Advances In Quantum Monte Carlo Shigenori Tanaka Stuart M Rothstein

cloughekoe 4 views 87 slides May 19, 2025
Slide 1
Slide 1 of 87
Slide 1
1
Slide 2
2
Slide 3
3
Slide 4
4
Slide 5
5
Slide 6
6
Slide 7
7
Slide 8
8
Slide 9
9
Slide 10
10
Slide 11
11
Slide 12
12
Slide 13
13
Slide 14
14
Slide 15
15
Slide 16
16
Slide 17
17
Slide 18
18
Slide 19
19
Slide 20
20
Slide 21
21
Slide 22
22
Slide 23
23
Slide 24
24
Slide 25
25
Slide 26
26
Slide 27
27
Slide 28
28
Slide 29
29
Slide 30
30
Slide 31
31
Slide 32
32
Slide 33
33
Slide 34
34
Slide 35
35
Slide 36
36
Slide 37
37
Slide 38
38
Slide 39
39
Slide 40
40
Slide 41
41
Slide 42
42
Slide 43
43
Slide 44
44
Slide 45
45
Slide 46
46
Slide 47
47
Slide 48
48
Slide 49
49
Slide 50
50
Slide 51
51
Slide 52
52
Slide 53
53
Slide 54
54
Slide 55
55
Slide 56
56
Slide 57
57
Slide 58
58
Slide 59
59
Slide 60
60
Slide 61
61
Slide 62
62
Slide 63
63
Slide 64
64
Slide 65
65
Slide 66
66
Slide 67
67
Slide 68
68
Slide 69
69
Slide 70
70
Slide 71
71
Slide 72
72
Slide 73
73
Slide 74
74
Slide 75
75
Slide 76
76
Slide 77
77
Slide 78
78
Slide 79
79
Slide 80
80
Slide 81
81
Slide 82
82
Slide 83
83
Slide 84
84
Slide 85
85
Slide 86
86
Slide 87
87

About This Presentation

Advances In Quantum Monte Carlo Shigenori Tanaka Stuart M Rothstein
Advances In Quantum Monte Carlo Shigenori Tanaka Stuart M Rothstein
Advances In Quantum Monte Carlo Shigenori Tanaka Stuart M Rothstein


Slide Content

Advances In Quantum Monte Carlo Shigenori Tanaka
Stuart M Rothstein download
https://ebookbell.com/product/advances-in-quantum-monte-carlo-
shigenori-tanaka-stuart-m-rothstein-4338138
Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.
Advances In Quantum Monte Carlo James B Anderson And Stuart M
Rothstein Eds
https://ebookbell.com/product/advances-in-quantum-monte-carlo-james-b-
anderson-and-stuart-m-rothstein-eds-4337930
Recent Advances In Quantum Monte Carlo Methods Part Ii William A
Lester Stuart M Rothstein Shige Tanaka Malvin H Kalos William A Lester
Jr
https://ebookbell.com/product/recent-advances-in-quantum-monte-carlo-
methods-part-ii-william-a-lester-stuart-m-rothstein-shige-tanaka-
malvin-h-kalos-william-a-lester-jr-51266992
Advances In Quantum Chemical Topology Beyond Qtaim Juan I Rodriguez
https://ebookbell.com/product/advances-in-quantum-chemical-topology-
beyond-qtaim-juan-i-rodriguez-49169278
Advances In Quantum Chemistry 56 1st Edition John R Sabin And Erkki
Brndas Eds
https://ebookbell.com/product/advances-in-quantum-chemistry-56-1st-
edition-john-r-sabin-and-erkki-brndas-eds-2175174

Advances In Quantum Mechanics Paul Bracken Editor
https://ebookbell.com/product/advances-in-quantum-mechanics-paul-
bracken-editor-36155264
Advances In Quantum Theory I Cotaescu
https://ebookbell.com/product/advances-in-quantum-theory-i-
cotaescu-4113416
Advances In Quantum Chemistry 43 John R Sabin Et Al
https://ebookbell.com/product/advances-in-quantum-chemistry-43-john-r-
sabin-et-al-4174946
Advances In Quantum Chemistry Per Olov Lowdin
https://ebookbell.com/product/advances-in-quantum-chemistry-per-olov-
lowdin-4174948
Advances In Quantum Chemistry Jr Sabin Eds
https://ebookbell.com/product/advances-in-quantum-chemistry-jr-sabin-
eds-4174950

Advances in Quantum
Monte CarloDownloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.fw001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.fw001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

ACS SYMPOSIUM SERIES 1094
Advances in Quantum
Monte Carlo
Shigenori Tanaka ,Editor
Kobe University
Kobe, Hyogo, Japan
Stuart M. Rothstein ,Editor
Brock University
St. Catharines, Ontario, Canada
William A. Lester ,Jr.,Editor
University of California, Berkeley
Berkeley ,California, United States
Sponsor ed by the
ACS Division of Physical Chemistry
American Chemical Society ,Washington, DC
Distributed in print by Oxford University Press, Inc.Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.fw001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Library of Congr ess Cataloging -in-Publication Data
Advances in quantum Monte Carlo /Shigenori Tanaka, Stuart M. Rothstein, William A.
Lester ,Jr.,editor[s] ;sponsored by the ACS Division of Physical Chemistry .
p. cm. --(ACS symposium series ;1094)
Includes bibliographical references and index.
ISBN 978 -0-8412 -2750 -7(alk. paper)
1. Monte Carlo method --Congresses. 2. Quantum chemistry --Congresses. 3. Quantum
theory --Congresses. I.Tanaka, Shigenori. II. Rothstein, Stuart M. III. Lester ,W.A. IV.
American Chemical Society .Division of Physical Chemistry .
QD462.6.M66A382 2012
541′.28 --dc23
2012001702
The paper used in this publication meets the minimum requirements of American National
Standard for Information Sciences—Permanence of Paper for Printed Library Materials,
ANSI Z39.48n1984.
Copyright ©2012 American Chemical Society
Distributed in print by Oxford University Press, Inc.
All Rights Reserved. Reprographic copying beyond that permitted by Sections 107 or 108
of the U.S. Copyright Act isallowed for internal use only ,provided that aper -chapter fee of
$40.25 plus $0.75 per page ispaid to the Copyright Clearance Center ,Inc., 222 Rosewood
Drive, Danvers, MA 01923, USA. Republication or reproduction for sale of pages in this
book ispermitted only under license from ACS. Direct these and other permission requests
to ACS Copyright Of fice, Publications Division, 1155 16th Street, N.W .,Washington, DC
20036.
The citation of trade names and /or names of manufacturers in this publication isnot to be
construed as an endorsement or as approval by ACS of the commercial products or services
referenced herein; nor should the mere reference herein to any drawing, specification,
chemical process, or other data be regarded as alicense or as aconveyance of any right
or permission to the holder ,reader ,or any other person or corporation, to manufacture,
reproduce, use, or sell any patented invention or copyrighted work that may in any way be
related thereto. Registered names, trademarks, etc., used in this publication, even without
specific indication thereof, are not to be considered unprotected by law .
PRINTED IN THE UNITED ST ATES OF AMERICADownloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.fw001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

For eword
The ACS Symposium Series was first published in 1974 to provide a
mechanism for publishing symposia quickly in book form. The purpose of
the series is to publish timely ,comprehensive books developed from the ACS
sponsored symposia based on current scientific research. Occasionally ,books are
developed from symposia sponsored by other organizations when the topic isof
keen interest to the chemistry audience.
Before agreeing to publish abook, the proposed table of contents isreviewed
for appropriate and comprehensive coverage and for interest tothe audience. Some
papers may be excluded to better focus the book; others may be added to provide
comprehensiveness. When appropriate, overview or introductory chapters are
added. Drafts of chapters are peer -reviewed prior to final acceptance or rejection,
and manuscripts are prepared in camera -ready format.
As arule, only original research papers and original review papers are
included in the volumes. Verbatim reproductions of previous published papers
are not accepted.
ACS Books DepartmentDownloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.fw001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Pr eface
The chapters in this monograph are contributions from the Advances in
Quantum Monte Carlo symposium held at Pacifichem 2010, International
Chemical Congress of Pacific Basin Societies. The symposium was dedicated to
celebrate the career of James B. Anderson, anotable researcher in the field.
Quantum Monte Carlo provides an ab initio solution to the Schroedinger
equation by performing arandom walk through configuration space in imaginary
time. Benchmark calculations suggest that its most commonly -used variant,
“fixed -node” dif fusion Monte Carlo, estimates ener gies with an accuracy
comparable tothat of high -level coupled -cluster calculations. These two methods,
each having advantages and disadvantages, are complementary “gold -standards”
of quantum chemistry .
There are challenges facing researchers in the field, several of which are
addressed in the chapters in this monograph. These include improving the
accuracy and precision of quantum Monte Carlo calculations; understanding
the exchange nodes and utilizing the simulated electron distribution; extending
the method to lar ge and /or experimentally -challenging systems; and developing
hybrid molecular mechanics /dynamics and Monte Carlo algorithms.
Weextend special thanks tothe Physical Chemistry Division of the American
Chemical Society and Silicon Graphics, Inc. that sponsored the symposium
and provided partial financial support to offset travel expenses of some of its
participants.
Shigenori Tanaka
Kobe, Hyogo, Japan
Stuart M. Rothstein
St. Catharines, Ontario, Canada
William A. Lester ,Jr.
Berkeley ,California, United States
ixDownloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.pr001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Editors’ Biographies
Shigenori Tanaka
Shigenori Tanaka received his B.S. degree in physics from University of
Tokyo in 1982 and his Ph.D. degree in physics from University of Tokyo (S.
Ichimaru) in 1986. After aYoung Research Fellow of Japan Society for the
Promotion of Science and serving two years as aResearch Associate atUniversity
of Tokyo, he joined in 1989 Toshiba Research and Development Center .As
aVisiting Associate, Tanaka also did research at the California Institute of
Technology (R.A. Marcus) from 1995−1996. Since 2004, he is aprofessor at
Kobe University .
Stuart M. Rothstein
Stuart M. Rothstein received his B.S. degree inchemistry from the University
of Illinois in 1964 and his Ph.D. degree in chemistry from the University of
Michigan (S.M. Blinder) in 1968. After aone year post -doc at Johns Hopkins
University (H.J. Silverstone) and serving one year as an Assistant Professor at
Swarthmore College, in 1970 he joined the Department of Chemistry at Brock
University .While on sabbatical leaves, Rothstein has done research at Uppsala
University (P.O. Löwdin), the University of Southern California (H.S. Taylor),
the University of California at Berkeley (W .A. Lester ,Jr.), and Kobe University
(S. Tanaka).
William A. Lester ,Jr.
William A. Lester ,Jr.was awarded aB.S. degree in 1958 and aM.S. degree
in 1959 in chemistry from the University of Chicago. He received his Ph.D. in
chemistry in 1964 from The Catholic University of America while employed at
the National Bureau of Standards. He has held appointments at the Theoretical
Chemistry Institute, University of Wisconsin, the Department of Chemistry ,UW
Madison, and the IBM San Jose Research Laboratory .He was Director of the U.S.
National Resource for Computation in Chemistry and Professor of Chemistry at
the University of California, Berkeley .He ispresently Professor of the Graduate
School, Department of Chemistry ,University of California, Berkeley .
©2012 American Chemical SocietyDownloaded by OHIO STATE UNIV LIBRARIES on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ot001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Chapter 1
Corr elated Sampling for Energy Differ ences in
Diffusion Quantum Monte Carlo
James B. Anderson
*
Department of Chemistry ,Pennsylvania State University ,
University Park, Pennsylvania 16802
*E-mail: [email protected]
Wedescribe amethod using correlated sampling in fixed -node
diffusion QMC calculations todetermine dif ferences inener gies
for similar molecular systems. The method is based on an
earlier scheme for calculating corrections totrial wave functions
applied to two or more systems with correlated calculations in
fixed -node QMC. The result is asignificant reduction in the
statistical error in comparing the ener gies of similar molecular
structures.
Intr oduction
The statistical error inherent in Quantum Monte Carlo calculations leads to
dif ficulties in calculating dif ferences in ener gies for similar systems. Analytic
variational calculations, for example, are free of this problem. In variational QMC
the statistical error in ener gy dif ferences may be reduced with use of correlated
sampling, but correlated sampling in diffusion QMC has been problematical. We
report here anew method which issuccessful for correlated sampling in dif fusion
QMC and produces accurate ener gy dif ferences for similar systems atthe dif fusion
QMC level.
Wehave described previously several versions of aquantum Monte Carlo
method for the direct calculation of corrections to trial wave functions (1,2).This
method gives -for fixed nodes -the dif ference δbetween atrue wave function
Ψand atrial wave function Ψ0,as well as the dif ference between the true ener gy
Eand the expectation value of the ener gy Evarfor the trial wave function. The
statistical or sampling errors associated with the Monte Carlo procedures as well
as any systematic errors occur only in the corrections. Thus very accurate wave
©2012 American Chemical SocietyDownloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

functions and ener gies may be corrected with very simple calculations. The nodes
are unchanged and the wave functions and ener gies for these systems are corrected
to the fixed -node values -those corresponding to the exact solutions for the fixed
nodes of the trial wave functions.
Wehave also described previously some correlated dif fusion QMC
calculations for similar structures of the H3
+ion (3).These took advantage of
scaled length and ener gy parameters, but were only mildly successful because
the initial correlation of electron positions was diminished as the calculations
proceeded.
In the work reported here we have combined the latest version of the correction
method (2)with the correlated sampling indif fusion QMC (3).Wehave carried out
corr ection calculations for cases of two similar systems with correlated sampling
of electron positions, moves, and multiplications wherever possible. Since the two
systems are slightly dif ferent, the initial similarity inwalker positions and weights
is not maintained exactly ,but the correction calculation is fairly short, and the
systems remain nearly similar over the length of the calculation.
Theor etical Basis
The derivation of the latest version of our correction scheme isgiven in Ref.
(2).Wegive abrief description here. The basic equation isthat for QMC with
importance sampling (4–7)where f=ΨΨ 0,the product of the true wave function
and atrial wave function, and Erefisareference ener gy,
Making use of the dif ference δ=Ψ–Ψ0and defining another function g=
δΨ 0=(Ψ –Ψ0)Ψ0we obtain
The term HΨ0/Ψ0isthe local ener gy Elocfor the trial wave function. The last
term in Eq. (2) isadistributed source term S,which may be positive or negative.
Itisconvenient to introduce the expectation value of the ener gy Evarfor the trial
function and write Sas acollection of terms
where
4Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

and where the subscript pindicates aregion Eloc<Evarand apositive particle feed,
the subscript nindicates aregion of Eloc>Evarand anegative particle feed, and
the subscript qindicates an additional particle feed, normally negative.
Weconsider the integrals of each of the terms in Eq. (2) over all accessible
space. Thus we define the integrals
and Eq. (2) becomes
Since the dif fusion and drift terms merely move particles within the
volume and the drift term prevents their crossing anodal surface, they make no
contribution to changes in Ig ,and they are zero and may be eliminated. The
multiplication term containing (Eloc–Eref)isapplicable to each particle fed to the
system. An average growth factor , , for each type of particle during
its lifetime in the system may be combined with the feed terms. Thus we obtain
Since particles fed at any point in the system tend to the same distribution
with increased time in the system, those of suf ficient age may be cancelled in
equal weights, positive with negative, regardless of their locations. For asteady
state and for acomplete cancellation of positive and negative particles at afixed
age we have ∂Ig/∂τ =0, and the ener gy Eisequal to Erefwhich isgiven by (after
rearrangement)
5Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Todetermine the ener gy one then needs to obtain only the ratios of the
integrals Ip/Iqand In/Iqrather than the individual values. Also needed are the
values of , and as well as Evar.
The ratios of the integrals may be determined by numerical integration,
typically by Metropolis sampling of Ψ0
2with accumulation of average values of
Ip/Iqand In/Iq.Their definition in terms of Evarensures that the absolute values of
these ratios be equal and uncertainty in their relative values iseliminated. Since
an accurate value of Evarisrequired, an analytic integration to determine Evaris
preferred. The values of , and are determined for sample feed particles
obtained in the Metropolis integrations. These are selected with probabilities
proportional to (in addition to Ψ0
2)the absolute values of the quantities [– (Eloc–
Evar)] pfor p,[– (Eloc–Evar)] nfor n,and unity for q.The particles are subjected to
dif fusion, drift, and multiplication (weight increase or decrease) for aperiod (age)
suf ficiently long to produce no further change in their average weights. Positive
particles of type p,fed in regions of Eloc<Evar,give values of >1. Negative
particles of type n,fed in regions of Eloc<Evar,give values of <1. Particles
of type qnormally give avalue very close to unity .
The ener gy Eisgiven by the known value of Evarand acorrection term. When
Ψ0isagood approximation to the true wave function, the correction term issmall
and any error in the correction term iscorrespondingly small. As Ψ0approaches
the true wave function the ratios Ip/Iqand In/Iqapproach zero and the values of ,
and approach unity .
Intr oducing Corr elation
Tocalculate the dif ference indif fusion QMC ener gies for two systems Aand B
the variational ener gies EvarAand EvarBfor the two systems and their dif ference are
determined incorrelated variational QMC calculations using Metropolis sampling.
Electron positions may be identical for the two systems or scaled in one way or
another to maintain similarity .The sampling may be carried out with avalue (or
weight) function Wgiven by WA+WB,the sum of squares of the two trial wave
functions. In this case the ener gy for each system isgiven by the weighted average
of its local ener gies using weights proportional to WAor WB.Ifthe two systems
are nearly the same the statistical error in the dif ference EvarA–EvarBwill be much
smaller than the statistical error inthe individual values EvarAand EvarB.The values
of Ip,In,and Iqas well as sample configurations of type p,n,and qare retained.
The next step isacorrection calculation for each of the trial wave functions
carried out as described in the section above using the correlated type p,n,and q
samples to determine values of , and for each of the systems. For these
correlated dif fusion QMC calculations the initial electron positions correspond but
the initial weights as found in the variational calculations may vary for the two
6Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

systems. The drift terms and the multiplication terms may also vary ,but ifthe
systems are similar the dif ferences will be small. The dif fusion terms and the points
of multiplication are matched for the two systems, so that electron movements and
multiplications are correlated as much as possible. Since the initial configurations
pand qare deviations from typical configurations, their local ener gies also deviate,
and the changes inweights and contributions tothe values of and are lar ge
atthe start of arun. Toavoid time -step error the time -step size must be very small
at the start. As the run proceeds the walkers approach normal behavior and the
values of and approach their asymptotic values.
ASimple Test: Neon vs. Neon (moved slightly)
Totest the method we calculated the dif ference in dif fusion QMC ener gies
for two systems Aand Bwhich are known to have the same ener gy: aneon atom
in each case (non -interacting) with the first located at the origin of coordinates
and the second displaced ashort distance from the origin. The displacement was
0.02 bohr .Each had atrial function consisting of afairly good SCF function and
aSchmidt -Moscowitz (8)Jastrow function centered on its nucleus. The two trial
functions were identical except for positioning. In the variational calculation the
sampling value for walker moves was taken as the sum of the squares of the trial
functions. The results are shown in Table 1.
Table 1.Corr elated DQMC: Ne vs. Ne shifted 0.02 bohr
Ener gies (hartr ees)
Ne
#1 Ne
#2
Evar –128.867 504 –127.867 268
(VQMC) ±0.000 090 ±0.000 086
Ecor –128.924 310 –128.924 195
(DQMC) ±0.001 100 ±0.001 100
Ecordif ference –0.000 115
±0.000 077
Eexact
a –128.9376 –128.9376
aReference (9).
Itmay be seen that the uncertainty in the dif ference for variational ener gies
is 16 microhartrees and that for dif fusion ener gies is 77 microhartrees, while
the corresponding uncertainties in the individual values are about 90 and
1100 microhartrees, respectively .The 77 -microhartree error in the calculated
dif ference in dif fusion ener gies is about the same as the indicated statistical
error in that quantity .This isthe desired result. That is, the known dif ference in
7Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

