx1 dx2 . . . dx2N 9 8

where a = a x and b = b x are two arbitrary phase space functions. The bracket dened in Eq. 1 is classied as nonHamiltonian 1416 since, in general, it does not obey the Jacobi relation, i.e., in general the Jacobiator J 0, where 21 J = a, b,c + b, c,a + c, a,b , 2

is the standard phase space volume element volume form 35 and the statistical weight w x is dened by dw = dt x . 10

with c = c x arbitrary phase space function in addition to the functions a and b, previously introduced . If J 0, the tensor Bij is said to dene an almost-Poisson structure 34 such systems have also been called pseudo-Hamiltonian 33 . An energy conserving and in general non-Hamiltonian phase space ow is then dened by the vector eld
2N

It has been shown that, provided the condition ew x Bij = 0, i = 1, . . . 2N 11

xj is satised, then

xi = xi,H =
j=1

Bij

H , xj

L =0

for every

12

where conservation of H x follows directly from the antisymmetry of Bij. It has previously been shown how equilibrium statistical mechanics can be comprehensively formulated within this framework 16 . It is also possible to recast the above formalism and the corresponding statistical mechanics in the language of differential forms 17,18 . If the matrix B is invertible this is true for all the cases considered here , with inverse ij, we can dene the 2-form 35 1 = 2
ijdx i

so that the volume element is invariant under each of the L 19 . The condition 11 is satised for all the cases considered below, so that, exploiting the decomposition in Eq. 6 , algorithms derived by means of a symmetric Trotter factorization of the Liouville propagator:
ns1 ns1

exp L =
=1

exp

exp exp Lnx

exp
=1

L ns 13

are not only time-reversible but also measure preserving. dx .

j

4
III. PHASE SPACE FORMULATION OF THE BK THERMOSTAT

The dynamics of Eq. 3 is then Hamiltonian if and only if the form Eq. 4 is closed, i.e., has zero exterior derivative, d = 0 35 . This condition is independent of the particular system of coordinates used to describe the dynamics. The structure of Eq. 3 can be taken as the starting point for derivation of efcient time-reversible integration algorithms that also preserve the appropriate measure in phase space 19 . Measure-preserving algorithms can be derived upon introducing a splitting of the Hamiltonian
ns

A phase space formulation of the BK thermostat can be achieved upon introducing the Hamiltonian HBK = K1 p p2 K2 p +V q + + + k BT 2m m m =H q,p + K1 p K2 p + + k BT m m + + 14a

14b

H=
=1

H ,

which in turn induces a splitting of the Liouville operator associated with the non-Hamiltonian bracket in Eq. 1 ,
2N

L xi = xi,H

=
j=1

Bij

H . xj

When the phase space ow has a nonzero compressibility

where q , p are the physical degrees of freedom coordinates and momenta , with mass m, to be simulated at constant temperature T, while and are the BK demons, with corresponding inertial parameters m and m , and associated momenta p , p 2224 . K1 and K2 provide the kinetic energy of demon variables, and for the moment are left arbitrary. Upon dening the phase space point as x = q , , , p , p , p = x1 , x2 , x3 , x4 , x5 , x6 , one can introduce an antisymmetric BK tensor eld as

036705-2

BULGAC-KUSNEZOV-NOS-HOOVER THERMOSTATS

PHYSICAL REVIEW E 81, 036705 2010

0 0 0 BBK = 1 0 G2

0 0 0 0 G1 p 0

0 0 0 0 0 G2 q

1 0 0 0 G1 0

0 G1 p 0 G1 0 0

G2 0 G2 q 0 0 0

LBK = 1

V V + G2 , q p q p p p + G1 , m q m p G1 , p p G2 , q p G1 K1 , m p p .

18a

LBK = 2 , 15

18b

LBK = kBT 3

18c

LBK = kBT 4 1 G1 K1 m p p

18d

where G1 and G2 are functions of system variables p , q only. Substituting BBK and HBK into Eq. 3 , we obtain the energy-conserving equations q= H G2 q,p p m K2 , p 16a 16b 16c K1 , p 16d 16e 16f

LBK = 5

18e

LBK = 6

1 G2 K2 G2 K2 + m p q m q p

18f

= 1 G1 K1 , m p p = 1 G2 K2 , m q p p= H G1 q,p q m

Upon choosing a symmetric Trotter factorization of the BK Liouville operator based on the decomposition
8

BK

=
=1

LBK

19

p = G1 p = G2

G1 H , k BT p p G2 H . k BT q q

a measure-preserving algorithm can be produced in full generality. In practice, a choice of K1, K2, G1, and G2 must be made in order obtain explicit formulas. In this paper, we make the following simple choices: G1 = p, G2 = q, K1 = p2 , 2 p2 . 2 20a 20b

The associated invariant measure for the BK ow is discussed in Appendix B.