DQMC ener gies (zero) isreproduced (as 0.000 115 ±0.000 077 hartrees) despite
uncertainties of 0.001 100 hartrees in the individual values.
Comparison of Two H2Molecules
Asecond example, for which the ener gy dif ference isknown nearly exactly ,
isthe case of two ground -state H2moleculas with slightly dif ferent internuclear
distances, the first with R=1.4000 bohr and the other with R=1.401 1bohr .In
this case the wave function is nodeless. The trial function was ahigh quality
single -determinant function with an optimized Jastrow term. The time -step and
distance parameters for the dif fusion calculations were scaled to place electrons
in geometrically similar positions –in effect by placing the nuclei in identical
positions in the scaled systems. The results are shown, together with those
of accurate analytic variational calculations by Cencek and Szalewicz (10 )for
comparison, in Table 2.
The table shows the desired effect: the uncertainty inthe calculated dif ference
in the DQMC ener gies (0.040 microhartrees) ismuch lower than the uncertainties
(4.800 and 5.300 microhartrees) in the DQMC ener gies of the calculations
considered separately .
Table 2.Corr elated DQMC: H2with R=1.4000 bohr vs. H2with R=1.401 1
bohr
Ener gies (hartr ees)
R=1.4000 R=1.401 1
Evar –1.173 434 517 –1.173 433 737
(VQMC) ±0.000 000 130 ±0.000 000 130
Ecor –1.174 470 644 –1.174 470 901
(DQMC) ±0.000 004 800 ±0.000 005 300
Ecordif ference –0.000 000 257
±0.000 000 040
Eexactdif ference
a 0.000 000 217
Eexact
a –1.174 475 714 –1.174 475 931
aReference (10 ).
The table shows the desired effect: the uncertainty inthe calculated dif ference
in the DQMC ener gies (0.040 microhartrees) ismuch lower than the uncertainties
(4.800 and 5.300 microhartrees) in the DQMC ener gies of the calculations
considered separately .
8Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Comparison of Two C10Rings
Athird example, for which the ener gy dif ference isnot known exactly ,isthe
case of two C10rings of slightly dif ferent radii. The ten carbon atoms were placed
evenly on circles, one with aradius of 2.04 bohr ,the other with aradius of 2.08
bohr .The circles were centered on the origin, and the length and time scales of
the second were adjusted to obtain geometric and dynamic similarity .The trial
wave function was asingle determinant function with aJastrow term, the same for
both systems. The effect of geometric scaling was to give dif ferent values for the
wave functions and local ener gies for corresponding electron configurations. The
results are shown in Table 3.
Table 3.Corr elated DQMC: C10cir cle (R=2.04 bohr) vs. C10cir cle (R
=2.08 bohr)
Ener gies (hartr ees)
R=2.04 R=2.08
Evar –379.4307 –379.4319
(VQMC) ±0.0067 ±0.0053
Ecor –379.4485 –379.4500
(DQMC) ±0.0130 ±0.0130
Ecordif ference +0.0015
±0.0013
Again the results show significantly lower statistical errors for the dif ferences
in variational QMC ener gies and the dif ferences in dif fusion QMC ener gies
than for the individual values of these ener gies. The uncertainty in the dif fusion
QMC dif ference is approximately equal to the indicated dif ference as expected
for aminimum in the ener gy.Additional calculations for slightly dif ferent
geometries indicate aminimum at aradius of about 2.08 bohr .Independent
analytical calculations (MP2, with substantially dif ferent trial functions) indicate
aminimum in the range of 2.08 to 2.12 bohr .
The Neon Dimer
The neon dimer represents aspecial challenge to quantum mechanics in
general, and this challenge might be met by correlated dif fusion QMC calculations.
One should expect significant cancellations in calculations for separated neon
atoms (Ne +Ne) vs. the Ne dimer (Ne -Ne) at or near its equilibrium separation
distance of about 5.8 bohr .The trial function used was the product of two single
atom functions (as used for the Ne atom above) with additional cross terms in the
Jastrow function. The results are shown in Table 4.
9Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Table 4.Corr elated DQMC: (Ne +Ne) vs. Ne 2(R=5.8 bohr)
Ener gies (hartr ees)
Ne +Ne Ne 2
Evar –257.538 127 –257.538 261
(VQMC) ±0.000 200 ±0.000 200
Ecor –257.845 542 –257.846 338
(DQMC) ±0.000 500 ±0.000 400
Ecordif ference –0.000 796
±0.000 700
In this case we were successful in demonstrating the effectiveness of
correlated variational QMC and obtained awell depth of 0.000134 +/-0.000009
hartrees, in good agreement with accepted values derived from experiment (11).
But, for the correlated dif fusion QMC the indicated well depth was 0.000796 +/-
0.000700 hartrees. With this lar ge value of the statistical error ,we conclude that
amuch better trial function isrequired for this dif ficult case. Weexpect to pursue
this case further .
Discussion
Itisclear that correlated sampling provides significant improvements in the
determination of dif ferences in dif fusion QMC ener gies of similar molecular
systems. Of course, the accuracy in the dif ferences in calculated ener gies depends
strongly on the degree of similarity of the systems compared. For closely similar
systems the dif ference method can provide (loosely) the derivative in ener gy with
respect to avariable describing the dif ference in the systems.
Wenote that the method of correlated sampling in dif fusion QMC described
here can be applied, in asimilar fashion, to the optimization of atrial wave
functions. In this case, we compute the dif ference in diffusion QMC ener gies for
the same system and two similar wave functions. Since the dif ference in ener gies
depends only on the node locations, the node structure may be optimized, thereby
providing asolution to the so-called ‘sign problem of quantum Monte Carlo’. We
are encouraged by our early investigations.
Acknowledgments
Support by the National Science Foundation (IGER TAward No. 9987589)
and by the Research Computing and Cyber Infrastructure (a unit of ITF) of the
Pennsylvania State University isgratefully acknowledged
10Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Refer ences
1. Anderson ,J.B.;Freihaut ,B. H. J.Comput. Phys. 1979 ,31 ,425 .
2. Anderson ,J.B. J.Chem. Phys. 2000 ,112 ,9699 .
3. Traynor ,C. A.;Anderson ,J.B. Chem. Phys. Lett. 1988 ,147 ,389 .
4. Metropolis ,N.;Ulam ,S. J.Am. Stat. Assoc. 1949 ,44 ,335 .
5. Anderson ,J.B. J.Chem. Phys. 1975 ,63 ,1499 .
6. Grimm ,R. C.;Storer ,R. G. J.Comput. Phys. 1971 ,7,134 .
7. Mentch ,F.;Anderson ,J.B. J.Chem. Phys. 1981 ,74 ,6307 .
8. Schmidt ,K. E.;Moscowitz ,J.W.J.Chem. Phys. 1990 ,93 ,4172 .
9. Chakravorty ,S. J.;Gwaltney ,S. R.;Davidson ,E. R.;Parpia ,F.A.;
Fischer ,C. F.Phys. Rev .A1993 ,47 ,3649 .
10. Cencek ,W.;Szalewicz ,K. Int. J.Quantum Chem. 2008 ,108 ,2191 .
11. Aziz ,R. A.;Slaman ,M. J.Chem. Phys. 1989 ,180 ,187 .
11Downloaded by 89.163.35.42 on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch001 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Chapter 2
Population Contr ol Bias with Applications to
Parallel Diffusion Monte Carlo
Jar on T.Kr ogel
*,1,2and David M. Ceperley
1,2
1Department of Physics, University of Illinois, Urbana -Champaign,
Urbana, Illinois 61801
2National Center for Super computing Applications (NCSA),
University of Illinois, Urbana -Champaign, Urbana, Illinois 61801
*E-mail: jkr [email protected]
The future of scientific computing will be driven by highly
distributed parallel machines with millions of compute nodes.
In order to take advantage of this already arriving wave
of computing capability we must identify and remove the
remaining barriers to parallel scaling in the Dif fusion Monte
Carlo algorithm. Toaddress these scaling issues in asimple
way ,we propose that atime delay be introduced into the
population control feedback. In order to assess this algorithm,
we investigate the behavior of population fluctuations and
the population control bias (which will emer ge into greater
relevance with lar ger physical systems and requirements of
higher accuracy) in amodel system for both the standard and
time delayed DMC algorithms. Wethen condense our findings
into asimple set of recommendations to improve the scaling of
DMC while managing the population control bias.
The Futur eof Parallel DMC
Modern parallel machines are composed of tens of thousands of SMP nodes
which will rapidly increase into the millions in coming years. Keeping pace
with such developments iscrucial to the success of scientific computing. Monte
Carlo methods are particularly well suited to make use of this ever increasing
computing power since they contain low serial dependency .In Markov Chain
Monte Carlo, afunction isevaluated along alar ge chain of randomly generated
©2012 American Chemical SocietyDownloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

particle configurations to obtain its mean. Conceptually ,we think of the chain
as arandom walk, meaning that it is formed by awalker which steps from
configuration to configuration. Typically ,straightforward replication of Monte
Carlo random walks allows near perfect scaling, however acentral feature of
any Monte Carlo algorithm will limit the practical efficiency gains realized when
scaling to millions of processors: the equilibration time of the random walks.
The bottleneck due to equilibration time arises from the pursuit to simulate
ever lar ger systems. Since the time to obtain new configurations is increased,
the equilibration time consumes alar ger fraction of the fixed serial run time. A
direct way tominimize the ratio of serial equilibration time and total run time isto
minimize the number of walkers per SMP node. For many Monte Carlo algorithms
itispossible to achieve the exact minimum of just one walker per node; however
this isnot the case for the standard Dif fusion Monte Carlo (DMC) method.
DMC isaMonte Carlo procedure to solve the many -body time independent
Schrödinger equation (1,2).Itiscurrently the most widely used Quantum Monte
Carlo algorithm for chemical and condensed matter systems. In the standard
algorithm, the population of walkers fluctuates through abirth -death process
known as branching and so load balancing challenges arise in the limit of few
walkers per node. For example, aset of nodes each running with asingle walker
will quickly evolve into asituation where nodes contain zero, one, or two or
more walkers each. In this circumstance, the overall efficiency can fall since all
walkers must complete astep before the trial ener gy iscollected and branching is
performed. In typical implementations, current codes use on the approximately
one hundred walkers per node to avoid load balancing inef ficiency ,however this
leads to equilibration inef ficiency as mentioned above. Though load balancing
operations cannot perfectly redistribute the load for arbitrary population sizes,
they will have to be performed frequently to prevent the situation from becoming
even worse. Since most implementations of load balancing rely on operations
within the Message Passing Interface (which impose synchronization at some
level), parallel efficiency will further suf fer with increasing machine size. At the
same time, the branching process must be tightly regulated to prevent chronic
shortages of walkers which could leave lar ge numbers of nodes idle.
Demanding smaller fluctuations in the DMC walker population (greater
stability) increases asystematic error within DMC as we detail below .However ,
with the advent of lar ger machines, agreater level of accuracy will inevitably be
pursued, requiring parallel algorithms with greater robustness and less bias. Thus
the population control bias, which has lar gely been neglected in the past, could
become increasingly relevant to future applications.
In this work, we explore the relationship between the population control
parameters and the stability and accuracy of the standard DMC method. Using
asimple and efficient model problem, the simple harmonic oscillator ,we obtain
empirical formulas describing the behavior of population stability and the
population control bias over arange of control parameters. Since we expect
population dynamics (and hence stability and bias) to depend more strongly on
the control method used than the particular details of the physical system, these
results should apply broadly to current DMC practice.
14Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Looking ahead to the rapidly approaching parallel efficiency dif ficulties
outlined above, we move on to apossible solution to the scaling problem. We
propose the introduction of afeedback delay in the population control mechanism
which removes synchronization requirements from the algorithm and allows load
balancing techniques that capitalize on keeping the time averaged load constant
rather than the instantaneous load. These features ease the tradeof fbetween serial
and parallel efficiency .As we must be cautious not to introduce unmanageable
instability or bias into the simulation, we repeat the model analysis on cases with
delayed feedback. These results are then consolidated into abrief set of practical
recommendations to improve the parallel efficiency and reduce the population
control bias in typical DMC simulations.
Other types of Quantum Monte Carlo, such as Reptation (3)and Path Integral
Monte Carlo (4),do not suf fer from population control bias. However ,those
methods have not been systematically applied to systems of many electrons, and
in particular ,there have been no studies of their relative efficiency ,e.g. how
their errors scale with computer time and number of electrons. In addition, these
methods have other dif ficulties in scaling to very lar ge number of processors. For
these reasons, we limit the discussion to Dif fusion Monte Carlo.
Review of the DMC Method
The Dif fusion Monte Carlo algorithm (1,2)can be viewed as atechnique to
iteratively refine the standard variational method (5).In the variational method,
an ansatz for the many -body wavefunction (ψT)is optimized to be as close as
possible tothe true ground state, usually by minimizing acombination of its ener gy
moments, .The Hamiltonian operator ,Ĥ, isgiven by
The variational ener gy obtained by integrating the local ener gy (EL=ψT
-1Ĥψ T)
over the trial probability density (f0=|ψ T|
2)isastrict upper bound on the ground
state ener gy of the system.
In DMC, the variational probability density isevolved forward in imaginary
time through the iterative application of ashort time Green’ sfunction.
15Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

The importance sampled (1,2,6,7)form of this propagator
optimally guides the trial probability density toward the mixed distribution (ψTψ0)
which is composed of the trial and exact ground state wavefunctions. In fact,
the probability density evolves toward the mixed distribution exponentially fast in
imaginary time.
The exponential factor on the right hand side of equation 5provides the first
glimpse of an instability in the DMC algorithm which will be addressed in the
context of population control. The trial ener gy,ET,is selected as close to the
ground state ener gy (E0)as possible to prevent the norm of the distribution from
vanishing or diver ging within the timescales of interest. The rapid evolution of the
probability density similarly means that the DMC ener gy
will also experience exponential conver gence from the variational ener gy atzero
time to the exact ground state ener gy atlar ge imaginary times.
Thus far ,we have considered only the formal outline of the DMC algorithm.
The conversion of the formal method into aMonte Carlo sampling process
introduces the practical issues of efficiency ,stability ,and accuracy .In practice,
the short time Green’ sfunction is approximately factorized into aproduct of
dif fusion and branching terms.
Here the magnitude of the timestep, τ,isdetermined by the acceptable level
of error in the breakup; later on τwill take on ameaning which dif fers from this
standard usage. The dif fusion term determines the spatial motion of the random
walks, causing them to under go Brownian motion which isbiased toward regions
of high probability as represented by the trial wavefunction. The branching term
acts as aconfigurational weight, which istypically implemented as abirth -death
process of the walkers.
The walker population atfinite imaginary time isrepresented by the norm of
the mixed distribution:
16Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Recalling the lar ge time limit of the probability density inequation 5,itisclear
that the walker population will perish or diver ge, depending on the value of the trial
ener gy.Yet, even ifthe trial ener gy isexact, local fluctuations in the branching
weights (eq 7)will cause the dynamic population to vary in an unbounded fashion
as imaginary time progresses. This is evident from the fact that the integral in
equation 8isapproximated by asum over all Monte Carlo trajectories (prior walker
populations) which involve products of the branching weights up totime t-τ.Thus
the future weight of asingle walker ,represented by the number of its descendants,
will exponentially disappear or diver ge depending on the local ener gy along its
future trajectory .The collective effect of unhindered branching isan instability in
the walker population, which isclearly intolerable from apractical perspective.
Several procedures have been invented to restrict the fluctuations of
the walker population at the expense of introducing additional bias into the
calculation. Such procedures generally fall into one of two camps: fixed
population methods, such as the “comb” (8,9)and reconfiguration algorithms
(10 ,11),or the dynamic adjustment of the trial ener gy (1,2,12 ,13 ).In this work,
our attention will be restricted to the latter ,as itisthe most widely used method
for population control in the DMC community .In the standard population control
procedure, the trial ener gy ismodified in the following way:
Here, the average indicates the current best estimate of the total ener gy and Pt
and P0are the current and tar get populations, respectively .
The calculation of the total ener gy and population iswhat binds together the
otherwise independent walkers, and causes aparallel bottleneck. The control
parameter ,g,issomewhat arbitrary and equals the typical number of generations
(or branching steps) required for the population to return to its tar get value. In the
absence of fluctuations, the population will return to its tar get value in exactly
gτ units of imaginary time. These population control adjustments perturb the
branching weights and hence introduce bias into any estimators. If the control
parameter istoo lar ge, the population fluctuations will be hard to control; ifitis
too small, there could be alar ge bias on the estimated ener gy.
The population control bias has been addressed in only ahandful of studies
to date, possibly because its magnitude has been small for the typical range of
physical system sizes (14).Though ithas been known for some time that this bias
ultimately scales exponentially with physical system size (13 ,15 ),itseems that
this fact isonly becoming appreciated more recently (16 ,17 ).Ithas been shown
that the bias will decrease linearly with tar get population size (12 ,13 ,15 ,18 )
and will decrease ifthe uncontrolled population fluctuations are small (10 ).The
population control bias can be systematically reduced by updating the trial ener gy
less often (12),increasing the tar get population size, loosening the dynamic control
(13 )(increasing g),or using importance sampling (12 ,13 ,18 )with abetter trial
wavefunction. Attempts to completely correct for the bias include extrapolation
in 1/P0(12 ,13)which requires many runs, and reweighting samples over apartial
17Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

history (13 ,15)in asingle run, though this isnot always more efficient than the
extrapolation approach (19).
Delayed Feedback DMC
The pursuit of higher accuracy is one of the main reasons lar ger machines
are developed for computational science. In the context of Quantum Monte Carlo
this means that continual effort must be given to minimize statistical error and
systematic biases. In order to meet these goals, we wish to reformulate Dif fusion
Monte Carlo inasimple way that removes the bottleneck toparallel efficiency and
does not exacerbate the population control bias. Any modifications to the DMC
algorithm should also maintain the stability of the algorithm, i.e. the population
fluctuations must be controllable. Finally ,we require that reproducibility be
maintained, specifically that simulation results should not depend on machine
conditions such as the number of nodes used or interruptions from background
daemon processes.
Asimple way of meeting these requirements isto introduce atime delay ,or
lag, in the population control feedback. Thus the trial ener gy used at time tonly
requires information from all walkers attime t-L.The opening of this window in
“time” postpones any synchronization demands, and may remove them entirely
provided that aload balancing can be done within this time window .This opens
the possibility of near perfect parallel scaling even for avery small number of
walkers per node. Introducing afeedback delay clearly satisfies the requirement
of reproducibility since information from dif ferent imaginary times isnot mixed.
Its impact on stability and bias isnot as obvious and requires further consideration
and testing.
In general, the introduction of afeedback delay should decrease the stability
of the simulation, meaning that fluctuations of the walker population should grow
lar ger .Crudely speaking, the effect of the time delay isto increase the population
control parameter (g)by the length of the delay (L),since itwill take Lmore steps to
reign inpopulation fluctuations than itwould otherwise. This loosening of control
will allow lar ger fluctuations, increasing the probability that the population will
explode, vanish, or simply become too volatile for effective load balancing.
The impact of afeedback delay on the population control bias issomewhat
more subtle. Tosee its effect, consider walkers entering aregion of very low local
ener gy.Without afeedback delay ,the population will immediately rise, causing
asharp response from the population control mechanism and hence alar ge bias.
Introducing atime delay will allow walkers to dif fuse back out of this biasing
region before branching, thus smoothing out its effect and reducing the bias. The
effect may or may not be lar ge, but at the very least the introduction of delayed
feedback should not increase the population control bias.
18Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Dependence of Stability and Accuracy on DMC Parameters
The impact of delayed population feedback on stability and accuracy needs
to be studied and quantified, with attention given to its interrelationship with
other parameters used in DMC. Relevant parameters include the tar get population
size, population control parameter ,branching frequency ,and trial ener gy update
frequency .In order to probe these relationships empirically ,we have developed
amodel that implements DMC for one of the simplest possible systems, the
one -dimensional harmonic oscillator .This model isjustified because the bias and
population fluctuations should depend more strongly on the control method used
than the actual details of the physical system.
Model System
The simple harmonic oscillator has the advantage that its imaginary time
propagator isknown exactly .This fact can be exploited to eliminate the timestep
error typically incurred by the breakup into dif fusion and branching terms (see
eq 7). With the timestep error removed, the only bias remaining is due to the
population control procedure, allowing us to study itin isolation and hence much
more accurately .In order to mimic DMC, the importance sampled propagator is
again factored into dif fusion (Gd)and branching (Gb)terms:
In the limit of small imaginary time, the results in equation 7are recovered.
Though the resulting DMC algorithm has no timestep error ,the timestep used
here still has an important connection to typical DMC simulation. In standard
DMC, there are three timescales which are relevant tothe discussion of population
control bias. The first is the Trotter timestep, τt,which appears in the standard
breakup given in eq 7.This timescale relates to the timestep error and ithas been
purposefully removed from our model. The other timescales are the branching
period, τb,which is the time between branching operations, and τu,the time
between updates of the trial ener gy (ET).Branching more often without adjusting
the trial ener gy should have no adverse effect on the bias, and so we set τb=τuin
our simulations. For the remainder of the work, we refer to this timescale as the
timestep or the update period and will write itas τ. The bias will increase with
more aggressive population control, and thus we expect to see alar ger population
control bias atsmall timesteps in our results.
The potential and trial wavefunction used for the model system are given by
with all quantities listed in dimensionless units (ħ=1,m=1/2). The exact ground
state ener gy for this system is equal to 1and the ground state wavefunction is
obtained when a=1. Since apoor trial wavefunction will increase the branching
19Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

rate, and thus more strongly manifest the population control bias, we use atrial
function with a=0.1 which badly oversamples low probability regions.
The exact importance sampled propagator ,or Green’ sfunction, isgiven by
Where cand sare defined as
The Green’ sfunction is then integrated, as in equation 10 ,to obtain the
effective branching weight and dif fusion transition probability .
The transition probability can be sampled exactly using aGaussian random
number generator .
Definitions of Stability and Bias
Let us define exactly what we mean by stability and bias in DMC. Defining
bias isstraightforward: itissimply the deviation away from the exact ground state
ener gy.
Stability ismost easily quantified in terms of its opposite: the volatility of the
walker population. Here we define volatility as
which isthe magnitude of population fluctuations.
Perhaps amore direct measure of stability isthe probability that asimulation
will not fail, meaning that the population remains within apractical range (we
have used ).Given the inexpensive nature of our calculations,
we also explore this quantity by repeating runs many times in volatile areas of
parameter space (which isspanned by g,P0,and τ). These quantities are linked
in the sense that isosurfaces of the volatility closely match those of the stability .
The reason for this close relationship isthat both quantities are determined by the
width of the population distribution. Thus conclusions drawn from the volatility
are valid reflections of how viable, or stable, DMC simulations will be.
20Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Results with Zer oFeedback Delay
Aseries of long simulations (3.34 x10
8samples) were performed over arange
of control parameters (g=1,2,4,8,16, 32, 64, 128), population sizes (P0=8,16,
32, 64), and update periods (τ=0.0025, 0.005, 0.01, 0.02, 0.04). These population
tar gets are very small, but were chosen so that the volatility and bias would be
increased and thus computed more accurately .
The relationship between volatility (ν), control parameter (g), tar get
population size (P0),and update period (τ) can be inferred from figure 1.The data
for various population sizes (dif ferentiated by symbol) lar gely lie on top of one
another ,demonstrating the inverse relationship between volatility and population
size. Simulations with P0=8were lar gely unstable and have been excluded
from the fits (though the 1/P0relationship still clearly held for these runs). The
data also support apower law relationship between the volatility and gτ. Slight
deviations from power law behavior were found for g<4(data not shown). As
gistaken below the critical threshold of 1, the population feedback overcorrects,
causing the population to oscillate within an exponentially growing envelope.
Thus, gcan be regarded as an independent parameter only as this instability is
approached (small g).Tosummarize, we find that:
Figur e1.Population volatility vs. combined population contr ol parameter and
update period.
21Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Asimilar relationship can be obtained for the population control bias. Figure 2
displays results for the bias over the same range of parameters in Figure 1.Again,
itisimmediately clear that the bias scales inversely with tar get population size.
The bias isclearly not apower law ingτ,but alogarithmic fit conforms tothe data
almost perfectly .The bias then approximately satisfies the empirical relation
where tfisaconstant roughly equal to3.0. Itshould be noted that this form cannot
hold in the limit of lar ge gτbecause the bias must strictly vanish in this regime. A
simple analysis of DMC using continuous weights rather than branching reveals
that the bias must transition to an inverse relationship in the lar ge gτlimit.
Figur e2.Population contr ol bias vs. combined population contr ol parameter
and update period.
Alhough we have studied the bias in asimple model problem, our results
should hold for other systems of typical interest involving singular potentials or
the fixed node constraint. These complicating factors will indeed affect the bias,
since they change the local ener gy distribution. However ,the central limit theorem
implies that these local ener gies, ifaveraged over enough steps, will conver ge to
the same distribution as the model case, but with adif ferent variance. Therefore
the effects of nodes and singular potentials should be contained only within the fit
22Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

constants in eq 18 ,and not in its functional form. Though this argument isfairly
general, explicit tests of such systems should be made to fully confirm its validity .
Results with Finite Feedback Delay
For the introduction afinite feedback delay (L),we expect that most of the
prior results will simply be modified by replacing gwith g+L .This expectation is
lar gely borne out by the data, especially inthe case of the volatility .Figure 3shows
filled stability contours of a64 walker simulation with unstable regions becoming
increasingly dark. Though the stability (probability of asuccessful run) isshown
here, the essential features are identical to the inverse volatility .Adramatic effect
of the feedback delay isseen inthe region of tight control (small g)and finite delay .
The strong instability seen in this region is amanifestation the critical
instability discussed earlier ,only here the system becomes unstable for all g<
L+1. In this regime, population fluctuations are overcorrected, similar to the
feedback experienced when amicrophone isplaced next to aloudspeaker .The
upper solid line inthe plot marks the expected isostability curve geff=g+Lshown
along the boundary between stability and instability .For systems with smaller
tar get populations, the stable (white) region becomes narrower and the predicted
contour matches well with the observed stability until the critical instability is
approached. Thus equation 17 is extended to obtain eq 20 as long as the time
delay satisfies eq 21 .
Provided Lischosen somewhat below this bound, the time delayed method
should be as stable as the standard algorithm with acontrol parameter of g+L.
Though the triangular stability region narrows as Lincreases, itshould be noted
that its size will increase proportional to the tar get population size.
Results for the population control bias demonstrate that afeedback delay does
reduce the bias, though the beneficial effects are partially checked by an increase
in volatility .Figure 4illustrates this situation by showing the real decrease in the
bias after the increased volatility isaccounted for .Time delayed simulations were
performed with afixed control parameter (g=32) over arange of delays (from 0to
32). The time delayed data are plotted atan effective control parameter (g+L),so
that data displayed ataparticular value of ghave the same volatility .Itappears at
first that adding atime delay equal tothe control parameter halves the bias relative
to the undelayed case. However ,after the data are shifted to reflect the loss of
stability ,the actual gains are smaller .Still, increasing the feedback delay can be
an effective means for reducing the population control bias provided the increased
volatility isacceptable.
23Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Figur e3.Stability vs. feedback delay and population contr ol parameter shown
with instability lines for P0=64 and τ=0.01. Brighter regions aremor estable.
Figur e4.Population contr ol bias vs. effective contr ol parameter for standar d
(black) and time delayed (red) DMC at τ=0.01 and P0=32.
24Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Practical Recommendations for DMC Simulations
Optimizing the balance between efficiency and accuracy isasignificant goal
of any discipline which employs lar ge scale simulation. Finding the right tradeof fs
in DMC isacase of selecting optimal values for the control parameters. Using
the insight gained from the model system, we are now in aposition to effectively
navigate the parameter space toward the optimal solution.
Anatural benefit of increasing machine size is that the tar get walker
population must scale in the same way ,providing an added reservoir of stability
to the simulation and damping the bias. The increase in the size of the simulated
physical system requires adecrease of the number of walkers per node in order
to restore the serial efficiency lost from longer equilibration times. The low
walker count per node increases the total idle time due to more interactions with
synchronization barriers and strains load balancing algorithms, further reducing
the parallel efficiency .
Introducing apopulation control feedback delay of suf ficient length should
push back the synchronization barriers inherent instandard DMC. Though the time
delay will increase the volatility of the walker population, and thus require frequent
attention from the load balancer ,there will be more time to get the balance right
before the ener gy must be accumulated. The load should be defined in this case as
the amount of time required toadvance all walkers toaspecific point inimaginary
time, such as Lsteps beyond the average projection time of the walkers.
The population control parameter (g)should be selected so as to satisfy the
lower bound in equation 21 while keeping itas small as isnecessary to reign in
the population fluctuations. Ifthe population control bias isaprimary concern,
the population correction in the trial ener gy could be updated less often and /or
standard correction methods such extrapolation (12 )or reweighting (13 )could be
employed.
Conclusion
Wehave established empirical relationships which reveal the dependence
of the population control bias and population volatility on the trial ener gy
update period, population control parameter ,and tar get population size. Though
established by athorough investigation of asimple model system, we expect
that these relationships will apply generally to Dif fusion Monte Carlo. Wehave
proposed asimple modification to the DMC algorithm, namely the use of delayed
feedback in the population control mechanism, which should increase the parallel
performance of DMC simulations. This modification, which requires minimal
change to existing codes, should function well with only afew walkers per node
which will reduce the fraction of CPU time spent on walker equilibration. Model
results have shown that the introduction of afeedback delay moderately reduces
the population control bias after the increase in population volatility has been
accounted for .Finally ,we have provided general recommendations to improve
serial and parallel efficiency while reducing bias by adjusting only the trial ener gy
update frequency ,population control parameter ,and feedback delay .Future work
25Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

will investigate the population control bias in real physical systems to assess the
adequacy of conclusions drawn from the model system.
Acknowledgments
Wegratefully acknowledge support from the U.S. Department of Ener gy
under contract No. DE -F605 -080R2335 and the National Science Foundation
under grant OCI -0904572. Helpful comments and suggestions made by Jeongnim
Kim and Bryan Clark have aided our work and we express our gratitude for their
insights and encouragement.
Refer ences
1. Ceperley ,D. M. ;Alder ,B. J.Phys. Rev .Lett. 1980 ,45 ,566 .
2. Reynolds ,P.J.;Ceperley ,D. M. ;Alder ,B. J.;Lester ,W.A. J.Chem. Phys.
1982 ,77 ,5593 –5603 .
3. Baroni ,S.;Moroni ,S. Phys. Rev .Lett. 1999 ,82 ,4745 –4748 .
4. Ceperley ,D. M. Rev .Mod. Phys. 1995 ,67 ,279 –355 .
5. McMillan ,W.L. Phys. Rev .1965 ,138 ,A442 .
6. Kalos ,M. H.;Levesque ,D.;Verlet ,L. Phys. Rev .A1974 ,9,2178 .
7. Mentch ,F.;Anderson ,J.B. J.Chem. Phys. 1981 ,74 ,6307 –631 1.
8. Davis, D. H. Lawrence Radiation Laboratory Report No. UCRL -6707, 1961.
9. Booth ,T.E.;Gubernatis ,J.E. Phys. Rev .E. 2009 ,80 ,046704 .
10. Buonaura ,M. C.;Sorella ,S. Phys. Rev .B1998 ,57 ,11446 –11456 .
11. Assaraf ,R.;Caf farel ,M. ;Khelif ,A. Phys. Rev .E2000 ,61 ,4566 .
12. Ceperley ,D. M. J.Comput. Phys. 1983 ,51 ,404 –422 .
13. Umrigar ,C. J.;Nightingale ,M. P.;Runge ,K. J.J.Chem. Phys. 1993 ,99 ,
2865 –2890 .
14. Drummond ,N. D.;Needs ,R. J.;Sorouri ,A.;Foulkes ,W.M. C. Phys. Rev .
B2008 ,78,125106 .
15. Hetherington ,J.H. Phys. Rev .A1984 ,30 ,2713 .
16. Warren ,G. L.;Hinde ,R. J.Phys. Rev .E2006 ,73 ,056706 .
17. Nemec ,N. Phys. Rev .B2010 ,81 ,0351 19 .
18. Cerf ,N.;Martin ,O. C. Phys. Rev .E1995 ,51 ,3679 .
19. Drummond ,N. D.;Radnai ,Z.;Trail ,J.R.;Towler ,M. D.;Needs ,R. J.Phys.
Rev .B2004 ,69,0851 16 .
26Downloaded by STANFORD UNIV GREEN LIBR on May 27, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch002 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Chapter 3
Enhancement of Sampling Efficiency in
ab Initio Monte Carlo Simulations Using an
Auxiliary Potential Energy Surface
Akira Nakayama
*and Tetsuya Taketsugu
Division of Chemistry ,Graduate School of Science, Hokkaido University ,
Sappor o060 -0810, Japan
*E-mail: akira [email protected]
An approach is developed to enhance sampling for ab
initio Monte Carlo and ab initio path integral Monte Carlo
calculations of molecular clusters by utilizing an approximate
potential as aguide to move in the configuration space more
efficiently .Two methods are introduced in this chapter ,where
the first one utilizes the interpolated potential ener gy surface
obtained by the moving least -squares method, and the second
one employs the molecular dynamics scheme to update the
system configuration in the context of hybrid Monte Carlo
method in which potential ener gy gradients are evaluated by
computationally less expensive ab initio electronic structure
methods. The sampling efficiencies in both methods are
demonstrated for awater molecule and hydronium cation.
I.Intr oduction
Computer simulations via molecular dynamics (MD) and Monte Carlo (MC)
methods are widely used inavariety of fields ranging from drug design tomaterials
science (1).Instead of using empirical force fields, ab initio molecular dynamics
or ab initio Monte Carlo methods, which solve the electronic Schrödinger equation
for nuclear potential ener gy (or its derivatives) as needed, are now being widely
used to investigate static and also dynamic properties of molecular systems from
first principles. The computational cost for such ab initio simulations is, however ,
still highly expensive when combined with statistical simulations since 10
5-10
6
©2012 American Chemical SocietyDownloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

simulation steps are usually required, thus limiting their applicability to relatively
small system sizes.
AMonte Carlo simulation is efficient in empirical force fields, especially
when the total potential isrepresented by asum of the additive interactions. In
each MC step, the pair interactions only between the ‘chosen’ atom and other
atoms are needed to evaluate, i.e .itisnot necessary to evaluate the potential of
the whole system. If one employs the primitive approach for ab initio Monte
Carlo calculations, ab initio potential ener gy must be calculated in each step even
ifonly one atom ischosen to move, which does not seem to be efficient compared
to simulations with empirical force fields. In contrast to ab initio Monte Carlo
calculations, ab initio molecular dynamics has the advantage that the whole
molecular configuration isupdated in each time step. However ,the evaluation of
the ab initio potential ener gy gradients requires additional computational cost,
which could be several times more than that of ener gy calculation. Furthermore,
the module for the analytical ener gy gradients is not available for some highly
correlated methods such as CCSD(T).
In this chapter we demonstrate that sampling efficiency for ab initio Monte
Carlo calculations isrealized by utilizing an approximate potential as aguide to
move in the configuration space more effectively .An auxiliary Markov chain is
generated on this approximate potential, and as an approximate potential, we adopt
an interpolated potential ener gy surface that isconstructed by the moving least -
squares method (2).Also we introduce another simple scheme in the context of
ab initio hybrid Monte Carlo (HMC) calculations, which also enhances sampling
efficiency by utilizing the potential ener gy gradients of an approximate potential
for updating the system configurations (3).
These schemes are applied to awater molecule and hydronium cation (H 3O
+)
and the effectiveness of the present approach isanalyzed in detail. Since quantum
effects of nuclei are important for hydrogen atoms, we also perform ab initio path
integral Monte Carlo (PIMC) calculations for these systems and the efficiency of
the method isagain discussed.
This chapter is organized as follows: Section II presents details of the
simulation methods including ashort review of the path integral Monte Carlo
method for the quantum treatment of nuclei. In Sec III the results obtained
by these methodologies are presented, along with the detailed analysis of the
efficiency of the method. Section IV contains the conclusions of this work.
II.Methodologies
Appr oximate Potential Method
The multiple Markov chain methods (MMC) (4)are originally developed to
enhance sampling in which Markov chains running at ahigher temperature are
used to promote transition among dif ferent regions of high probability density .
The approximate potential (AP) method exploits this idea and uses an auxiliary
Markov chain that moves on the approximate and computationally less expensive
potential. In this method, aMonte Carlo importance sampling isperformed on this
approximate potential, and after afixed number of steps, the dif ference between the
28Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

original and approximate potential isused as acriterion to restore sampling with
respect tothe original potential. With this method, correlation inthe Markov chain
on the original potential issignificantly reduced. Since the approximate potential
ischosen such that the computational cost for evaluation ismuch less expensive
than the original one, significant computational saving isattained. Iftimie et.al .
utilized this idea in the ab initio simulation of proton transfer reactions and used
molecular mechanics potentials as an approximate potential (5,6).Gelb applied
this scheme to Lennard -Jones fluids and showed that the speedups by almost a
factor of four are obtained (7).Bandyopadhyay showed that this method can be
used to determine stationary points on the potential ener gy surface of molecular
systems (8).
In the AP method, the key to efficient sampling is finding asuitable
approximate potential. Itis desirable that the approximate potential is close to
the original potential within aregion of configuration space accessible atagiven
temperature. This results in ahigh acceptance ratio and ensures an effective
sampling. When the approximate potential isidentical tothe original potential, the
acceptance ratio isalways one. On the other hand, ifthe approximate potential is
appreciably dif ferent from the original one, the Markov chain on the approximate
potential may spend long periods of time in the region that is less significant
for the system. In some cases this could lead to less efficient sampling than the
conventional MC method.
Webriefly describe the AP method for acanonical ensemble of classical
system, whose probability density isgiven as
where isthe reciprocal temperature inunits of the Boltzmann constant.
Here Ris the configuration of the system and Zis the partition function. The
extension to quantum systems using imaginary time path integral formulation,
which will be described below ,is straightforward. The procedure for the AP
method is as follows. Suppose that the current configuration of the system is
R
(i).An auxiliary Markov chain isgenerated from R
(i)with the probability density
proportional to where Vapisthe approximate potential. After a
fixed number of updates on the approximate potential, one attempts to move the
system configuration R
(i)to the final configuration of the auxiliary Markov chain
(R
(j))with probability
One can see that the transition from R
(i)to R
(j)is performed based on the
dif ference between the original and approximate potentials. With this criterion the
microscopic reversibility for the system configuration
29Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

issatisfied (5,7),where Pisthe transition probability .The acceptance probability
given ineq. (2) isclosely related tothat of the parallel tempering method (4,9,10 )
(also known as the replica exchange method (11)).Ifthe above move isaccepted,
another auxiliary Markov chain isgenerated from R
(j),and ifthe move isrejected,
the current configuration of the system goes back to the original configuration
R
(i)and another attempt is made. It is noted that the limiting distribution for
the system does not depend on the form of the approximate potential, provided
that the updating scheme on the approximate potential isergodic. In the current
implementation of ab initio MC, ab initio calculation is performed at the final
configuration of the auxiliary Markov chain on the approximate potential.
The IMLS Appr oach
The approximate potential ener gy surface isconstructed by the interpolation
method using the moving least -squares approach (12 ,13 ).The modified Shepard
interpolation method (14 )is widely used for construction of potential ener gy
surface and itis usually more efficient than the IMLS method. The modified
Shepard interpolation method requires the derivatives normally up to the second
order to obtain areliable potential ener gy.The ab initio calculations of these
derivatives at reference points are computationally expensive for some highly
sophisticated electronic structure methods. The IMLS method, on the other hand,
yields asuf ficiently accurate potential only with the potential ener gy.The IMLS
isrelatively expensive since one needs to solve aweighted -least -squares equation
at each evaluation point, but still much less expensive than ab initio calculations
by several orders of magnitude.
The interpolated potential ener gy at anuclear configuration Zisrepresented
by alinear combination of linearly independent basis functions bi(Z)as
where Misthe total number of basis functions. The coef ficients aiare obtained by
minimizing the following weighted deviations:
Here, Ndis the number of reference data points and Vref(Z
(i))is the ener gy
value atpoint Z
(i)in the dataset. The weight function wi(Z)decays with distance
so that the closer data points have alar ger weight than the more distant
ones. Note that the coef ficients aiare implicitly dependent on Zbecause of the
weight function. In this study ,the coordinates are taken to be the reciprocal
internuclear distances (i.e .,Zk=1/rij),which iscommonly used in the context of
the IMLS interpolations and also in the modified Shepard interpolations.
30Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