Algorithm for BK dynamics

20c

In order to derive a measure preserving algorithms, the rst step, following Eq. 5 , is to introduce a splitting of HBK: HBK = V q , 1 HBK = 2 p2 , 2m 17a 17b 17c 17d 17e BBK = K2 =

20d

In terms of Eqs. 20a 20d , the antisymmetric BK tensor becomes 0 0 0 1 0 q 0 0 0 0 1 0 0 0 0 0 0 1 0 0 p 0 1 0 0 0 q 0 1 0 0 0 , 21

HBK = kBT , 3 HBK = kBT , 4 HBK = 5 HBK = 6 K1 p , m K2 p . m

0 p

1 0

17f

and the Hamiltonian reads p2 p2 HBK = H q,p + + + k BT 2m 2m + . 22

A measure-preserving splitting of the Liouville operator then follows from Eq. 6 :

036705-3

ALESSANDRO SERGI AND GREGORY S. EZRA

PHYSICAL REVIEW E 81, 036705 2010

The split Liouville operators now simplify as follows: V BK = V L1 +q , q p q p p2 BK = p + , L2 m q m p BK = k T L3 B , 23a

Using the so-called direct translation technique 36 we can expand the above symmetric breakup of the Liouville operator into a pseudocode form, ready to be implemented on the computer: q i p q+ p 4m

23b

BK :UB

23c

p + Fp 4 p 2m p 2m

BK = k T L4 B BK = p L5 m

,
2

23d

p p exp q q exp

p p p + p m m +

23e

ii

p + 2m + p 2m p 4m

BK :UC

BK = p q + p L6 q m m

p2 . m p

23f

For the purposes of dening an efcient integration algorithm, we combine commuting Liouville operators as follows: iii
BK LA

q p

q+

BK :UB

BK + BK = F q L1 L4

+ Fp

24a

p + Fp 4 p+ F p + Fp q+ + p 4m p 4m
BK :UB

BK LB

BK + BK = p + Fp L2 L3 , p m q + p m

24b

iv

p p

BK :UA

BK LC

BK + BK = p p p q + p L5 L6 p m q m m

, q 24c v

where F q = V/ q, Fp = q V kBT, q 25a

p 25b

p + Fp 4 p 2m p 2m

Fp = Dening UBK

p2 kBT. m

p p exp 25c q q exp vi 26 p + 2m + p 2m

BK :UC

= exp BK , L

= A , B , C, one possible reversible measurewhere preserving integration algorithm for the BK thermostat is then U
BK BK = UB

BK UC

BK UB

BK UA

q vii 27 p

q+

p 4m

BK :UB

BK UB

BK UC

BK UB

p + Fp 4

036705-4

BULGAC-KUSNEZOV-NOS-HOOVER THERMOSTATS IV. BULGAC-KUSNEZOV-NOS-HOOVER DYNAMICS

PHYSICAL REVIEW E 81, 036705 2010 Algorithm for BKNH dynamics

The BKNH Hamiltonian H

BKNH

The Hamiltonian can be split as HBKNH = V q , 1 HBKNH = 2 p2 , 2m 31a 31b 31c 31d 31e

p2 K1 p K2 p = H q,p + + + m m 2m + k BT + + 2kBT 28

is simply the BK Hamiltonian augmented by the Nos variables , p with mass m . With the antisymmetric BKNH tensor
BKNH