The minimization condition with respect toaileads tothe following equation:
where , ,and Wis a
diagonal matrix whose elements are .Biscomposed of values of
the basis functions at Z
(i)and given by
The interpolated potential for agiven configuration Zisobtained by solving
the normal equation Eq. (6) ,which isimplemented by using the singular value
decomposition (SVD).
In this work, we use polynomials up to the second order of each component
of Zas the basis functions. The weight function of the form
isemployed throughout this study .
Path Integral Monte Carlo Method
The imaginary time path integral formulation of quantum statistics provides a
conceptual and computationally practical route for studying the quantum nature of
systems atthermal equilibrium (15).Based on Feynman’ snotion that aquantum
system can be mapped onto aclassical model of interacting “polymers” with path
integrals, the path integral Monte Carlo method has proven to be extremely useful
for studying finite temperature properties of many -particle systems. Quantum
effects such as zero -point motion and tunneling as well as thermal fluctuations are
included rigorously inthis method and itisnow seeing widespread use inavariety
of applications to chemical systems. Acombination of path integral techniques
with ab initio calculations has been realized for systems where the empirical force
fields isnot available.
31Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

In the path integral Monte Carlo calculations, the expectation value of an
operator Âcorresponding toan observable atthermal equilibrium can be expressed
as
where Pisthe number of time slices and isthe canonical partition
function. Here we assume that the system of interest consists of Natoms with the
Cartesian positions .In the above equation, Âisassumed
to be afunction of the position operator and cyclic periodic boundary condition
RP+1=R1is imposed. AMonte Carlo sampling is performed according to the
distribution function given by
The staging PIMC algorithm (16,17 )ishighly efficient as asampling method
in PIMC, thus we employ this technique in aMC sampling on the approximate
potential. When applying the AP method to PIMC simulations, the acceptance
probability for the update of the system configuration from Rto R′isgiven by
Note that the harmonic spring terms cancel out. Therefore, the acceptance
criterion isbased on the dif ference between the ab initio and approximate potential
at the current and trial configurations.
Hybrid Monte Carlo
The hybrid Monte Carlo method combines the advantages of both the MD
and MC methods, allowing the global update of the system configuration with
reasonable acceptance ratio (18 ,19).Itis an exact method and does not suf fer
from the finite step -size errors of MD simulations. In an update process of the
system configuration, the time -reversible and area -preserving MD algorithm needs
to be used to ensure the detailed balance and the commonly used velocity Verlet
algorithm satisfies this requirement. The initial momenta P
(i)are drawn randomly
from the Boltzmann distribution at the given inverse temperature βand after a
fixed number of MD steps, the final configuration R
(j)isaccepted according to the
following criterion
32Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

where P
(j)are momenta atthe final configuration and ρisthe tar get distribution. In
this study ,ρisgiven as exp(− βV ),therefore the acceptance probability isrewritten
as in the right -hand side of eq. (12) .
Here we propose amethod which employs acomputationally inexpensive
(approximate) potential ener gy surface Vapand its gradient for an MD update part,
but uses the same acceptance probability given above to ensure that the tar get
distribution isstill determined under the potential V.Note that P
(j)are the momenta
which evolve under the potential Vapfrom initial phase space variables P
(i)and
R
(i).In the context of ab initio simulations, one can use ener gy gradients of less
expensive ab initio potentials for an MD update part. The numerical evaluation
of the ener gy gradients requires additional computational cost, which could be
several times higher than that of the ener gy calculation, thus there would be agreat
saving in computational effort ifthe above approach isutilized. For example, the
potential ener gies are calculated at the level of CCSD(T) /cc-pVTZ, but ener gy
gradients of HF /cc-pVDZ level are used for an MD update. Itnevertheless gives
the correct distribution determined by CCSD(T) /cc-pVTZ.
III .Results and Discussion
Simulations Using Interpolated Potential Energy Surface
Ab Initio Simulations on H2O
‘On -the -fly’ ab initio Monte Carlo calculations were performed on an
H2Omolecule. The ab initio calculations were performed at the MP2 level of
theory using the MOLPRO2008.1 package (20 ).The segmented DZP basis set
augmented with dif fuse functions (21 )was employed. The reference points for
interpolation were sampled by classical molecular dynamics calculations at a
constant temperature of 300 Kwith the massive Nosé -Hoover thermostat chains.
The time step of 1.0 fswas employed and all configurations and potential ener gies
along the trajectory were stored as the reference data. This time step is not
appropriate from the viewpoint of the ener gy conservation, but itcan explore the
configuration space more quickly .Atotal of 100 reference points were stored.
The ab initio MC calculation with Ns=10 and Np=100 was performed for atotal
of 1,000,000 steps.
Table Ishows the obtained average of the geometrical parameters and
potential ener gies along with these statistical errors and relative efficiencies. The
statistical errors are estimated by block averages. The result of the conventional
MC calculation (without using AP) isalso displayed for comparison. The zero of
the potential ener gy corresponds to the ener gy at the equilibrium structure. The
ab initio PIMC calculations were also carried out for atotal of 100,000 steps.
Reference points were generated by the ab initio molecular dynamics simulation
at T=2700 Kas described above. AMonte Carlo sampling on the approximate
33Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

potential was performed with Ns=96 and Np=100 using the staging algorithm
with path length L=6.
The obtained results are summarized in Table I,along with the result of the
conventional PIMC method (without using AP) employing the staging algorithm
with L=6. Since we used the primitive form for the kinetic ener gy estimator ,the
lar ge statistical error for kinetic ener gy was observed incomparison tothe potential
ener gy.The virial estimator ispreferable for the kinetic ener gy,but since itrequires
potential ener gy derivatives, we have not employed ithere. The acceptance ratio
for the update of the system configuration is around 0.91 with acurrent set of
parameters, which isalso ameasure of the accuracy of the IMLS potential.
In this example on H2O, an increase in efficiency ,which ispropotional to the
inverse of the square of the statistaical error (the relative value with respect to
the conventional method isshown in the table), by roughly an order of magnitude
is obtained for both the classical and quantum simulations. Itis noted that the
above AP simulations take only around 1-2% more in cpu time than that of the
corresponding conventional calculations.
Table I.Average of geometrical parameters and energies along with
these statistical err ors and relative efficiencies for ab initio classical MC
(1,000,000 steps) and ab initio PIMC (100,000 steps) calculations on an H2O
molecule. AP isthe result of calculation with the appr oximate potential. The
equilibrium geometrical parameters areROH=0.9640 Åand θ=104.00
degr ee. The efficiency isrelative to its value without using AP .
classical MC
(AP)
classical MC
(without AP)
PIMC
(AP)
PIMC
(without AP)
<ROH>(Å) 0.9669 − 0.9817 −
<θ>(deg) 103.95 − 103.79 −
<V>(K) 454.0 452.4 3180.1 3170.7
σV(K) 0.6 1.9 3.4 10.5
efficiency 8.9 1.0 9.9 1.0
<K>(K) − − 3960.8 4077.6
σK(K) − − 23.4 73.3
efficiency − − 9.8 1.0
34Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Ab Initio Simulations on H3O
+
The hydronium cation H3O
+has been studied extensively inrelation toEigen’ s
proton transfer mechanism inwater .This cation isafloppy molecule and involves
the Walden inversion between pyramidal structures inthe C3vsymmetry .Previous
calculations by ab initio path integral molecular dynamics showed that the Walden
inversion isenhanced significantly by the inclusion of quantum effects of nuclei
(22 ).
Figur e1.Distribution functions of the distance between (a) the Oatom and
H-H-Hplane and (b) the Oand Hatoms. The solid and dashed lines arethe
results of PIMC and classical MC calculations, respectively .
The ab initio MC and PIMC calculations were performed with Ns=20 and
Ns=96 for atotal of 1,000,000 and 100,000 steps, respectively .Atotal of 500
reference points were generated by the molecular dynamics simulation atT=2700
Kand the nearest Np=100 points were used for the fitting. Figure 1shows the
distribution function of the distance between the Oatom and H-H-Hplane and
between the Oand Hatoms. The increased value atzero for quantum simulation
in Figure 1(a) reflects the fact that the inversion isenhanced significantly by the
quantum treatment of nuclei.
Again, in order to demonstrate the effectiveness of our method, the
conventional MC and staging PIMC calculations were performed for 1,000,000
and 100,000× P/(L-1) steps, respectively .Table II summarizes the obtained
geometrical parameters and ener gies, along with these statistical errors and
relative efficiencies. As seen in the table, an increase in efficiency by an order of
magnitude isobtained for the expectation values of ener gy.The acceptance ratio
for the update of the system configuration was ~0.81 and ~0.82 for classical MC
and PIMC simulations, respectively .Itis noteworthy that simulation with the
acceptance ratio of around 0.8 still gives asignificant reduction in computational
cost. One can improve this rate further by increasing reference points. With
the current set of parameters, the computational cost isincreased by only 1-2%
compared to the conventional calculation.
35Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Table II.Average of geometrical parameters and energies, along with these
statistical err ors (in kelvin) and relative efficiencies for ab initio simulations
on H3O
+.AP isthe result of calculation with the appr oximate potential. The
efficiency isrelative to its value without using AP .
classical MC
(AP)
Classical MC
(without AP)
PIMC
(AP)
PIMC
(without AP)
<ROH>(Å) 0.9845 − 1.0007 −
<θ>(deg) 112.08 − 112.31 −
<V>(K) 928.2 932.9 5194.5 5183.5
σV(K) 1.1 3.8 4.7 13.7
efficiency 12.9 1.0 8.4 1.0
<K>(K) − − 5826.1 6035.8
σK(K) − − 28.4 93.4
efficency − − 10.8 1.0
Hybrid Monte Carlo and Hybrid Path Integral Monte Carlo Simulations
As asimple illustration of the method, we first show the results of ab initio
HMC calculations on an H2Omolecule atT=300 K. The tar get level of ab initio
method is set to MP2 /cc-pVDZ. For simplicity ,asingle step was used for an
MD part in one MC cycle and atotal of 50,000 MC steps were taken. Ab initio
MD calculations at aconstant temperature of 300 Kwere also carried out for
comparison at the level of MP2 /cc-pVDZ.
Figure 2(a) shows the acceptance ratio as afunction of atime step by
employing various auxiliary potentials obtained from dif ferent (low -level) ab
initio methods and basis sets. Itis seen that good acceptance ratio is obtained
even when the time step is increased to 1.5 ~2.0 fs. Figure 2(b) shows the
statistical errors of potential ener gies (σV)and OH bond lengths (σR)as afunction
of acceptance ratio. Agood conver gence isachieved when the acceptance ratio is
around 40 ~60 %. The statistical errors are reduced in most cases in comparison
to ab initio MD simulations with Δt =0.1 fs. Considering that the CPU time
decreases dramatically as alevel of ab initio calculations islowered for computing
gradients and that the efficiency of the method isproportional to the inverse of the
square of statistical errors, asignificant computational saving isachieved. Note
that increasing atime step of ab initio MD calculations will certainly reduce the
statistical errors, but will induce lar ge systematic errors due to finite step -size. In
ab initio MD, the total ener gy deviates from its initial value by ~1.2 kcal /mol for
the entire run when the time step isincreased to 0.5 fs.
36Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Ab initio path integral hybrid Monte Carlo calculations (23)were also
performed for H2Oand H3O
+molecules with 16 imaginary -time slices and
compared with the ab initio path integral MD simulations (24 )(MP2 /cc-pVDZ)
ataconstant temperature of 300 K. The normal mode coordinates (25 )were used
in all calculations and atotal of 50,000 steps were taken for both path integral
HMC and path integral MD calculations. The results obtained are shown in
Table III and the performance isessentially similar to that obtained in ab initio
HMC simulations. Again, ab initio path integral HMC calculations with auxiliary
potentials offered asignificant increase in computational efficiency .
Figur e2.(a) Acceptance ratio as afunction of time step in ab initio HMC
simulation. MP2 /cc-pVDZ //RHF /ST O-3G indicates that the tar get distribution is
calculated at the level of MP2 /cc-pVDZ and the ener gy gradients areobtained by
RHF /ST O-3G for an MD update. (b) Statistical err ors of potential ener gies (filled
symbols) and OH bond lengths (hollow symbols) as afunction of acceptance
ratio. The err ors areestimated by the block averages using 50 blocks in both
the ab initio HMC and MD simulations. The err ors of ab initio MD calculations
areshown at zer o, wher ethe higher (lower) values corr espond to simulations
with Δt =0.2 fs(Δt =0.1 fs). The average values of potential ener gy and OH
bond length taken from the run of minimum err ors are<V> =0.906 kcal /mol and
<R> =0.9678 Å, respectively .
37Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Table III .Acceptance ratio and statistical err ors of potential energies (σ V
in units of kcal /mol) and OH bond lengths (σ Rin units of Å) for H2Oand
H3O
+molecules in ab initio path integral HMC and MD simulations. The
average values of potential energy and OH bond length taken from the run
of minimum err ors are<V> =6.022 kcal /mol and <R> =0.981 1Åfor H2O,
and <V> =10.013 kcal /mol and <R> =1.0019 Åfor H3O
+.
H2O Δt % σV×10
-2 σR×10
-4
MP2 /cc-pVDZ //MP2 /cc-pVDZ
1.0
2.0
94.4
58.4
3.0
1.8
3.2
1.7
MP2 /cc-pVDZ //RHF /cc-pVDZ
1.0
2.0
67.2
38.6
3.8
2.8
3.9
2.5
MP2 /cc-pVDZ //RHF /ST O-3G
1.0
2.0
57.0
29.5
4.3
3.0
4.3
3.2
path integral MD 0.1
0.2


12.5
4.6
5.9
4.5
H3O
+
MP2 /cc-pVDZ //MP2 /cc-pVDZ
1.0
2.0
93.8
55.8
4.7
3.2
3.3
2.6
MP2 /cc-pVDZ //RHF /cc-pVDZ
1.0
2.0
65.5
35.2
6.4
3.7
6.5
2.1
MP2 /cc-pVDZ //RHF /ST O-3G
1.0
2.0
73.1
43.0
5.6
3.6
4.6
2.7
path integral MD 0.1
0.2


15.8
9.5
5.7
4.5
IV .Concluding Remarks
In this chapter ,we demonstrated that the efficiency of ab initio Monte Carlo
and ab initio path integral Monte Carlo calculation is enhanced significantly
using asuitable approximate potential. In the first part, the interpolated potential
ener gy obtained by the IMLS method was used as an approximate potential
and this scheme was applied to awater molecule and small protonated water
clusters. It was observed that the statistical errors were reduced by almost a
factor of three in most of the calculations presented in this work, which translates
into areduction of the computational cost by an order of magnitude. One of
the attractive features of the method is that potential ener gy gradients are not
required in the calculations, which enables one to perform ab initio simulations
with highly correlated electronic structure methods such as CCSD(T), where the
evaluation of ener gy gradients ishighly demanding or the module for analytical
gradients is not available. This will open the door for quantum simulations
with highly accurate ab initio potential. Another feature isthat one can use any
form of the Monte Carlo sampling on the approximate potential. If the system
has the multiple minima separated by high ener gy barriers, itusually exhibits
aquasi -ergodic behavior .This problem can be circumvented by employing the
38Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

advanced sampling method, such as J-walking (26 ),multicanonical sampling
(27 ),and parallel tempering method (4,9,10 ).
In the second part, we introduce asimple scheme in the context of hybrid
Monte Carlo and hybrid path integral Monte Carlo methods, where the ener gy
gradients are obtained from the computationally less expensive method, and
in this chapter ,we employ alower level ab initio method for updating system
configurations. Again, asignificant increase in computational efficiency is
obtained. Itisnoted that the similarity of the tar get and auxiliary potentials isof
course akey factor which determines the efficiency of the method. One needs to
find acompromise between accuracy of an auxiliary potential and computational
cost for evaluating ener gy gradients. One of the notable advantages of the present
method is again that itdoes not require ener gy gradients calculations for tar get
(high -level) ab initio potentials.
Ab initio simulations have the strength in which one can have direct access
to the electronic property of molecules. The detailed understanding of these
properties, such as electronic char ge or dipole moment, will provide deeper
insight into the microscopic structure of complex molecular systems.
Acknowledgments
A.N. acknowledges financial support by aGrant -in-Aid for Young Scientists
from the Ministry of Education, Culture, Sports, Science and Technology (MEXT ,
Japan). The authors thanks Dr .Motoyuki Shiga for fruitful discussions and part
of the present work isacollaboration with him.
Refer ences
1. Frenkel, D.; Smit, B. Understanding Molecular Simulation :Academic Press:
San Diego, 1996.
2. Nakayama ,A.;Seki ,N.;Taketsugu ,T.J.Chem. Phys. 2009 ,130 .
3. Nakayama ,A.;Taketsugu ,T.;Shiga ,M. Chem. Lett. 2009 ,38 .
4. Geyer ,C. J.;Thompson ,E. A. J.Am. Stat. Assoc. 1995 ,90 .
5. Iftimie ,R.;Salahub ,D.;Wei,D. Q.;Schofield ,J.J.Chem. Phys. 2000 ,113 .
6. Iftimie ,R.;Schofield ,J.J.Chem. Phys. 2001 ,114 .
7. Gelb ,L. D. J.Chem. Phys. 2003 ,118 .
8. Bandyopadhyay ,P.J.Chem. Phys. 2005 ,122 .
9. Lyubartsev ,A. P.;Martsinovski ,A. A.;Shevkunov ,S. V.;Vorontsov -
Velyaminov ,P.N. J.Chem. Phys. 1992 ,96 .
10. Marinari ,E.;Parisi ,G. Eur ophys. Lett. 1992 ,19 .
11. Hukushima ,K.;Nemoto ,K. J.Phys. Soc. Jpn. 1996 ,65 .
12. Maisuradze ,G. G.;Thompson ,D. L.;Wagner ,A. F.;Minkof f,M. J.Chem.
Phys. 2003 ,119.
13. Kawano ,A.;Guo ,Y.;Thompson ,D. L.;Wagner ,A. F.;Minkof f,M. J.Chem.
Phys. 2004 ,120 .
14. Ischtwan ,J.;Collins ,M. A. J.Chem. Phys. 1994 ,100 .
39Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

15. Feynman, R. P.Statistical Mechanics ;Addison -Wesley: Redwood City ,CA,
1972.
16. Pollock ,E. L.;Ceperley ,D. M. Phys. Rev .B1984 ,30 .
17. Sprik ,M. ;Klein ,M. L.;Chandler ,D. Phys. Rev .B1985 ,31 .
18. Duane ,S.;Kennedy ,A. D.;Pendleton ,B. J.;Roweth ,D. Phys. Lett. B1987 ,
195 .
19. Mehlig ,B.;Heermann ,D. W.;Forrest ,B. M. Phys. Rev .B1992 ,45 .
20. Werner ,H.-J.; Knowles, P.J.; Lindh, R.; Manby ,F.R.; Schütz, M.; et al.
MOLPRO ,version 2008.1; apackage of ab initio programs.
21. Segmented Gaussian Basis Set. http: //setani.sci.hokudai.ac.jp /sapporo /.
22. Tachikawa ,M. ;Shiga ,M. J.Chem. Phys. 2004 ,121 .
23. Tuckerman ,M. E.;Berne ,B. J.;Martyna ,G. J.;Klein ,M. L. J.Chem. Phys.
1993 ,99.
24. Shiga ,M. ;Tachikawa ,M. ;Miura ,S. J.Chem. Phys. 2001 ,115.
25. Cao ,J.;Martyna ,G. J.J.Chem. Phys. 1996 ,104 .
26. Frantz ,D. D.;Freeman ,D. L.;Doll ,J.D. J.Chem. Phys. 1990 ,93 .
27. Ber g,B. A.;Neuhaus ,T.Phys. Lett. B1991 ,267 .
40Downloaded by UNIV OF GUELPH LIBRARY on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch003 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Chapter 4
Recent Results in the Exact Treatment of
Fermions at Zer oand Finite Temperatur e
Norm M. Tubman,
*,1,2Jonathan L. DuBois,
1and Berni J. Alder
1
1Lawr ence Livermor eNational Laboratory ,Livermor e,CA 94550
2Department of Physics, University of Illinois, Urbana -Champaign,
Champaign, Illinois 61820
*E-mail: [email protected]
Wepresent release -node quantum Monte Carlo simulations of
the first row diatomic molecules and assess how accurately
their ground state ener gies can be obtained with current
computational resources. An explicit analysis of the
fermion -boson ener gy dif ference shows astrong dependence on
the nuclear char ge, Z, which in turn determines the growth of
the variance of the release -node ener gy.Weshow that efficiency
gains from maximum entropy analysis are modest and that
extrapolation to the ground state is tractable only for the low
Zelements. For finite temperatures we discuss what can be
gleaned from the structure of permutation space for interacting
Fermi systems. Wethen demonstrate improved efficiency in
the exact path integral Monte Carlo treatment of liquid
3He by
using importance sampling to deemphasize the contribution of
long permutation cycles to the partition function.
Exact Methods at T=0
The fundamental goal in the field of ab initio simulations is to perform
electronic calculations to high accuracy or,even better ,exactly .Tosimulate exact
methods an exponentially increasing amount of resources seems to be needed,
and thus in practice system sizes are often severely restricted. For example, two
well -known methods which are in principle exact are configuration interaction
and density functional theory .Here, the exponential computational complexity is
manifest in the formulation of configuration interaction (1),and isless obvious in
©2012 American Chemical SocietyDownloaded by STANFORD UNIV GREEN LIBR on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch004 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

DFT where itarises as aproblem in generating exact functionals (2).As aresult,
practical uses of these algorithms do not yield exact results.
Release -node quantum Monte Carlo (RN -QMC) isasimulation method that
allows the eigenstates of aHamiltonian to be sampled without any systematic
bias. The method iscomputationally expensive, however ,and there are significant
limitations to what can feasibly be simulated. For quantum Monte Carlo (QMC)
calculations this has lead to the development of the popular but approximate
fixed -node dif fusion Monte Carlo (FN -DMC). Despite the possibility of
improved accuracy ,RN -QMC has only been used in arelatively limited number
of simulations. Nonetheless there have been some notable applications and
algorithmic developments.
One such application in which RN -QMC has been successful isthe electron
gas (3)where hundreds of electrons have been simulated without conver gence
problems. Conver gence of aRN -QMC calculation is dependent on two decay
parameters given by τ1=E
F
0,E
B
0,which determines the imaginary -time growth
of the variance, and τ2=E
F
1,E
B
0,which determines the slowest imaginary -time
decay of the excited antisymmetric components relative to the fermion ground
state. The ener gies E
F
0,E
B
0,and E
F
1are the fermion ground state, the boson
ground state and the first excited fermion state, respectively .Hamiltonians
in which the fermion -boson ener gy gap is small are well suited for RN -QMC
simulations, and the free electron gas isone such Hamiltonian. Nuclear char ge
centers change the situation significantly such that relatively small atoms have
lar ge fermion -boson ener gy gaps. The lar gest molecular RN -QMC calculations
to date have been performed on systems of around 10 electrons. However ,
even in these cases, conver gence is not always attained. In the first RN -QMC
calculations of molecules (4)several systems were simulated including H2Oand
Li 2molecules. Later RN -QMC calculations included systems such as HF (5),
LiH (6),and H2+H (7)in which RN -QMC was able to reach higher accuracies as
aresult of algorithmic modifications and increased computational resources. Our
goal is to consider the range of Hamiltonians that can be practically simulated
with RN -QMC with current computational power and modern algorithms. In
particular we have applied the method to the simulation of the first row dimers
with an accuracy goal of 10
−3[a.u.].
Released Node Quantum Monte Carlo
Wecan understand how RN -QMC achieves exact results by considering the
eigenfunction expansion of the imaginary time propagator
For ET=E0the asymptotic, t→∞, limit of this operator gives the ground state
of the Hamiltonian. Technically aRN -QMC calculation isonly conver ged in this
limit, however in practice we consider aRN -QMC calculation conver ged when
the slope of the release -node ener gy estimator iszero towithin the statistical error .
This occurs when the standard error of our ener gy estimate at agiven imaginary
42Downloaded by STANFORD UNIV GREEN LIBR on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch004 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

time is lar ger than the final ener gy dif ference to be decayed in the asymptotic
limit. As imaginary time increases with repeated application of the propagator in
equation (1),all excited states decay relative tothe ground state. Ifatrial function,
ΨT(X),isused for importance sampling, the asymptotic distribution will be:
For astandard molecular Hamiltonian the ground state wave function is a
boson state and the fermion ground state isan excited state of the Hamiltonian.
Like all other excited states, itdecays exponentially relative to the ground state
with application of the imaginary -time propagator .In certain cases we can
accurately measure the fermion ground state as it decays in imaginary time.
Todetermine the fermion ground state we can make aprojection onto the
antisymmetric subspace during this decay process.
The antisymmetric projection is problematic for many Hamiltonians. The
release -node method is an example of atransient algorithm since the fermion
ground state is decaying exponentially in the limit of lar ge imaginary time.
The release process is initiated with the introduction of anodeless guide wave
function, ΨG(X),such that the walkers will equilibrate to the boson ground state.
During this process an antisymmetric trial wave function isused to project out the
antisymmetric signal. This causes asign problem that manifests in the simulation
as exponentially growing noise in the release -node ener gy estimate.
The release -node ener gy can be calculated in asimilar form to the FN -DMC
This estimator involves the sum of positive and negative terms, given by the
sign of the trial wave -function as seen in the term R=ΨT(X)/ΨG(X),which isthe
reweighting factor for the guiding wave function. The weight of the ith walker is
calculated as
where the product is over all previous positions the walker has traversed. The
terms EL
Tand EL
Gare the local ener gies of the trial wave -function and the guiding
function respectively .The form of the local ener gy is given by EL
G,T=Ψ
−1
G,T
HΨ G,T.Itcan be shown that the variance of the transient ener gy E0
RNisproportional
to agrowing exponential with imaginary time (8,9):
The release -node estimate for the ener gy will start decaying from the FN -
DMC ener gy,and eventually itwill become flat, while the variance will continue to
grow as given in equation (4) .The conver gence rate from the FN -DMC ener gy to
the fermion ground state ener gy isdif ferent for each of the component eigenstates
43Downloaded by STANFORD UNIV GREEN LIBR on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch004 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

present inthe fixed -node wave function. Ideally one would like afast decay of the
fermion excited states, and aslow increase of the variance, i.e. (E1
F−E0
F)≫(E0
F
−E0
B).In this ideal situation one can hope to sample E0
Fatalar ge imaginary time
with relatively small variance.
Fermion -Boson Gaps
The actual cost of aRN -QMC calculation is somewhat complicated by the
introduction of atrial wave function. Agood trial wave function can significantly
decrease the contamination from the excited states, improving the conver gence of
the release algorithm. In the limit that the trial wave function has the correct nodal
structure, FN -QMC gives the ground state ener gy and RN -QMC will be flat as a
function of imaginary time. Therefore generating high quality RN -QMC results
involves abalance of computing the best wave function possible and running
release -node for as much imaginary time as possible. Once atrial wave function
isoptimized and put into arelease -node calculation itisthe growth of the variance
as afunction of imaginary time that prevents acalculation from conver ging. Itcan
be seen from equation (5) that this isindependent of the trial wave function.
With this behavior of the variance it is important that the excited states
decay away before the simulation isoverwhelmed with noise. For the first -row
dimers, excluding Li 2,we have estimated the amount of time needed to conver ge
our starting FN -DMC starting wave functions
11to an accuracy greater than
10
−3[a.u.] will be greater than 1[a.u.]
−1.Amore accurate estimate of the
conver gence times isdependent on the magnitude of the excited state components
of the FN -DMC wave function. Since the rate of growth of the variance is
determined by the fermion -boson ener gy gap, itisimportant to understand how
aHamiltonian influences this ener gy gap. Werecently made some calculations
of the fermion -boson ener gy dif ferences for the first -row dimers shown in Table
I.The fermion ener gies are calculated from FN -DMC and the boson ener gies
are calculated with unrestricted Dif fusion Monte Carlo (DMC) using nodeless
guiding wave functions. The boson ener gies are measured without any systematic
errors, while the fermion ener gies have systematic errors corresponding to errors
in the nodal surfaces. These systematic errors are estimated to be much smaller
than 1[a.u.] and are much smaller than the scale of the fermion -boson ener gy
dif ferences. The fermion -boson ener gy dif ferences cover arange of about two
orders of magnitude across the first -row dimers. As previously mentioned, we
would like to simulate at least 1[a.u.]
−1of imaginary time during the release
process, however the size of these fermion -boson gaps imply that the variance
will grow to intractable sizes for most of the first row dimers well before 1[a.u.]
−1
of simulation time.
In arecent RN -QMC study of the first row dimers (10 ),for agiven set of trial
wave functions (11),we were able to achieve an accuracy of 10
−4[a.u.] for Li 2,
however we were not able to achieved our desired accuracy of 10
−3for any of the
other first row dimers.
44Downloaded by STANFORD UNIV GREEN LIBR on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch004 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Table I.Fermion -Boson energy gaps for the second row dimers. Boson
ground states aremeasur ed without any systematic err ors with DMC, while
the fermion energies aretaken as the estimated exact values from refer ence
(11).All energies arein Hartr ee.
Impr oving Release Node
While performing RN -QMC on better wave functions will always lead
to better results, there are several methods have been proposed to reduce the
computational cost of RN -QMC –some of which have the promise of removing
the exponential scaling of the algorithm. In this section we consider afew of
the more promising techniques which are imaginary time projections and walker
cancellation.
Projection techniques are based on sampling quantities during the release
process, other than the RN -QMC ener gy estimator ,to project out the ground state
ener gy.Itisknown that some of these quantities, which are called imaginary time
correlators, have simple imaginary -time dependence of the eigenvalue spectrum.
After generating samples of acorrelator of interest, we can fit the data and extract
out the ground -state ener gy.The benefit of such an approach is that ahighly
accurate estimate of the ground -state ener gy might be possible with alimited
amount of imaginary -time data. The idea of fitting imaginary time data is a
general concept that has been applied very broadly in computational physics and
other fields (12 ).As far as applications of this for RN -QMC calculations, only the
LiH molecule and various model Hamiltonians have tested these ideas (6,13 ,14 ).
Projecting out the ground state ener gy with this imaginary time data is
equivalent to performing an inverse Laplace transform. The inverse Laplace
transform isknown to be sensitive to noisy data and we used aMaximum Entropy
technique in our calculations to reduce our sensitivity to noise. Weshowed
significantly better results than our standard release -node calculations (10 ),as
we were able to generate ground state ener gies for Li 2,Be 2,and B2with this
approach. However ,due tothe nature of the inverse Laplace transform, our results
were too noisy to properly project out ground state ener gies for the rest of the first
45Downloaded by STANFORD UNIV GREEN LIBR on May 28, 2012 | http://pubs.acs.org Publication Date (Web): February 6, 2012 | doi: 10.1021/bk-2012-1094.ch004 In Advances in Quantum Monte Carlo; Tanaka, S., et al.; ACS Symposium Series; American Chemical Society: Washington, DC, 2012.

Another Random Document on
Scribd Without Any Related Topics

vasta sanotusta tai muista syistä, ei muistanut kukaan nähneensä
näillä huulilla, jotka ympäröivät hänen pientä suutansa. Hän ei
myöskään puhunut muuta kuin mitä hänen välttämättömästi tarvitsi
puhua. Hän puhui käsivarsiaan ja sormiansa liikutellen. Kun hän
kirjureinsa luona oli, tarkasteli hän heidän laskujaan, ja hänen
terävät silmänsä huomasivat paikalla, jos virheitä löytyi; ja jos
semmoisia löytyi, tuikkasi hän laihan sormensa virheen kohdalle ja
iski silmänsä onnettomaan laskuntekiään. Hän oli itse kassansa
hoitaja, ja kun hän siitä maksoi ostettuja tavaroita tahi alamaisilleen
palkan, punnitsi hän aina joka rahan käsissänsä, ennenkuin hän sen
antoi. Jos joku hänen kirjureistaan tarvitsi ennen määräaikaa rahaa
ja tätä kauppaneuvokselta pyysi, tarkasti hän konttorikalenteria, joka
seinällä riippui, ja seurasi sitä tuolla laihalla sormellaan kulumassa
olevasta päivästä maksun määräpäivään saakka; siinä pysähtyi
sonni, ja se pysähdys oli kauppaneuvoksen vastaus. Aina, kun illalla
kello löi 8, pyyhkäsi hän itse konttorikalenterista kulumassa olevan
päivän pois, sillä kello 8 illalla suljettiin aina hänen konttorinsa. —
Tämmöinen oli kauppaneuvos tavoiltaan, mutta kumminkin pitivät
ne, jotka häntä tunsivat, häntä suuressa kunniassa; sillä hän oli
samallainen itseänsä kuin muitakin kohtaan: jäykkä, vakava ja suora.
»Järjestys!» oli sana, joka kuului useimmiten, kun ei hän voinut
puhumatta olla, ja ankara järjestys vallitsi hänen huoneessaan.
Jouluaatto oli kulumassa. Kauppaneuvoksella itsellä ja kaikilla
hänen alamaisillaan oli tänä päivänä suunnattomasti työtä. Ukko Loo
hyöri ja pyöri paikasta paikkaan, jokaisen töitä tarkastellen, jottei
erehdystä tulisi missään. Ostajia meni ja toisia tuli. Jokaisessa
huoneessa oli vähäiset markkinat. Mutta kumminkaan ei ollut
kauppaneuvos hyvällä tuulella. Jo aamulla oli hän suuttunut, sillä
järjestys hänen talossaan oli rikottu. Järjestykseen tietysti kuuluu,
että jokainen maksaa saadun tavaran, ennenkuin sen mukaansa vie.

Aamupäivällä oli puodissa tuo järjestys rikottu. Kauppaneuvos oli
näet nähnyt, miten eräs eukko oli esiliinansa alle pistänyt
lankakääryn ja oli sitä vieden hiipimässä ulos. Tämän huomattuansa
samosi kauppaneuvos eukon kimppuun, korjasi aika kiireesti
lankansa talteen, talutti sitten eukon ovelle, työnsi hänen ulos ja
antoi hänelle lisäksi aika potkun. »Järjestys, järjestys kaikissa!»
huusi hän silloin. Tästä järjestyksen pidosta oli puodissa syntynyt
kaikuva nauru, mutta se ei kauppaneuvosta harmittanut. Harmissaan
hän oli vain siitä, että järjestys oli rikottu. Mutta tähän harmiin liittyi
toinenkin, josta kohta saamme tilaisuuden kertoa.
Että kauppaneuvos ei kuitenkaan ollut saituri, vaikka sitä
ylläsanotusta voisi päättää, näkyi siitäkin, miten hän perhettään
kohteli. Mitä ikänä hänen rouvansa ja tyttärensä tahtoivat, sen he
kyllä saivat vastustamatta, mutta ainoastaan sillä ehdolla, että he
kauppaneuvokselta sitä pyysivät. Hän, hän itse eikä kukaan muu, sai
antaa, ei kukaan, ei edes hänen oma rouvansakaan saanut ottaa.
Jouluaaton edellisinä päivinä oli sekä rouvalla että tyttärellä ollut
hirmuisen paljon tarpeita. Lista kädessä kulki kauppaneuvos kahden
puotipojan kanssa huoneesta huoneesen, kooten ylikertaan meneviä
tavaroita, ja aina vaan oli milloin rouvalla, milloin tyttärellä uusia
tarpeita. Vakavana täytti kauppaneuvos heidän vaatimuksiaan. Kello
7:stä aamulla puolipäivään asti toimitti hän väsymättä näitä ja muita
töitään. Nyt kumminkin kuultiin hänen entistä useimmin huutavan
syvällä äänellä: »Järjestys!» vaikka hänellä tähän ei näkynyt olevan
mitään erityistä syytä. Aina tuon tuostakin vilkasi hän ulos
akkunasta, ja silloin aina kuultiin tuo: »Järjestys!» Puotipojat tätä
kummaksuivat, sillä ei koskaan ennen ollut kauppaneuvos pitänyt
lukua siitä, mitä ulkona kadulla tapahtui. Vihdoin arvasivat he, että
kauppaneuvos odotti jotakin.