HBKNH = kBT , 3 HBKNH = kBT , 4 HBKNH = 5 HBKNH = 6 K1 p , m K2 p , m p2 , 2m

0 0 0 = 0 1 0 G2 0

0 0 0 0 0 G1 p 0 0

0 0 0 0 0 0 G2 q 0

0 0 0 0 0 0 0 1

1 0 0 0 0 G1 0 0

0 G1 p 0 0 G1 0 0 p

G2 0 G2 q 0 0 0 0 p

0 0

31f

0 1 0 p p 0 29 yields LBKNH = 1 ,

HBKNH = 7

31g 31h

HBKNH = 2kBT , 8

The measure-preserving splitting 19 of the Liouville operator L = BBKNH ij HBKNH , xi xj 32

we obtain from Eq. 3 equations of motion for the phase space variables x = q , , , , p , p , p , p = x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 : q= H G2 q,p p m K2 , p 30a 30b

V V + G2 , q p q p p p + G1 , m q m p G1 , p p G2 , q p K1 , p p

33a

LBKNH = 2

33b

= 1 G1 K1 , m p p = 1 G2 K2 , m q p p = , m H G1 q,p p= q m K1 , p

LBKNH = kBT 3 30c LBKNH = kBT 4 30d 30e 30f LBKNH = 6 30g 30h LBKNH = 7 p m LBKNH = 5 1 G1 K1 m p p

33c

33d

p G1 K1 + m p p m

G1 H p p k BT , p = G1 p m p p = G2 p = p m G2 H p p k BT , q m q K1 p + p m K2 2kBT. p

33e 1 G2 K2 G2 K2 + m p q m q p + p m K2 , p p 33f p p p p , p p m m . 33g

Here, a single Nos variable is coupled to both of the BK demons and . The associated invariant measure is discussed in Appendix C.
036705-5

LBKNH = 2kBT 8

33h

ALESSANDRO SERGI AND GREGORY S. EZRA

PHYSICAL REVIEW E 81, 036705 2010

At this stage, we leave the general formulation and adopt the particular choice of K1, K2, G1, and G2 given in Eq. 20 . The antisymmetric BKNH tensor becomes 0 0 0 BBKNH = 0 1 0 q 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 p 0 0 1 0 0 0 0 p q 0 1 0 0 0 0 p 0 0 0 1 0 p p 0 34 and the Hamiltonian simplies to
2 2 2 BKNH = H q,p + p + p + p + k T H B 2m 2m 2m

BKNH LB
BKNH LC

BKNH + BKNH = p + Fp L2 L3 p m q

37b

BKNH + BKNH + BKNH L5 L6 L8 = p p p p q + p m q m m + p m + Fp p , 37c

0 p

, where F q = V , q

38a

1 0

Fp = q + + 2 BT . 35

V kBT, q

38b

Fp =

p2 kBT, m

38c

The split Liouville operators are now V BKNH = V +q , L1 q p q p p2 BKNH = p + , L2 m q m p BKNH = k T L3 B , 36a

Fp =

p2 p2 + 2kBT. m m

38d

In LA there appears an operator with the form Li = pk p i + F pi mk pi , 39

36b

36c

where k , i = , for LA. The action of the propagator associated with this operator on pi is derived in Appendix A, and is given by e Li p i = p ie
pk/mk

+ F pie

pk/2mk

BKNH = k T L4 B BKNH = p L5 m

36d The apparently singular function 36e pk 2mk

1

pk 2mk

sinh

pk . 2mk 40

p2 p p + , p m p m + p2 , m p

sinh

pk 2mk

41

BKNH = p q + p L6 q m m BKNH = p L7 m

36f

p p p p , p p m m .

36g

is in fact well behaved as pk 0, and can be expanded in a Maclaurin series to suitably high order 37 . In our implementation we used an eighth order expansion. The propagators for the BKNH dynamics can now be dened as UBKNH = exp BKNH , L 42

BKNH = 2k T L8 B

36h

For the purposes of dening an efcient integration algorithm, we combine commuting Liouville operators as follows:
BKNH LA

= A , B , C. One possible reversible measurewhere preserving integration algorithm for the BKNH thermostat can then be derived from the following Trotter factorization: U
BKNH BKNH = UB

BKNH UC

BKNH UB

BKNH UA

BKNH + BKNH + BKNH L4 L7 L1 =F q p + p m p p p + p + Fp p m m p 37a

BKNH UB

BKNH UC

BKNH UB

43

The direct translation technique gives the following pseudocode:

036705-6

BULGAC-KUSNEZOV-NOS-HOOVER THERMOSTATS

PHYSICAL REVIEW E 81, 036705 2010

q i p p q ii

q+

p 4m

p
BKNH :UB

p exp q exp + + p 2m p 2m

p + Fp 4 p exp q exp + + p 2m p 2m p 2m p 2m

, q vi

p 2m p 2m
BKNH :UC

BKNH :UC

p + Fp 2 p 4m

q vii p

q+

p + Fp 2 q q+ p 4m

BKNH :UB

p + Fp 4

iii p p p iv p

BKNH :UB

p + Fp 4 p+ F q p + Fp + p m p m

V. BULGAC-KUSNEZOV-NOS-HOOVER CHAIN

BKNH :UA

For simplicity, we explicitly treat only the case, in which the p and p demons are each coupled to a standard NosHoover thermostat length one . It would be straightforward to couple each of the demons to NH chains 32 , and the general case can be easily inferred from what follows. Dene the Hamiltonian HBKNHC = H q,p + + k BT p2 p2 K1 p K2 p + + + m m 2m 2m + . 44

p exp p q+ 4m p + Fp 4

+ +

q v p

BKNH :UB

Upon dening the phase space point x = q , , , , , p , and the p , p , p , p = x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 , x9 , x10 antisymmetric BKNHC tensor

0 0 0 0 BBKNHC = 0 1 0 G2 0 0

0 0 0 0 0 0 G1 p 0 0 0

0 0 0 0 0 0 0 G2 q 0 0

0 0 0 0 0 0 0 0 1 0

0 0 0 0 0 0 0 0 0 1

1 0 0 0 0 0 G1 0 0 0

0 G1 p 0 0 0 G1 0 0 p 0

G2 0 G2 q 0 0 0 0 0 0 p

0 0 0 1 0 0 p 0 0 0

0 0 0 0 1 0 0 p 0 0 , 45

036705-7

ALESSANDRO SERGI AND GREGORY S. EZRA

PHYSICAL REVIEW E 81, 036705 2010

associated non-Hamiltonian equations of motion are HBKNHC xj

BBKNHC 0 46 0 0 0 = 0 1 0 q 0 47a 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 p 0 0 0 1 0 0 0 0 0 p 0 q 0 1 0 0 0 0 0 0 p 0 0 0 1 0 0 p 0 0 0 0 0 0 0 1 0 0 p 0 0 49 ,

xi =

BBKNHC ij

with i = 1 , . . . , 10.
Algorithm for BKNHC chain dynamics

0 p

Splitting the BKNHC chain Hamiltonian as HBKNHC = V q , 1 p , 2m

2

1 0

HBKNHC = 2

47b

the Hamiltonian p2 p2 p2 p2 + + + HBKNHC = H q,p + 2m 2m 2m 2m

HBKNHC = kBT , 3 HBKNHC = kBT 4

47c

+ k BT

+ +

50

and associated Liouville operators , 47d V BKNHC = V +q , L1 q p q p p2 BKNHC = p + , L2 m q m p BKNHC = k T L3 B 47f BKNHC = k T L4 B 47g BKNHC = p L5 m p , 51d 51e 51f 51g 51h 51i 51j p , 51a 51b 51c

HBKNHC = 5

K1 p , m

47e

K2 p , HBKNHC = 6 m

HBKNHC = 7

p2 , 2m

p2 p p + , p m p m + p2 , m p

HBKNHC = kBT , 8 p2 HBKNHC = , 9 2m

47h

BKNHC = p q + p L6 q m m BKNHC = p L7 m

p p , p m p ,

47i

BKNHC = k T L8 B BKNHC = p L9 m

BKNHC = k BT , H10

47j

p p , p m p .

we obtain the corresponding measure-preserving splitting of the Liouville operator HBKNHC . xi xj

BKNHC = k T L10 B

We combine commuting Liouville operators as follows: L = BBKNHC ij 48

BKNHC LA

BKNHC + BKNHC + BKNHC L4 L9 L1 =F q p + p m + p p + Fp m p , 52a

At this stage we go directly to Eq. 20 . The antisymmetric Nos-Hoover-Bulgac-Kusnezov tensor becomes