Päivä oli puolessa. Kauppias oli juuri menossa ylikertaan
päivälliselle, jota hänen talossaan aina syötiin kello 1. Järjestystä
noudatettiin siinäkin. Joka ei heti, kun kello löi 1, istunut pöydässä,
hän jäi päivällisettä.
Juuri kun hän oli menossa ruoalle, tuli postipoika tuoden koko
joukon kirjeitä. Kauppaneuvos silmäsi päällekirjoituksia pikaisesti. Ne
olivat näöltään kauppakirjeitä. Mutta niiden seassa oli yksi, joka ei
näyttänyt kauppakirjeeltä. Sen avasi hän heti ja luki sen kiireesti.
»Järjestys! Järjestys!» huusi hän ja mumisi sitten: »Vahinko … ilma
… myrsky … Hän tulee kumminkin tänään!»
Sitten riensi hän kiireesti ylös ja ennätti juuri sijallensa pöytään,
kun kello löi. Että hän nyt oli hyvillänsä, sen näki jokainen, ken
hänen tunsi. Joku, jota hän odotti, oli tulossa. Hän oli jo päättänyt,
että tämä odotettu oli järjestyksen rikkoja, ja siitä oli hän ollut
harmissaan; mutta nyt oli harmi kadonnut. Pöydässä, jossa hänen
kirjurinsakin istuivat, ei puhuttu sanaakaan. Semmoinen oli järjestys.
Pöydästä noustua kirjurit heti palasivat alikertaan. Kauppaneuvos,
kuten hänen tapansa oli, jäi nytkin rouvansa ja tyttärensä luo kello
kolmeen saakka.
Rouva Loo oli samassa määrässä lihava kuin hänen miehensä
kauppaneuvos oli laiha. Se oli jotakin, se! Muuten oli hän varsin
tavallinen, jokapäiväinen, vähän kärtyinen ja kielekäs. Akkunassa
viistopeilinsä edessä istui hän uteliaana, kuten vanhat rouvat ainakin,
silmäillen siihen, nähdäksensä keitä kadulla kulki. Tänään ei hänellä
kumminkaan ollut sanottavasti aikaa tähän. Läähättäen oli hän
työskennellyt joulupuuhissa; ja että joulua oli aikomus viettää
juhlallisesti ja suuresti, sen huomasi jokainen, ken näki ylikerran
huoneet. Siellä oli vielä puolisen aikana kaikki melkein ylösalaisin.

Mutta tähän tämmöiseen epäjärjestykseen ei kauppaneuvos
suuttunut. Hän tiesi, että epäjärjestys aina tällaisissa asioissa käy
järjestyksen edellä.
Kun kirjurit olivat lähteneet, lähestyi ukko rouvaansa ja sanoi:
»Tulee pian!»
»No, taivas olkoon kiitetty! Minä jo ajattelin … minä jo pelkäsin…
Kuule, Lovisa!… Hän tulee pian! Mutta sano, Loo, miksi on hän
viipynyt, miksi ei hän tullut eilen… Eilen olisi hänen pitänyt tulla,
niinhän oli päätetty … mikä esti?»
»Tuli vahinko. Oli paha ilma, myrsky. Ei uskaltanut lähteä… Odotti
toista laivaa».
»Ja tänään tulee hän! Lovisa! Huuda pian piiat ja palvelijat! Sali on
kuntoon asetettava. Hän tulee, Lovisa, kuulitko, hän tulee, tyttö!
Etkö ole iloinen, Lovisa?»
Tyttö, jota lihava rouva iloitsemaan kehoitti, oli nuori, 24 tahi 25
vuotinen. Hän oli pitkä ja siten siinä isänsä kuva; mutta mikään muu
hänessä ei muistuttanut isästä. Hänen kasvoistaan olisi myös
turhaan haettu jotakin, josta olisi häntä äitinsä kaltaiseksi voitu
sanoa. Äidin kasvot olivat säännölliset, vaikka niitä rumensi
liikanainen lihavuus. Lovisan kasvot olivat rumat. Hänen silmänsä
olivat pienet ja aina vetiset; hänen poskensa olivat valkeankeltaiset,
hänen suunsa suuri, ja hampaansa, mitkä aina näkyivät, olivat
mustat, arvattavasti liiallisesta herkkujen syömisestä.
Äitinsä puhetta ja kysymystä kuullessaan punastui Lovisa, jos voi
sitä punastumiseksi sanoa, että hänen valkeankeltaisista poskistaan
valkoinen väri katosi. Hänen pienet, vetiset silmänsä katselivat

maahan, ja ainoa sana, minkä hän sai sanotuksi, oli: »Mamma!»
Siinä sanassa ja äänessä, millä se sanottiin, oli muka olevinaan
naisellista ujoutta ja nuhdetta, mutta oikeastaan niissä ei ollut
mitään.
»Niin, niin; rakas Lovisani!» lörpötteli rouva. »Nyt jätät sinä kohta
vanhat vanhempasi!» Ja rouvan ääni värähteli, kun hän sen sanoi.
Nenäliinaan tarttui hän liikutettuna ja pyyhkäsi silmiänsä. »Älä
unohda meitä! Muista niitä neuvoja, joita isältäsi ja minulta olet
saanut. Ota minut esikuvaksesi kaikissa seikoissa, mihin maailmassa
voit joutua, niin tulet onnelliseksi ja teet miehesi asunnon
paratiisiksi, niinkuin minä olen tehnyt Loon… Eikö se ole totta, Loo?»
»Hm!» mumisi kauppaneuvos… »Järjestys, järjestys kaikissa,
sanon minä!»
»Sinä kuulet sen nyt itse, Lovisa», jatkoi rouva, joka tällä välin oli
vetäytynyt tavalliselle vahtijapaikalleen akkunan viereen. »Sinä
kuulet sen, sinä siunattu lapsi… Mutta siinäpä sinä seisoa töllöttelet
kuin mieletön! Etkö kuule, Lovisa! Hän tulee, ylkäsi tulee!»
»Mutta minä en ole häntä koskaan nähnyt!» muistutti siihen
Lovisa. Tässä lauseessa piti oleman jonkinlaista vastustusta, mutta
sitä ei siinä ollut.
»Hm, hm, järjestys!» lausui kauppaneuvos. »Jos minä kalliita
tavaroita ulkomaalta ostan, olenko minä niitä nähnyt, ennenkuin ne
tänne tulevat ja ovat minun? Jos minä luotan siihen
kauppahuoneeseen, jolta tavarat ostan, olen minä vakuutettu
tavaran hyvyydestä. Järjestys kaikissa! Minä tahdon uskoa, että sinä
luotat minuun ja pidät hyvänä sen yljän, jonka sinulle olen hankkinut
… poh! Vaikka et ole häntä nähnyt… Poh! Järjestys!»

Kauppapojat ja kirjurit olisivat jääneet suu auki, jos olisivat tämän
kauppaneuvoksen puheen kuulleet, sillä kuukauden kuluessa ei ollut
kauppaneuvos niin monta sanaa lausunut, jos ei lukuun oteta tuota
alinomaista muistutusta: »Järjestys!» Rouvakin jäi aluksi
vastaamatta, sillä niin hämmästyi hänkin. Mutta pian tointui hän
taasen.
»Sinä kuulet, Lovisa! Sinä kuulet», huusi hän, »mitä isäsi on
tehnyt. Suutele isän kättä ja kiitä! Oi, Jumala paratkoon!» jatkoi
rouva ja vei nenäliinan taasen silmillensä. »Jumala paratkoon! Ei ole
kaikilla, ei monella, tuskin ainoallakaan niin helliä vanhempia kuin
sinulla! Huomenna! Taivaan isä! Jo huomenna olet sinä, Lovisa,
antanut rakkautta uhkuvan sydämesi pois. Sinun ei enää tarvitse
totella isääsi ja äitiäsi, ainoastaan häntä, jota sinun tulee miehenäsi
kunnioittaa ja…»
»Ppruuuu! Järjestys! Vasta ylihuomenna!» pisti kauppaneuvos
väliin.
»Jaa … niin, ylihuomenna! Jumalan kiitos! Vasta ylihuomenna!»
tulvasivat sanat rouvan suusta. »Ylihuomenna, isäsi syntymä- ja
nimipäivänä, kun tänne kokoontuu koko kaupunki, silloinpa, silloinpa
aukenevat kaikkein silmät, kun isäsi sinut kihlaa nuorelle, kauniille
Hirtsille… Mutta tässä sinä vielä töllistelet, niinkuin et tietäisi, mitä
kihlatun naisen tulee tehdä, joka ylkäänsä odottaa… Pian pukemaan
päällesi paremmat vaatteet, ettei ylkäsi tarvitse sinua hävetä!»
Kun Lovisa tämän käskyn kuuli, lähti hän pian kamarista. Isä ja äiti
jäivät sinne kahden. Rouvan kieli kävi entistä vauhtiansa. Ukko Loo
käveli edestakaisin lattialla. Kuunteliko hän rouvansa lörpötyksiä, on
vaikea sanoa.

Vihdoin seisahtui hän ja katseli kelloaan. Hänen oli aika lähteä
konttoriin. Hän vetäysi ovelle.
»Minä onneton ihminen!» huusi rouva, kun hän tämän huomasi.
— »Löytyykö maailmassa onnettomampaa ihmistä kuin minä olen!
Tässä olen puhunut tyttäremme tulevaisuudesta, ja mieheni ei edes
viitsi kuunnella, mitä puhun. Oi, minä olen onnettomin kaikista, jotka
alttarin edessä ovat miehen valtaan antauneet!» Ja taasen pyyhkäsi
rouva nenäliinalla silmiään.
Mutta samalla pudotti hän nenäliinan ja hypähti ylös. Hän oli
katsellut akkunapeiliin ja siellä nähnyt jotakin outoa.
»No, kaikki taivaan pyhät!» huusi hän. »Tässä me puhumme,
niinkuin ei olisi muuta tehtävää, ja tuossa tulee odotettu! Minä
sanon, Loo, mene pian häntä vastaan, hän jo ajaa pihaan!»
Kauppaneuvos oli jo avannut oven, kun hän tämän kuuli.
Kiireemmästi kuin ennen puikahti hän siitä nyt ulos, ja vasta rapuista
kuului hänen vastauksensa: »Järjestys!»
Mutta ennenkuin seuraamme kauppaneuvosta odotetun kosian
luo, käykäämme tätä katsomassa. Meidän on palaaminen ajassa pari
päivää takaperin tullaksemme tuntemaan häntä sekä niitä seikkoja,
jotka ovat saaneet hänet kosiana esiintymään.
* * * * *
Pari päivää ennen jouluaattoa oli kaupungin kestikievaritaloon
tullut tuntematon vieras. Että hän oli Ruotsista kotoisin, sen huomasi
hänen puheensa murteesta. Hän oli kaunis, nuori mies, silmät
elävät, välkkyvät, »Minä olen kauppias, kuljen kauppa-asioilla ja

tulen täällä viipymään pari päivää», sanoi hän kestikievarin isännälle,
joka kohteliaasti otti häntä vastaan. »Nimeni on Troberg, jos joku
tahtoisi minua tavata».
Kestikievarin isäntä ei tuossa vieraassa nähnyt mitään kummaa.
Vieras tuli ja meni, meni ja tuli näiden kahden päivän kuluessa.
Harvoin oli hän kotona. Mutta missä hän kävi, sitä ei saanut isäntä
tietää. Jouluaaton aamuna kutsutti hän luoksensa isännän. »Onko
teillä ketään, jonka voisi lähettää kruununvouti Trobergin luo? Hän
on vähän sukua minulle, ja minä tahtoisin häntä tavata ennenkuin
lähden», puhui hän. Kestikievari myöntyi tähän, ja vähää ennen
puolta päivää tuli tuo kaivattu vieras.
»No, sen sanon!» huusi vouti, istuessaan kamarissa kahdenkesken
ruotsalaisen kanssa. »Sinä täällä, täällä Suomessa, etkä tule minun
luokseni asumaan! Ja mitä tämä nimen salaaminen tietää, vai soiko
nimi Troberg paremmalle korvissasi kuin oma nimesi Hirts?»
»Anna anteeksi, veli», sanoi vieras nauraen, »etten luoksesi kohta
tullut. Sen sijaan saat minulle tänä iltana, huomenna ja ehkäpä vielä
ylihuomennakin osoittaa vierasvaraisuuttasi. 'Järjestys!' niinkuin
kauppaneuvos Loo sanoo. Miten on, onko silkkihame valmis ja
salaisuuteni tallessa?» kysyi hän yhä nauraen.
»On; mutta mitä varten tämä ilvenäytelmä? Mitä, Herran nimessä,
sinä silkkihameella teet ja mitä varten on se neiti Loolle annettava?
Jos olisit ollut täällä ennen tahi jos olisit nähnyt tuon neitiviirin,
luulisinpä, että sinä kosiana ilmestyisit!»
»Kuka sanoo, etten ole häntä nähnyt?» kysyi viekkaasti nuori
herra.

»Sinä et ole täällä ennen käynyt eikä hänkään Ruotsissa».
»Mutta jos sinulle sanon, että minä ja hän Tapaninpäivänä
vietämme kihlajaisiamme, niin mitä sinulla on siihen sanomista?»
»En tietysti voisi muuta kuin syvään kumartaa ja toivottaa onnea…
Mutta älä sano minua Trobergiksi, jos minä sitä uskon tahi jos tästä
vähintäkään ymmärrän».
»Ei sinun ole suuresti tarvis mitään ymmärtääkään, kunhan vain
ymmärrän minä, jota asia koskee. Olkoon kumminkin järkesi
valaistukseksi sanottu, että minä olen tullut tänne viedäkseni täältä
muassani kaupunkinne kauniimman kaunottaren».
»'Järjestyksen' tyttären!… Kauniimman!» — Ja vouti purskahti
nauruun täyttä kurkkuansa.
»Älä naura! Minäkö tulen kihlauttamaan itseäni ja lähtisin
pitkänokkaisena! Mutta kokkapuheet hiiteen… Kuule, viinuri! Pullo
punssia!» huudahti nuori herra ovea avaten.
Kun oli haluttu makujuoma saatu, täytti herra Hirts lasit, nousi
vakavana ja lausui: »Ei sanaakaan, ennenkuin olen lopettanut, ja
jollet silloin ymmärrä enempää kuin nyt, niin sanonpa sinua
tuhmaksi. Nyt puheen aluksi: neiti Lovisa Loon malja, minun
morsiameni malja, ja se pohjaan!»
Ystävät joivat maljan pohjaan, vouti kumartuen ja vaiti ollen,
mutta uteliaana.
»Ja nyt kuule ja kummastele!» jatkoi vieras. »Sinä muistat, miten
ja milloin me Marstrandin kylpylaitoksessa tutustuimme, miten siellä
ystävyytemme alkoi. Siitä ei ole sen enempää puhumista, paitsi että

minä tämän ystävyyden nojassa uskalsin sinuun turvata tuossa
silkkihame-asiassa. Sinä täytit pyyntöni, ja minä kiitän sinua. Kun
Marstrandista tulin kotiin, oli minun maallinen onneni seppäin
kynsistä juuri päässyt. Isäni ja kauppaneuvos Loo olivat sitä
takoneet, ja minun onneni oli valmiina ahjosta lähtenyt neiti Lovisa
Loon muodossa. Sinä tiedät, että kauppahuoneet Hirts ja Loo ovat jo
kymmeniä vuosia olleet keskinäisissä kaupoissa. Ukkojen välillä oli
ystävyys syntynyt, vaikka mahdotonta on minun ymmärtää, mihin
tämä ystävyys nojautuu, kummallisia, itsepäisiä ja omituisia kun ovat
ukot molemmat. Sinä näit isäni. Hän on kunnon ukko, ja vaikka
kummallinen, niin ei kumminkaan niin kummallinen kuin Loo, jos
oikein häntä käsitin. No niin! Minä tulen Marstrandista, olen tuskin
saanut tervehdykset sanotuksi, ennenkuin minut kutsutaan isäni luo.
Kuule ja hämmästy! 'Albert!' — sanoo isäni — 'sinä olet nyt kohta
siinä ijässä, että sinun sopii mennä naimisiin. Kuule loppuun! Minä
en siedä, että puhettani katkaistaan, ennenkuin loppuun olen
puhunut, sen tiedät', ärjäsi hän, kun minä tahdoin avata suuni. —
'Sinä olet nyt 21 vuotinen ja valmis ja vapaa avioliittoon. Asian näin
ollen olen hakenut ja löytänyt sinulle morsiamen. Sinun ikäisesi on
hän; hyvässä perheessä on hän kasvanut, oppinut järjestystä ja,
siisteyttä, Albert. Sinun morsiamesi on neiti Lovisa Loo Suomesta.
Varoissa vetää kauppahuone Loo vertoja meidän huoneellemme. Kun
ovat häät vietetyt, saat itse valita, tahdotko täällä vai Suomessa
ruveta asumaan. Nyt olen kaikki puhunut. Kauppaneuvos Loo on
täällä. Hän rakastaa järjestystä ja tiedät siis, mitä sinun nyt on
tehtävä'. — 'Tiedän' — vastasin minä. 'Minun tulee pyytää
kauppaneuvokselta morsiamekseni neiti Lovisa Loo, ja sen teen minä
heti. Tulkaa, isä!' — Olisit nähnyt, miten iloiseksi isäni kävi, kun hän
tämän kuuli. Hän, joka ei koskaan näytä, mitä hänen rintakuorensa
sisällä asuu, huusi nyt: 'Ne sanat, Albert, ne tuottavat, sinulle 10,000

riksiä. Käy konttorissa, niin saat ne. Huvittele niillä vuoden loppuun
saakka. Sitten vietät hääsi. Mutta hääpäivänä pistät sinä
hulivilisuuden säkkiin, jonka minä suljen sinetilläni'. — Minä olin juuri
silloin, niinkuin tiedät, suuressa rahapulassa. Tämä Lovisa Loo ja
10,000 riksiä tulivat niinkuin taivaasta. Kiireesti riensin ukko Loon
luo, syleilin häntä ja suutelin hänen pergamenttihuuliansa, ja
paikalla, ennenkuin hän ennätti tokaista järjestystään, olin minä
pyytänyt siveän ja kauniin neiti Lovisan aviovaimokseni,
rakastaakseni häntä myötä- ja vastoinkäymisissä. Nytpä nousi ilo!
Ukko Loo näki järjestyksen vaativan, että hänkin puolestansa syleilisi
ja suutelisi minua. Jolleivät nuo 10,000 riksiä olisi eläneet mielessäni,
olisin purskahtanut nauruun ja nauranut itseni kuoliaaksi, kun tämän
syleilyn ja suutelon sain. Ojentaen pitkiä käsiänsä, joita voi
paremmin ankkuritouviin verrata kuin elävän ihmisen jäseninä pitää,
tuli hän minua kohti marssien suorana, vakavana, kuten kenraali
armeijansa etupäässä, ja kun hän minun saavutti, teki hän … näin ja
näin!»
Vouti oli kummastellen ja hymyillen kuunnellut ystävänsä
kertomusta. »Sinä riivattu!» huusi hän, kun nuori Hirts syleillen ja
suudellen näytti, miten ukko Loo oli häntä syleillyt ja suudellut. »Sinä
riivattu!» Ja hän purskahti nauruun, joka ei tahtonut loppua saada.
»Appeni malja, ja pohjaan sekin!» esitti Hirts. Trobergilla ei ollut
mitään tätä maljaa vastaan. Se juotiin siis, ja nuori Hirts jatkoi
naurettavalla vakavuudella: »Niin olin minä kihlattu, sillä neiti
Lovisan suostumusta ei muka tarvittukaan. Minä kumminkin olin
toista mieltä, mutta olin liian viisas sitä ilmoittamaan. Minä kutsuin
ukko Loota rakkaaksi apekseni, mutta tämä soti järjestystä vastaan;
vaikka tahtoisin panna nuo 10,000 riksiä äyristä menemään, että
tämä järjestyksen rikkominen oli ukolle varsin mieluinen. 'Vasta

häitten jäljestä appi!' sanoi hän. 'Järjestys, järjestys kaikissa!' Sitten
tuotiin sampanjaa. Kolmen miehen siinä juotiin niin, että jo pelkäsin
järjestyksen menevän hiiteen. Mutta eipä niinkään, vaikka meni se
ainakin minulta: sillä kun ukot huomasivat saaneensa tarpeeksi
eivätkä uusia pulloja tuottaneet, kiiruhdin minä konttoriin, sain rahat
ja teurastin kurkkuni ja tulevan aviosäätyni terveydeksi koko joukon
hopeakauloja. Voit arvata, että aikani nyt oli hauska ja iloinen. Päivät
kuluivat, niin etten tahtonut uskoa itseäni, kun huomasin syksyn
tulleeksi. Mutta nyt, kun lähestyi aika, jona minun tuli täyttää
lupaukseni, alkoi väliin vähän arveluttaa. Olisin kirjoittanut sinulle ja
ilmoittanut tämän kummallisen seikan, mutta ukkojen kehoituksen ja
päätöksen mukaan olin pakoitettu olemaan vaiti. Asia oli näet jäävä
meidän kolmen salaisuudeksi aina siihen päivään saakka, jolloin
kihlaus tulisi julkisesti tapahtumaan. Kuten sanoin, päivät kuluivat, ja
pian oli leikistä tuleva tosi. Silloin tuli ukko Loolta kirje, jossa hän
määräsi kihlauksen tapahtuvaksi Tapaninpäivänä. Minun tulisi pari
päivää ennen joulua saapua hänen luoksensa. Kun isäni tämän
kirjeen sai, kehoitti hän minua antamaan jonkun kalliin joululahjan
neiti Loolle. Ostin valkoista silkkiä, parasta mitä sain, ja kuu en ollut
ikinä morsiantani nähnyt, enkä siis osannut toimittaa sitä sopivasti
ommelluksi, lähetin pakan sinun luoksesi, kuten muistat, pyytäin,
että jollakin lailla toimittaisit hameen valmiiksi ja morsiamelleni
sopivaksi. Minä olin valinnut juuri tällaisen lahjan hyväilläkseni
ukkovaariani. Se onnistuikin, vaikkei enää rahoja lähtenyt. Ukko kiitti
minua, kun ilmoitin tahtovani kihlauspäivänäni nähdä morsiameni
siinä puvussa, minkä rakkauteni oli hänelle valinnut. — Joko nyt
ymmärrät salaisuuden?»
»En, totta maar, vieläkään ymmärrä», vastasi ystävä. »Sinä olet
täällä, et asu appesi luona, etkä lupaa jouluasikaan siellä viettää!»