036705-8

BULGAC-KUSNEZOV-NOS-HOOVER THERMOSTATS
BKNHC LB

PHYSICAL REVIEW E 81, 036705 2010

BKNHC + BKNHC + BKNHC L2 L3 L7 p p + = m q m p + p + Fp m p , 52b

Li =

pk p i + F pi mk

pi

54

where k , i = , for LA and k , i = , for LB. Again following the derivation in Appendix A, we nd e Li p i = p ie
pk/mk

BKNHC LC

BKNHC + BKNHC + BKNHC + BKNHC L8 L10 L5 L6 = + p p p p q + p m q m m p m + Fp p + Fp V , q p , 52c

+ F pie

pk/2mk

pk 2mk

sinh

pk . 2mk 55

where F q = Fp = q Fp =

The function 2mk 1sinh order expansion 37 . The propagators UBKNHC

pk

pk 2mk

is treated through an eighth

53a 53b 53c 53d

= exp BKNHC , L

56

V kBT, q

with = A , B , C can now be introduced. One possible reversible measure-preserving integration algorithm for the BKNHC chain thermostat is then U
BKNHC BKNHC = UB

p2 kBT, m

p2 Fp = kBT, m p2 kBT. Fp = m Both in the form

BKNHC LA

BKNHC UC

BKNHC UB

BKNHC UA BKNHC UB BKNHC UC BKNHC UB

. 57

53e

and

BKNHC LB

there appears an operator with

In pseudocode form, we have the resulting integration algorithm:

q i p

q+ +

p 4m p 4m
/4 p /m BKNHC :UB

p e

+ F p e 4

/4 p /2m

p 4 2m

sinh

p 4 2m

p q

p exp q exp + p 2m

p 2m p 2m

ii p p

p + 2m

BKNHC :UC

p + Fp 2 p + Fp 2
036705-9

ALESSANDRO SERGI AND GREGORY S. EZRA

PHYSICAL REVIEW E 81, 036705 2010

q iii p

q+ +

p 4m p 4m
/4 p /m BKNHC :UB

p e p

+ F p e 4

/4 p /2m

p 4 2m

sinh

p 4 2m

p+ F + p m
p /m BKNHC :UA

iv p

p e p 4m p 4m

+ F p e

p /2m

p 2m

sinh

p 2m

q v p

q+ +

BKNHC :UB

p e

/4 p /m

+ F p e 4 p q

/4 p /2m

p 4 2m

sinh

p 4 2m

p exp q exp + p 2m

p 2m p 2m

vi p p p 4m p 4m
/4 p /m

p + 2m

BKNHC :UC

p + Fp 2 p + Fp 2

q vii p

q+ +

BKNHC :UB

p e

+ F p e 4

/4 p /2m

p 4 2m

sinh

p 4 2m

VI. NUMERICAL RESULTS

In its simplicity, the dynamics of a harmonic mode in one dimension is a paradigmatic example for checking the chaotic ergodic properties of constant-temperature phase space ows and the correct sampling of the canonical distribution. It is well known that it is necessary to generalize basic Nos-

Hoover dynamics 1,8,9 to thermostats such as the NosHoover chain 32,37 in order to produce correct constanttemperature averages for systems such as the harmonic oscillator. Some time ago, BK dynamics was devised to provide a deterministic thermostat for systems such as classical spins

036705-10

BULGAC-KUSNEZOV-NOS-HOOVER THERMOSTATS
0.08
1.04 1.03 H
BKNH

PHYSICAL REVIEW E 81, 036705 2010

BKNHC

0.06 0.04 0.02 0 P

1.02 1.01 1 0 500

4 3 2 1 0 -1 -2 -3 -4 -4-3-2-1 0 1 2 3 4 q

BK

1000 1500 2000 2500 t

0 0.5 1 1.5 2 2.5 3 3.5 4 r

FIG. 1. Comparison of the total extended Hamiltonian versus time normalized with respect to its value at t = 0 for the harmonic oscillator undergoing simple Bulgac-Kusnezov dynamics HBK , NH controlled Bulgac-Kusnezov dynamics HBKNH , and BulgacKusnezov-Nos-Hoover chain dynamics HBKNHC . Two curves have been displaced vertically for clarity. The time-reversible measure-preserving algorithms developed in this paper conserve the extended Hamiltonian to high accuracy in all three cases.

FIG. 2. Radial phase space probability for a harmonic oscillator under Bulgac-Kusnezov dynamics. The continuous line shows the theoretical value while the black bullets display the numerical results. The inset displays a plot of the phase space distribution of points along the single trajectory used to compute the radial probability.