»Kuule siis loppuun, kuule ja kummastele! Mutta sitä ennen neiti
Lovisan malja, neiti Lovisan, josta ei ikinä tule rouva Hirtsiä. Vapaa
on neiti Lovisa, hurraa!»
»Niin se oli» — jatkoi Hirts, kun oli malja juotu. »Aika kului, ja
minun oli lähteminen vaihettamaan onneni aviosäädyn kahleiksi. En
voinut, en kehdannut sotia vastaan. Minäpä vielä kiiruhdin lähtöäni.
Vaivaloinen oli kulku Ahvenanmeren yli joulukuun pakkasessa; mutta
saavutin kumminkin vahingotta kaupungin, missä nuori mies oli
kahleisiin pantava. Olin tiellä paljon miettinyt, miten pääsisin tuosta
pulastani, ja vihdoin oli minussa syntynyt päätös. Oliko mielestäsi
siinä mitään kummaa, etten tänne tultuani paikalla rientänyt
kauppaneuvoksen luo, että ensin tahdoin nähdä sen olennon, jolle
olin määrätty? Oikeata nimeäni ilmaisematta olen nyt täällä ollut
kolmatta päivää, olen tarkastellut, nähnyt ja ymmärtänyt, ettei neiti
Lovisa Loo sovi Albert Hirts'ille. Se on totinen tosi!»
»Mutta kuinka arvelet päästä vapaaksi hänestä?» kysyi vouti. »Veli
parka! Luulenpa, että olet kutsunut tänne minun jostakusta
tärkeämmästä syystä, kuin vielä osaan arvata».
»Aivan niin! Ja kaikkein ensiksi, laita että nämä kirjeet tulevat
kauppaneuvos Loolle. Ne ovat kauppa-asioita koskevia» — lisäsi hän
nauraen — »ja jos ne eivät saa arvoani laskeumaan, niin en
ymmärrä järjestystä … vähintäin 80 prosentilla! Muuta en nyt pyydä
sinulta, mutta huomenna ja ylihuomenna voit ystävääsi auttaa».
»Eikö muuta nyt? Vähälläpä pääsen, alussa ainakin» — sanoi
vouti. »Kirjeet saat uskoa minulle. Mutta sano, oletko neitiä nähnyt,
ja mimmoinen oli hän mielestäsi? Tahtoisinpa vielä lopuksi kuulla,
mitä nyt mietit ja mitkä keinot olet keksinyt vapaussodassasi?»

»Neidin olen kolmasti nähnyt. Jos hänen itkeneistä silmistänsä voi
hänen sisällisiä avujansa päättää, niin on hän todellakin verrattava
tuohon syntejään surevaan Magdaleenaan j.n.e., j.n.e. Mutta minun
käsitykseni on aivan kykenemätön, jotta voisin hänen kuvansa
maalata niin kummallisella värisekoituksella kuin pitäisi. Siis
päätökseksi: Häneen sidottuna uhkaa minua onni niin suuri, niin
taivaallinen, että minä, joka halpuuteni tunnen, en ensinkään ole
siihen mahdollinen. Mitä viimeiseen kysymykseesi tulee, niin milloin
olet kuullut tahi nähnyt, että ilvenäytelmissä viimeisellä näytöksellä
aletaan? Hillitse siis uteliaisuutesi!»
Vielä vähän aikaa juttelivat ystävät, sitten erosivat he. Vouti otti
ollaksensa rakkauden jumalan sanansaattaja, niin kuin hän tointansa
nimitti, kun kirjeet taskussa matkaansa lähti.
Ne kirjeet, mitkä kauppaneuvos Loo sai juuri silloin, kun hän
puoliselle meni … mistä postikonttorista ne tulivat, sen arvannee
lukija.
Kauppaneuvos Loo oli ennättänyt konttorinsa ovelle, kun odotettu
vieras tuli häntä vastaan, kiireesti ja iloisesti. »Järjestys, Jumalan
nimeen, järjestys!» huusi ukko tervetuliaisiksi, kun hänen tuleva
vävynsä oli hellyyttä osoittavalla syleilyksellään kaataa hänet. Nuori
herra silloin tasaantui, ja nyt, jos olisi vouti ollut näkemässä,
mimmoiselta hänen ystävänsä näytti, kun hän siinä seisoen piti
kädessään kauppaneuvoksen laihaa, luista kättä, ei hän olisi
saattanut nauruaan hillitä. Jos olisi Hirts parka saanut kaikessa
järjestyksessä heitetyn kauhallisen kylmää vettä vasten naamaansa,
ei olisi hän hämmästyneemmältä näyttänyt. Hän oli näet luullut
saavansa lähtökäskyn paikalla, oikein ankaran, ja nyt … nyt otettiin
hän vastaan niinkuin rakas vävy ainakin.

»Tämä maa on, hitto vie, noitien … tahi on minut Troberg
pettänyt», ajatteli hän. Mutta ajatuksiaan ei hän saattanut näyttää.
Hän oli siinä verrattava kortinlyöjään, joka koko onnensa yhdelle
kortille panee ja näkee tapanneensa. Vähitellen tointui hän
kumminkin, ja kun nyt ukko Loo saattoi hänet ylikertaan, oli hän
jälleen sama mies kuin ennenkin. Ukko Loo näytti iloiselta, sen
verran kuin sitä hänessä voi näkyä. Hän olisi tosin tahtonut ottaa
vastaan tulevan vävynsä konttorissaan; mutta eipä hän näyttänyt
panevan suuresti pahaksi, että järjestys tällä kohdalla tuli vähän
rikotuksi.
Nuori herra Hirts saatettiin kamareihin, jotka häntä varten olivat
määrätyt. Sill'aikaa kuin hänen kapusäkkiään kannettiin ylös, kertoi
hän nyt täällä terveisiä vanhalta Hirts'iltä, puhui, kuten kunnollisen
kauppiaan tulee, hinnoista ja tavaroista ja miellytti vanhaa
kauppaneuvosta etenkin mainitsemalla joitakuita kauppa-asioita,
joissa kauppahuone Hirts oli voittanut summattomasti. »Se oli minun
ensi työni, ensimäinen askeleeni kauppa-alalla», lopetti hän. »Isäni
suostumuksella tosin panin onnen vaiheelle melkein koko
kauppahuoneen omaisuuden; mutta se onnistui, ja onnistui niin
hyvin, etten osannut sitä todeksi uskoa ennenkuin itse näin onneni».
»Järjestys, ennen kaikkia järjestys!» lausui siihen kauppaneuvos.
»Mutta vähäinen varovaisuuskaan noissa asioissa ei haittaa!»
»Aivan oikein!» vastasi Hirts. »Ja vastaisuudessa en panekaan
kaikkia onnen nojaan niin kuin nyt. Mutta» — keskeytti hän —
»ettekö tästä jo entisestään tiedä? Kumminkin kirjoitettiin teille siitä
jo pari viikkoa ennen lähtöäni».
»En ole saanut!… Ah! Posti tuli äsken, enkä ole vielä ennättänyt
lukea muuta kuin sinun kirjeesi».

»Jumalan kiitos!» huokasi nuori Hirts.
»Mutta järjestys!» lisäsi kauppaneuvos. »Menen toimiini. Pue
päällesi!» —
Kauppaneuvos meni lähemmin silmäilemään noita kirjeitä. Mutta
kun hän tuli konttoriinsa, oli hänellä siellä toista toimitettavaa. Posti
oli lähtevä kello 6, ja järjestys vaati, että postiin pantavat kirjeet sitä
ennen olivat sinne jätetyt. Kauppaneuvoksella oli tapana lukea läpi
kaikki kirjureittensa kirjoittamat kauppakirjeet, ennenkuin ne
lähetettiin. Tästä syystä jäivät äsken tulleet kirjeet vieläkin
avaamatta.
Sillä aikaa nuori Hirts pukeusi kosia-pukuun. Valkoinen huivi
kaulassa, yllään uusi hännystakki, joka tuli hajuvedeltä, astui hän
sitten saliin. Siellä häntä jo kauan oli vanha kielekäs rouva odottanut.
Tosin tämä saliin astuminen vähän loukkasi järjestystä, sillä tietysti
olisi kauppaneuvoksen tullut esittää Hirts rouvalleen. Mutta mitäpä
siitä; Hirts sen teki itse.
»Te, armollinen rouva, ehkä tiedätte kuka minä olen, nimittäin
kauppias
Hirts Ruotsista; samaten minä arvaan edessäni näkeväni
kauppaneuvos
Loon puolison», puhui hän nöyrästi kumarrellen.
»Te arvaatte oikein», vastasi lihava rouva ja katseli ihastellen
kaunista, nuorta herraa. »Minä olen rouva Loo… Te olette kauan
viipynyt, hyvä herra; matka meren ylitse ei liene hupaista tällä aikaa
vuodesta. Jäät ja myrskyt! Huu! Minua hirvittää, kun niitä ajattelen».

»Aivan niin! Mutta sille, joka kulkee semmoisissa asioissa kuin
minä, ei ole mitään hirvittävää». — Ja taasen kumarteli nuori herra
syvään.
»Minä tiedän nämä asiat», vastasi rouva. »Meillä, miehelläni ja
minulla, ei ole minkäänlaisia salaisuuksia. Hän osaa lukea minun
silmistäni, mitä minä ajattelen, ja minä myös hänen silmistään hänen
ajatuksensa».
»Sitä en, hitto vie, minä osaa!» ajatteli Albertimme. Mutta kovaa
sanoi hän: »Ken teidät, armollinen rouva, näkee, osaa heti, jos
vähäsenkin tavallista järkeä on, nähdä, että teissä asuu kaikki ne
avut, jotka kaunista sukupuolta kaunistavat. Ja minä olen onnellinen,
tämänkin tähden suuresti onnellinen, jos asiani täällä onnistuu».
Nuori herra Hirts, — voimme sen kohta sanoa — oli voittanut
rouvan suosion suurimmassa mitassa, ja tämä hänen viimeinen
lauseensa teki sen niin taatuksi, että Hirts varmaankin olisi
hämmästynyt, jos olisi arvannut sen.
»Mieheni on teitä kiittänyt, ja nyt minäkin voin sanoa, että tämä
kiitos ei ole ylellinen», lausui rouva, »vaikka te, kuten näen, vähän
osaatte imarrella. Minä, niinkuin sanoin, tiedän asianne, ja
heittäkäämme siis pois kaikki liikanainen kursasteleminen. Te tulette
tänne … älkää punastuko, herra … te tulette tänne meidän
tyttäremme tähden!»
»Niin!» vastasi Hirts hiljaa teeskennellyllä äänellä ja loi silmänsä
ujosti alas. »Punastuinko? Se on yhtä suuri valhe, kuin se totuus,
että sinä, eukko, painat vähintäinkin neljä sentneriä», ajatteli hän
itsekseen.

»Emme tahdo tehdä mitään tyttäremme tahtoa vastaan tässä
asiassa. Hän valitkoon tulevan tukensa itse. Mutta, niinkuin
rakastetun äidin tulee, tiedän minä, että jalon tyttäreni sydän on
vapaa, ja voinpa, sillä minä tunnen tyttäreni, voinpa hänen
puolestaan vakuuttaa, että semmoinen mies kuin te olette, on
vaarallinen naissydänten rauhalle».
Nuori Hirts kumarsi. »Jos, mitä tyttäreenne tulee, minä tässä
asiassa todellakin onnistuisin, niin olisin onnellisin maailmassa».
Nyt aukeni ovi, ja saliin astui tuo onnellisuutta tuottava neito
Lovisa Loo. Hän oli joko äitinsä tahdosta tahi oman itsensä halusta
pukeutunut kauniimpaan pukuun. Hänkin oli saanut tiedon siitä, että
heillä oli vieraana outo herra, jota ei kukaan tuntenut. Hän arvasi,
kuka tämä oli. Mutta ennenkuin hän astui saliin, oli hän avaimen
reijästä kurkistellut saliin ja nähnyt nuoren kosiansa, ja tämä näky
sai hänen sydämensä sykkimään. Hän pyyhkäsi hienolla nenäliinalla
silmiään, jotta ei niiden vetisyyttä heti havaittaisi. Ja nyt seisoi hän
siinä ujona.
»Kah! Tässä hän on, josta puhumme!» huusi vanha rouva, kun
hän tyttärensä huomasi. — »Tyttäreni Lovisa ja kauppias Hirts, isäsi
ystävän poika», esitteli rouva Loo.
Nuori herra astui kiireesti neidin luo ja puristi hellästi hänen
kättään. Mutta niinkuin nuorukaisen tuossa tilaisuudessa tulee, ei
hän sanonut mitään.
Neiti Lovisa Loo loi maahan katseensa. Sanaakaan ei hänellä ollut
sanoa.

Vanha lihava rouva hymyili ja pyyhkäsi liikutettuna silmiänsä.
Niinkuin hän sittemmin monisesti kertoi, muisti hän tuossa nyt sitä
hetkeä, jolloin Loo oli hänen kättään ensi kerran pidellyt. Yhtä ujona
oli hänkin silloin seisonut, kuin nyt hänen tyttärensä tässä. Rouva
muisti myöskin, että tällaisissa tapauksissa on parasta jättää
asianomaiset kahden kesken. Kun hän siis oli alun nähnyt, nousi hän
ja sanoi emännän äänellä: »Mutta, herra Hirts! Suokaa anteeksi! En
muistanutkaan, että tulette pitkältä matkalta ja että teillä luultavasti
on sekä nälkä että jano». — Tämän sanottuaan hän astua
kämpyröitsi salista, vähät huolien siitä, ettei Albert sanonut nälkää
eikä janoa kärsivänsä.
»Sen minä uskon! Joka armaansa vieressä seisoo, ei muista nälkää
eikä janoa», mumisi rouva mielihyvillään itseksensä.
Siinä seisoi nyt nuori mies ja nuori nainen, ja mikä pahinta oli,
molemmat tiesivät, mitä varten he siinä seisoivat. Mutta samallainen
mieli ei heissä vallinnut, ja se oli kyllä paha. Albert olisi mieluummin
häilynyt vähäisessä aluksessa keskellä myrskyistä Ahvenan merta.
Lovisa sitä vastaan tunsi taivaallisen esimaun tulevasta onnestaan.
Albert mietti, miten puhetta alkaa, ja Lovisa mitä ja miten vastata
kysymykseen, jonka tuo nuori herra tässä varmaankin pian tekisi. Ja
tätä miettiessään seisoivat he näin pari pitkää minuuttia.
Vihdoin oli nuori Hirts päässyt selville. »Se tyttö», ajatteli hän,
»joka näin ilman mitään antaupi tuntemattomalle, ei ole minkään
arvoinen, eikä siis luettane minulle syyksi, että minä tämän
ilvenäytelmän päätän niinkuin olen alkanutkin».
»Neiti Lovisa!» alkoi hän hiljaisella, vapisevalla äänellä. »Te ehkä
tiedätte, mitä vanhempamme ovat meistä päättäneet. Vähän minä
tässä kumminkin kysyn vanhempaini tahtoa. Olen nähnyt teidät, ja

ensi silmäys sanoi minulle, että ilman teidän rakkauttanne onneni
maailmassa aina on oleva vaillinainen. Tässä, tämmöinen olen,
nuori, kokematon! Sanokaa, voitteko tulevaisuudessa, sillä nyt en voi
sitä vielä pyytääkään, voitteko tahi luuletteko voivanne tulla
onnelliseksi minun puolisonani?» — Ja vielä entistä hellemmästi
puristi Albert Lovisan kättä.
Lovisa oli miettinyt, mitä tuohon kysymykseen vastata; mutta nyt,
kun tuli aika vastata, eipä hänellä ollut muuta sanomista kuin
hiljainen, mutta pitkään vedetty »jaa … a!»
Temppu oli niin muodoin tehty, ja Albert ja Lovisa olivat kihloissa.
Lovisa mietti, miten naapuri, hänen tyttärensä ja kaikki tuttavat
häntä kadehtisivat, miten nuoret herrat, jotka suurimmaksi osaksi
olivat häntä halveksineet, suuttuisivat häneen, kun näkisivät, minkä
kauniin ja uljaan sulhon hän oli saanut, ja Lovisa paran sydän paisui
enemmän kostonhimosta kuin rakkaudesta. Albert taasen
puolestansa ajatteli: »Mitä hittoa sanon hänelle?» Ja hän mietti vielä
lisäksi, pitikö liitto vahvistettaman tuolla tavallisella kosimasuutelolla!
»Enpä siitä voine päästä … mutta jääköön se niin kauvaksi kuin
mahdollista», lopetti hän keskustelun oman itsensä kanssa.
Hän olisi siinä ollut kumminkin pahassa pulassa, ellei Lovisa itse
olisi häntä vapauttanut. »Ylihuomenna on isäni syntymä- ja
nimipäivä», sanoi hän. »Isäni ja äitini tahtovat, ettei kukaan tästä
meidän kihlauksestamme saa vähintäkään vihiä, ennenkuin
ylihuomenna, jolloin meillä pidetään suuret pidot. Olkaamme siis
varuillamme, ettei kukaan huomaa mitään, ei edes palkollisetkaan!»
»Oi armaani, miten voin niin kauvan salata, mitä sydämessäni
asuu! Jos saisin tehdä, mitä sydämeni vaatii, niin tahtoisin koko

maailmalle huutaa onneani. Mutta minä jo näen sen, että sinun
vähinkin tahtosi on oleva minun lakini, ja minä koitan osoittaa sinulle
tänä odotuksen ja toivon aikana saman verran kylmyyttä, kuin
lempeä asuu sydämessäni. Suudelman, lempisuudelman vaadin
palkakseni».
»Niin, niin silloin! Ylihuomenna!», vastasi morsian. Mutta tuskinpa
olisi ollut hänellä mitään vastaan antaa palkinto paikalla, jos olisi ylkä
sitä vaatinut. »Kumminkaan» — lisäsi hän — »ei sinun tarvitse
liikanaista kylmyyttä osoittaa, sillä varmaankaan ei kukaan saata
aavistaa, mitä näin vähässä ajassa on tapahtunut».
»Se nyt oli totta!» ajatteli Albert, joka töin tuskin uskoi todeksi,
mitä neljännestunnissa oli kuullut. Myönnytykseksi hän vaan puristi
Lovisan kättä. Sen keinon oli hän keksinyt muka tunteidensa tulkiksi,
ja se riitti; sillä neiti Lovisa oli nyt niin vakuutettu tulevasta
rouvuudestaan kuin neito saattaa olla.
Salissa oli jo Albertin sinne astuessa ollut jotenkin hämärä; nyt oli
hämärä muuttunut pimeäksi, mutta kumminkaan ei ollut hän eikä
myöskään Lovisa sitä huomaavinaan. Tuollainen puolipimeys on aina
rakastavaisille tervetullut. Ja vaikka se nyt tällä kerralla ei ollutkaan
Albertin mieleen, oli se kumminkin mieluista toisesta, jota ei meidän
tarvinne kertoa.
Kauan aikaa saivat nuoret tässä kahden kesken istua, ennenkuin
kukaan heitä häiritsi. Vihdoin lienee rouvalle aika hänen askareissaan
tullut tukalaksi, sillä yhtäkkiä avasi hän oven. Lovisa silloin heti
hypähti sohvalta, missä hän oli armaansa vieressä istunut, ja
teeskennellyllä täyssydämisyydellä juoksi hän äitinsä luo ja painoi
kasvonsa hänen rintaansa vasten.