23,24 . To ensure efcient thermostatting, BK found it necessary to introduce several demons per thermostatted degree of freedom, where each demon was taken to have a different and in principle complicated coupling to the system degree of freedom 23,24 . In the present work, we keep the form of the system-thermostat coupling as simple as possible, in order to facilitate the formulation of explicit, reversible and measure-preserving integrators 19 . It is then of interest to investigate the ability of our BK-type thermostats to produce the correct canonical sampling in the case of the harmonic oscillator. Interest in harmonic modes is also justied by the possibility of devising models of condensed matter systems in terms of coupled spins and harmonic modes, as already done in quantum dynamics with so-called spin-boson models 38 . We therefore investigate the performance of our integration schemes on the simple one-dimensional harmonic oscillator. For the particular calculations reported here, the oscillator angular frequency, all masses and kBT were taken to be unity. The time step in all cases was = 0.0025, and all runs were calculated for 106 time steps, starting from the same initial conditions: harmonic oscillator coordinate q = 0.3, all other phase space variables zero at t = 0. The measure-preserving algorithms derived here result in stable numerical integration for all the three cases treated: BK, BKNH, and BKNHC chain dynamics. Figure 1 shows the three extended Hamiltonians normalized by their respective initial time value versus time. All three Hamiltonians are numerically conserved by the corresponding measurepreserving algorithm to very high accuracy which is maintained in all the three cases . However, the basic BK phase space ow is not capable of producing the correct canonical sampling for a harmonic mode. This can be easily checked since the canonical distribution function of the harmonic oscillator is isotropic in phase space and its radial dependence can be calculated exactly. Details of this way of visualizing the phase space sampling have already been given in 14,15 . Figure 2, displaying the comparison between the theoretical and the calculated value of the radial probability in phase space,

clearly shows that the BK dynamics is not able to produce canonical sampling. A look at the inset of Fig. 2, showing the phase space distribution for the harmonic mode, also immediately shows that the dynamics is not ergodic. The same analysis has been carried out for BKNH and BKNHC phase space ows, and these are displayed in Figs. 3 and 4, respectively. Within numerical errors, both BKNH and BKNHC thermostats are able to produce the correct canonical distribution for the stiff harmonic modes. Introduction of a single, global Nos-type variable in the BKNH thermostat effectively introduces additional coupling between the two demon variables. The effectiveness of the BKNH thermostat is consistent with our ndings results not discussed here that introduction of explicit coupling between demons in BK thermostat dynamics also leads to efcient thermostatting of the harmonic oscillator.
VII. CONCLUSIONS

We have formulated Bulgac-Kusnezov 23,24 , NosHoover controlled Bulgac-Kusnezov, and Bulgac-KusnezovNos-Hoover chain thermostats in phase space by means of
0.08 0.06 0.04 0.02 0 0 0.5 1 1.5 2 2.5 3 3.5 4 r P
p 4 3 2 1 0 -1 -2 -3 -4 -4-3-2-1 0 1 2 3 4 q

FIG. 3. Radial phase space probability for a harmonic oscillator under Nos-Hoover controlled Bulgac-Kusnezov dynamics. The continuous line shows the theoretical value while the black bullets display the numerical results. The inset displays a plot of the phase space distribution of points along the single trajectory used to compute the radial probability.

036705-11

ALESSANDRO SERGI AND GREGORY S. EZRA

0.08 0.06 0.04 0.02 0 0 0.5 1 1.5 2 2.5 3 3.5 4 r P
p

PHYSICAL REVIEW E 81, 036705 2010

4 3 2 1 0 -1 -2 -3 -4 -4-3-2-1 0 1 2 3 4 q

pi

exp

pk p i + F pi mk

pi

pi ,

A3a

=pie

pk/mk

mk F p 1 e pk i

pk/mk

A3b

sinh =pie
pk/mk

+ F pie

pk/2mk

FIG. 4. Radial phase space probability for a harmonic oscillator under Bulgac-Kusnezov-Nos-Hoover chain dynamics. The continuous line shows the theoretical value while the black bullets display the numerical results. The inset displays a plot of the phase space distribution of points along the single trajectory used to compute the radial probability.

pk 2mk . pk 2mk

A3c

APPENDIX B: INVARIANT MEASURE OF THE BK PHASE SPACE FLOWS

The phase space compressibility of the phase space BK thermostat is

BK =

non-Hamiltonian brackets 14,15 . We have derived timereversible measure-preserving algorithms 19 for these three cases and showed that additional control by a single NosHoover thermostat or independent Nos-Hoover thermostats is necessary to produce the correct canonical distribution for a stiff harmonic mode. Measure-preserving dynamics of the kind discussed here is associated with equilibrium simulations where, for example, there is a single temperature parameter T . Stationary phase space distributions associated with nonequilibrium situations are much more complicated than the smooth equilibrium densities analyzed in the present paper 11,39,40 . Nonequilibrium simulations of heat ow could be carried out by extending the present approach to multimode systems e.g., a chain of oscillators coupled to BK-type demons with associated NH thermostats corresponding to two different temperatures 4143 . The techniques presented here for derivation and implementation of thermostats have been shown to be efcient and versatile. We anticipate that analogous approaches can be usefully applied to systems of classical spins coupled to both harmonic and anharmonic modes.

BBK HBK 1 G1 K1 1 G2 K2 ij = . xi xi m p p m q p B1

Upon introducing the function

BK HT = H +

K1 K2 + , m m

B2

one can easily nd that

BK = BK 1 dHT , kBT dt

B3

so that the invariant measure in phase space reads d = dx exp

t BK =dx exp HT ,

dt

BK

B4a

B4b + , B4c

=dx exp HBK exp as desired.

APPENDIX A: OPERATOR FORMULA

We wish to determine the action of the propagator associated with the Liouville operator Eq. 39 . This is equivalent to solving the evolution equation recall i k pk dpi = p i + F pi , dt mk from t = 0 to t = . Integrating, we have mk pk ln pi + F pi pk mk giving

APPENDIX C: INVARIANT MEASURE OF THE BKNH PHASE SPACE FLOWS

The phase space compressibility of the NH controlled Bulgac-Kusnezov thermostat is BBKNH HBKNH ij BKNH = xi xi = 1 G1 K1 1 G2 K2 p 2 . m p p m q p m C1

A1

Upon introducing the function = ,

0

A2

BKNH HT =H+

p2 K1 K2 + + , m m 2m

C2

we have
036705-12

BULGAC-KUSNEZOV-NOS-HOOVER THERMOSTATS
BK 1 dHT , kBT dt

PHYSICAL REVIEW E 81, 036705 2010

BKNH =

C3

BKNHC =

BBKNHC HBKNHC ij xi xi 1 G1 K1 1 G2 K2 p p . m p p m q p m m D1

so that the invariant measure in phase space is

d = dx exp
t

dt

BKNH

C4a

Upon introducing the function

BKNHC =H+ HT

p2 p2 K1 K2 + + + , m m 2m 2m
BK 1 dHT , kBT dt

D2

BKNH , =dx exp HT

C4b

we have
BKNHC =

D3

=dx exp HBKNH exp

+ +2

C4c

so that the invariant measure in phase space reads d = dx exp dt

t BKNHC =dx exp HT , BKNHC

D4a D4b + + + . D4c

APPENDIX D: INVARIANT MEASURE OF THE BKNHC CHAIN PHASE SPACE FLOWS

The phase space compressibility of the Nos-HooverBulgac-Kusnezov chain is

=dx exp HBKNHC exp

1 S. Nos, Prog. Theor. Phys. 103, 1 1991 . 2 D. Frenkel and B. Smit, Understanding Molecular Simulation: From Algorithms to Applications, 2nd ed. Academic, New York, 2001 . 3 B. Leimkuhler and S. Reich, Simulating Hamiltonian Dynamics Cambridge University Press, Cambridge, 2004 . 4 K. Chen and D. P. Landau, Phys. Rev. B 49, 3266 1994 . 5 A. Bunker, K. Chen, and D. P. Landau, Phys. Rev. B 54, 9259 1996 . 6 H. G. Evertz and D. P. Landau, Phys. Rev. B 54, 12302 1996 . 7 B. V. Costa, J. E. R. Costa, and D. P. Landau, J. Appl. Phys. 81, 5746 1997 . 8 S. Nos, J. Chem. Phys. 81, 511 1984 . 9 W. G. Hoover, Phys. Rev. A 31, 1695 1985 . 10 J. Jellinek, J. Phys. Chem. 92, 3163 1988 . 11 D. J. Evans and G. P. Morriss, Statistical Mechanics of Nonequilibrium Liquids Academic, New York, 1990 . 12 M. E. Tuckerman, C. J. Mundy, and G. J. Martyna, Europhys. Lett. 45, 149 1999 . 13 M. E. Tuckerman, Y. Liu, G. Ciccotti, and G. J. Martyna, J. Chem. Phys. 115, 1678 2001 . 14 A. Sergi and M. Ferrario, Phys. Rev. E 64, 056125 2001 . 15 A. Sergi, Phys. Rev. E 67, 021101 2003 . 16 A. Sergi and P. V. Giaquinta, J. Stat. Mech. 2007, P02013 2007 . 17 G. S. Ezra, J. Math. Chem. 32, 339 2002 . 18 G. S. Ezra, J. Math. Chem. 35, 29 2004 . 19 G. S. Ezra, J. Chem. Phys. 125, 034104 2006 . 20 A. Bulgac and D. Kusnezov, Ann. Phys. 199, 187 1990 . 21 J. E. Marsden and T. S. Ratiu, Introduction to Mechanics and

Symmetry Springer-Verlag, New York, 1999 . 22 A. Bulgac and D. Kusnezov, Phys. Rev. A 42, 5045 1990 . 23 D. Kusnezov, A. Bulgac, and W. Bauer, Ann. Phys. 204, 155 1990 . 24 D. Kusnezov and A. Bulgac, Ann. Phys. 214, 180 1992 . 25 J. Frank, W. Z. Huang, and B. Leimkuhler, J. Comput. Phys. 133, 160 1997 . 26 T. Arponen and B. Leimkuhler, BIT Numerical Math. 44, 403 2004 . 27 S. H. Tsai, M. Krech, and D. P. Landau, Braz. J. Phys. 34, 384 2004 . 28 R. I. McLachlan and D. R. J. ONeale, J. Phys. A 39, L447 2006 . 29 W. G. Hoover, C. G. Hoover, and D. J. Isbister, Phys. Rev. E 63, 026209 2001 . 30 F. Legoll, M. Luskin, and R. Moeckel, Arch. Ration. Mech. Anal. 184, 449 2007 . 31 F. Legoll, M. Luskin, and R. Moeckel, Nonlinearity 22, 1673 2009 . 32 G. J. Martyna, M. L. Klein, and M. E. Tuckerman, J. Chem. Phys. 97, 2635 1992 . 33 J. D. Ramshaw, J. Non-Equilib. Thermodyn. 16, 33 1991 . 34 A. C. da Silva and A. Weinstein, Geometric Models for Noncommutative Algebra AMS, New York, 1999 . 35 B. Schutz, Geometrical Methods of Mathematical Physics Cambridge University Press, Cambridge, 1980 . 36 M. Tuckerman, B. J. Berne, and G. J. Martyna, J. Chem. Phys. 97, 1990 1992 . 37 G. J. Martyna, M. E. Tuckerman, D. J. Tobias, and M. L. Klein, Mol. Phys. 87, 1117 1996 . 38 A. J. Leggett, S. Chakravarty, A. T. Dorsey, M. P. A. Fisher,

036705-13

ALESSANDRO SERGI AND GREGORY S. EZRA A. Garg, and W. Zwerger, Rev. Mod. Phys. 59, 1 1987 . 39 W. G. Hoover, J. Chem. Phys. 109, 4164 1998 . 40 J. R. Dorfman, An Introduction to Chaos in Nonequilibrium Statistical Mechanics Cambridge University Press, Cambridge, 1999 . 41 C. J. Mundy, S. Balasubramanian, K. Bagchi, M. E. Tucker-

PHYSICAL REVIEW E 81, 036705 2010 man, G. J. Martyna, and M. L. Klein, Nonequilibrium Molecular Dynamics Wiley-VCH, New York, 2000 , Vol. 14, pp. 291397. 42 W. G. Hoover, Mol. Simul. 33, 13 2007 . 43 W. G. Hoover and C. G. Hoover, J. Chem. Phys. 126, 164113 2007 .

036705-14