»Joko se on tapahtunut?» kysyi rouva kuiskuttaen Lovisan
korvaan, mutta kumminkin niin kovalla äänellä, että Albert sen kuuli.
»Jo!» ennätti Lovisa vastata, kun samassa nuori kosia tarttui
rouvan käteen ja lausui: »Lovisa on minun; suokaa, äiti, meille
siunauksenne!»
Vanha rouva levitti sylinsä, sulki siihen molemmat nuoret ja iski
läiskäyksen, joka oli muka suudelma olevinaan, sekä Lovisan että
Albertin huulille.
Mitä Albert siitä piti, sen annamme lukiaimme arvata. Hyväpä oli
— voimme sen sanoa — että salissa oli pimeä, muuten ehkä olisi
hänen virnistyksensä kenties osoittanut hänen rakkautensa oikean
laadun.
Saliin tuotiin nyt kynttilöitä. Jouluaaton iltaa valmistautuivat kaikki
viettämään. Miten Albert talonväen puuhatessa saatettiin ruokasaliin
ja miten häntä siellä hyvänä pidettiin, se jääköön kertomatta,
samaten kuin miten suuri, kauniisti koristettu ja sadoilla kynttilöillä
varustettu kuusi kannettiin suureen saliin,
Sillaikaa, kun tätä kaikkea ylikerrassa valmistetaan, on meidän
käyminen alikerrassa katsomassa, mitä siellä kauppaneuvos on
toimittanut ja toimittaa, kosk'ei häntä kuulu ylikertaan tulevaksi,
vaikka häntä rouva ja Lovisa levottomina odottavat.
Kauppaneuvos on vähässä ajassa saanut kirjureinsa kirjeet
luetuiksi. Nyt, kun hän siitä toimesta on päässyt, käy hän
pikimältään katsomassa, miten puotipojat täyttävät velvollisuuksiaan.
Hän on erittäin hyvällä tuulella, jota hän sillä osoittaa, että hän aina
tuon tuostakin hieroo käsiään.

Kirjeet ovat avattavat. Tämä työ on kauppaneuvoksen mielestä
varsin tärkeä. Hän pyyhkii silmälasiansa ja istuu konttorin perällä
olevalle sohvalle, asettaen kirjeet viereensä.
Pari, kolme on hän jo avannut; ne eivät sisällä muuta kuin
ilmoituksia kauppaneuvoksen kirjoittamain kirjeitten perilletulosta.
Nyt avaa hän neljännen ja lukee sen.
Koko hänen olentonsa muuttuu: hän pyyhkii uudestaan silmälasia,
ikäänkuin epäillen tokko hän on nähnyt oikein, ja lukee taasen.
Hänen vartalonsa käy suoremmaksi; hänen silmänsä säihkyvät; hän
nousee ja käy kiireesti pari kertaa lattian poikki sanaa sanomatta;
järjestyksen huutokaan ei nyt pääse hänen suustaan. Sitte istuu hän
taasen, avaa jäljellä olevat kaksi kirjettä; mutta niiden lukeminen
näkyy vaan yhä enemmän häntä suututtavan. Sillä kun hän ne on
lukenut, puristaa hän ne kaikki kolme yhteen kokoon ja pistää ne
taskuunsa, pukeutuu kiireesti turkkiinsa, valjastuttaa pikaisesti
hevosensa ja lähtee ajamaan.
Kirjurit olivat kummastellen nähneet, minkä vaikutuksen viimeksi
luetut kirjeet isännässä tekivät. He halusivat uteliaasti tietoa, minne
kauppaneuvos aikoi. He saivatkin tietää sen, sillä kovasti huusi
kauppaneuvos ajomiehelle: »Aja kestikievaritaloon!»
»Siellä on luultavasti joku matkustavainen kauppias!» sanoivat
kirjurit toisilleen.
Antakaamme kauppaneuvoksen ajaa omia matkojaan ja
käykäämme katsomaan, mitä nuo kummat kirjeet sisälsivät. Kaksi
niistä on Ruotsista, kolmas kauppaneuvoksen omasta kaupungista.
Kirjeet, jotka Ruotsista ovat, kuuluvat näin. Ensimmäinen:

»Herra Kauppaneuvos Loo! Surulla täytyy minun ilmoittaa,
että Kauppahuone Hirts nyt on perikadon partaalla. Nuori
kauppias Albert Hirts, huoneen johtajan poika, on joko isänsä
suostumuksella tahi tämän tietämättä antautunut asioihin,
jotka ovat syöneet kauppahuoneen kaikki varat, jotta ennen
joulua on Kauppahuone konkurssiin menevä. — Lähempiä
tietoja vasta. Nyt vaan päätteeksi ilmoitan, että eilen kolme
vekseliä, yhteensä 10,000 riksiä, protesteerattiin Hirtsin
Kauppahuonetta vastaan. Kunnioittaen
— —».
Toinen:
»Herra Kauppias Loo! Kun oli viime kirjeeni jo postiin pantu,
tapasin nuoren kauppiaan A. Hirtsin. Hän kehui olevansa
kihloissa Teidän tyttärenne kanssa ja valmiina lähtemään
Suomeen häitänsä viettämään. Teidän varoillanne aikoo hän
asemaansa parantaa. Hänen velkansa ovat niin suuret, että
olen katsonut velvollisuudekseni varoittaa teitä liittymästä
häneen. Muutenkin hoetaan hänestä kaikenlaisia asioita,
joista jokainen häpäisee häntä ja anastaa häneltä kaikkein
rehellisten ihmisten kunnioituksen. Joku näistä jutuista
saattaa olla valhe; mutta pelkäänpä, että hän on syyllinen
mokiin niihin seikkoihin, joista ei vielä tarkemmin tiedetä.
Kunnioittaen — —».
Kolmas, joka oli samasta kaupungista, missä kauppaneuvos asui,
kuului:

»Herra Kauppaneuvos! Ystäväni —— kehoituksesta ilmoitan
täten, että nuori kauppias A. Hirts 21 p:nä tätä Joulukuuta
saapui tänne ja asettui Troberg nimisenä asumaan
kestikievarissa. Sekä eilen että toissapäivänä on hän enimmät
ajat ollut ulkona, missä, sitä en ole saanut tietää. Minulta on
hän teidän maksettavaksenne lainannut 400 riksiä. Kun olen
valmiina lähtemään Poriin, en ole voinut suusanallisesti näitä
Teille ilmoittaa.
Kunnioittaen N——».
Näin kuuluivat ne kirjeet, jotka saivat kauppaneuvoksen
jouluaattona ulos lähtemään. Kello oli puoli kahdeksan, kun hän
palasi. Hän oli käynyt kestikievarissa ja kruununvouti Trobergin
luona, ja nyt, kun hän tuli takaisin, oli hän jotenkin vakautunut
kirjeiden todenperäisyydestä, vaikka ei hän ensinkään tuota — — ja
N——ää tuntenut. Kruununvouti, jolle kauppaneuvos oli näyttänyt
vastasaadut kirjeet, oli heti ymmärtänyt asian laidan; oli sanonut
varsin vähän tuntevansa Hirtsiä, ainoastaan kerran nähneensä hänen
eräässä kylpylaitoksessa Ruotsissa ennenkuin nyt kestikievarissa,
missä hän oli nuhdellut häntä siitä, että hän ei käyttänyt omaa
nimeänsä, vieläpä uhannut häntä, kun kuuli hänen ottaneen nimen
Troberg. — »Rahaa hän minultakin halusi lainata», lisäsi vouti, jonka
vähän kävi sääliksi kauppaneuvoksen tuska, mutta joka kumminkaan
ei tahtonut ilmoittaa ystävänsä salaisuutta, vaikka hänen kyllä
mielensä teki.
Näillä tiedoilla tuli kauppaneuvos kotiinsa, tuli vihan vimmassa ja
ennätti tuskin konttorissa saada turkit yltään, ennenkuin hän kiiruhti
ylikertaan. Puoti oli kello 6 suljettu, ja useammat kirjurit, joilla ei
ollut kotoa kaupungissa ja jotka tämän tähden olivat

kauppaneuvoksen perheeseen joulua viettämään kutsutut, olivat jo
suureen saliin saapuneet.
Isän astuessa sisään istui neiti Lovisa pianon ääressä, soitti ja
lauloi äänellä, joka todellakin oli kauniimpi kuin olisi voinut luulla.
Albert, joka seisoi laulajan vieressä, ei siis varsin liikaa sanonut, kun
hän morsiantansa ylisti ja kiitteli. — Lovisa lauloi juuri ukko Loon
tullessa erinomaisen liikuttavaa romansia kahdesta rakastavaisesta,
jotka molemmat surmasivat itsensä, kun eivät toisiansa saaneet.
    »Voik' ihanampaa olla
    Kuin kuolla kanssa armaansa,
    Ja yhtyä taas tuolla
    Ijäisen ilon taivaassa!»
lauloi hän.
»No, vihdoin viimein olet täällä!» huusi lihava rouva, kun hän
miehensä huomasi. »Missä, Jumalan nimessä, olet ollut? Mikä tärkeä
syy sai sinun tämmöisenä iltana ulos kaupungille, varsinkin kun
meillä on niin harvinainen vieras, ystäväsi poika Ruotsista».
Kauppaneuvos ei siihen vastannut mitään, silmäili vaan pikaisesti
ympärilleen ja lähestyi pianoa, jonka vieressä Albert seisoi. Albert oli
ennättänyt mennä häntä vastaan muutamia askeleita. Lovisa taasen
innostuneena jatkoi isästään huolimatta:
    »Ei loista lempi täällä.
    Tääll' yksin suru vallitsee…»
»Hm!» ärjäsi isä. »Lopeta siihen! — Minulla on vähän sanomista
Teille», lausui hän, kääntyen Hirtsiin. »Seuratkaa minua!» Ja ukko

Loo meni Albertin kamariin.
»Korvat auki nyt, Albert!» ajatteli nuori herra itsekseen, kun hän
kauppaneuvosta seurasi.
Mitä he kamarissa puhuivat, tulisi liian pitkäksi sanasta sanaan
kertoa. Vaikka viha ilman mitään järjestystä kiehui kauppaneuvoksen
sydämessä, ei hän kuitenkaan heti tätä vihaansa näyttänyt. Hän
vaan tarttui siihen keskusteluun, joka heillä oli ollut ennenkuin hän
iltapäivällä oli palannut konttoriinsa. Hän rupesi tiedustelemaan
tuosta summattomasta voitosta, joka oli saatu nuoren Hirtsin
toimella, ja kun Albert nyt suoraan tunnusti, ettei hän vielä ollut
siihen asiaan ryhtynyt, mutta että hän heti naimisensa jäljestä olisi
siihen ryhtyvä, sittenkuin olisi saanut kuulla appensa neuvoja, ja että
hän nyt vaan ilmanaikojaan huvikseen oli kertonut kauppa-asian jo
tehdyksi sekä että hän näin kyllä voi hyvin tehdä, koska asia oli niin
varmaan onnistuva, että rahat olivat jo saatuina pidettävät — niin
kauppaneuvos äkkiä keskeytti häntä huutaen: »Niin olet sinä
valehtelia, petturi, konna! Ja kaikki on totta, mitä sinusta tiedän…
Naimisesi jäljestä!… Siitä naimisesta ei tule mitään, ei tässä eikä
tulevassa maailmassa!»
Albert oli kummastellen katsovinansa kauppaneuvosta. »Ja
miks'ei?» väitti hän. »Minähän olen voittanut neiti Lovisan sydämen.
Ettehän toki peruuta sanojanne tuon pienen, viattoman valheen
tähden, joka kumminkin kohta on todeksi kääntyvä. Itsehän te
minua vävyksenne olette pyytänyt… Mitä järjestystä tämä on?»
»Järjestys! Vaiti! Lue nämät ja vastaa sitten!» ärjäsi ukko.
Värisemistä teeskennellen otti Albert vastaan kirjeet. Niitä
lukiessaan koitti hän ikäänkuin hämmästyneenä purra huuliansa.

Sitten iski hän silmänsä kauppaneuvokseen; mutta kun hän tämän
silmät havaitsi itseensä käännetyiksi, käänsi hän omat silmänsä heti
toisaalle. Näin teeskennellen sai hän kauppaneuvoksen yhä
varmemmin uskomaan kirjeiden totuutta. Ja mitä voikaan
kauppaneuvos muuta uskoa, kun Albert kirjeet luettuansa siinä seisoi
niinkuin ainakin kukistunut, joka ilkitöistä on tavattu. Kauppaneuvos
ei enää epäillyt rahtuakaan. Nuori Hirts oli suurin konna maailmassa,
ja semmoiselleko isä antaisi tyttärensä?
Kauppaneuvos hymyili — jota hänen ei kukaan muistanut ennen
tehneen — hymyili, niinkuin se hymyilee, joka on ollut
joutumaisillaan uhkaavaan vaaraan, mutta omalla nerollaan on tullut
vaaran suuruuden tietämään ja on osannut sitä välttää — hymyili ja
katsoi nyt sanaa sanomatta kukistunutta uhriansa. Albert siinä vähän
aikaa seisoi sanaa virkkamatta hänkään. Mutta yhtäkkiä ojensi hän
vartaloaan ja sanoi: »Herra kauppaneuvos! Uskokaa minua, tämä
kaikki, mitä näissä kirjeissä luette, on sulaa valhetta alusta
loppuun!» Mutta tämän sanoi hän äänellä, joka myönsi kaikki
todeksi. Kauppaneuvos ainakin piti tämän puolustuksen kokeena
päästä vapaaksi kanteesta, jonka selvimmät todistukset olivat
todeksi näyttäneet.
»Jos et olisi ystäväni Hirtsin poika, ystävän, jonka tämä käytöksesi
ja rikoksesi varmaankin hautaan vie, panettaisin minä sinut kiinni ja
lähettäisin takaisin … mutta nyt, mene, kunnoton, mene minne
tahdot, kunhan vaan paikalla siirryt minun huoneestani…»
Kauppaneuvos lausui tämän hänelle kyllä pitkän lauseen niin
liikutettuna, että se koski Albertin sydäntä.
»Te niinmuodoin ajatte minun ulos jouluaattona huoneestanne, ja
syynä siihen on, että uskotte jonkun kadehtivan valehtelian

nimettömiä kirjeitä enemmän kuin minua. Minä en voi tätä estää,
mutta kun on ehtinyt joku viikko kulua, kadutte varmaan, mitä nyt
vihassa teette». Kun Albert tämän sanoi, oli hän itse liikutettu ja olisi
nyt aivan mielellään tunnustanut, että hän itse, eikä kukaan muu, oli
kirjoittanut nuo valheelliset kirjeet. Mutta silloin olisi hänen täytynyt
mainita, mitkä syyt hänen olivat siihen saaneet, ja silloin… Jumala
tiennee, miten silloin olisi käynyt. Hän vähän mietti tunnustaa kaikki,
vaan neiti Lovisa muistui samassa hänen mieleensä ja: »ennen
kuolema, kuin elää sidottuna siihen kyöpeliin», oli hänen
päätöksensä.
»Minua on pahasti loukattu» — lisäsi hän, kun kauppaneuvos ei
mitään vastannut. »Ja uskokaa minua! Tämä loukkaus on niin suuri,
etten ikinä enää ottaisi neiti Lovisaa, vaikka se teille kuinkakin
mieleen olisi».
»Mieleen! Järjestys, poika!» huusi kauppaneuvos, joka nyt oli
päässyt entiselleen takaisin.
»Järjestys, sanon minä! Kauniilla puheilla ei minua petetä. Ulos
huoneestani paikalla tahi… Järjestys!»
»Sanokaa vihanne ja sokeutenne syyksi, että minä menen
äskeiselle morsiamelleni jäähyväisiä sanomatta» — virkkoi Hirts ja
alkoi sitoa kiinni kapusäkkiään. — »Entisen ystävyytenne nojassa
pyydän, että suotte minulle hevosen kestikievariin».
»Se odottaa teitä jo kadulla», oli kauppaneuvoksen vastaus.
Albert päätti nyt näyttäytyä närkästyneeltä. Kapusäkkiään kiinni
sitoessaan ei hän enää puhunut mitään, ja kun tämä työ oli tehty,
lähti hän, kapusäkkiä kädessään kantaen, jäähyväisiä saamatta.

»Jumalan kiitos, että se on loppunut!» — huokasi hän ovesta
etuhuoneesen päästyään. »En toiste rupea tämmöisiin. En!» — Pari
kolme askelta oli hän etuhuoneessa astunut eteenpäin, niin jo
seisahtui hän äkkiä. Salin oven edessä porstuassa näki hän naisen,
joka tämän seisahduksen sai aikaan, naisen, joka juuri aukaisi salin
oven ja heitti sisälle vähäisen vakan.
Tämä nainen oli nuori neito, niin kaunis, niin ihana, ettei ollut
Albert ennen ikinä semmoista nähnyt. Kun neitonen kuuli avattavan
kamarin ovea, joka vei porstuaan, käänsi hän kasvonsa Albertin
puoleen, ja heidän silmäyksensä kohtasivat toisiaan. Tämä
silmänluonti kesti vain sekunnin.
»Kuulutteko taloon?» kysyi Albert astuessaan neitosen luo, joka
samaten kuin hän itsekin oli valmis kulkemaan rappusia alikertaan.
»En!» vastasi neito, jonka olisi tehnyt mieli nuorelta herralta
samaa kysyä, mutta tuota tekemästä kumminkin ujoutensa hänet
esti.
»Te siis omasta puolestanne annatte joululahjoja
kauppaneuvokselle?»
Neitonen ei ennättänyt vastata, sillä ihan samassa aukaistiin salin
ovi sisäpuolelta, ja nyt näkivät Albert ja neitonen, jotka porstuassa
seisoivat, vanhan kauppaneuvoksen, joka vakka kädessä astui
heidän luoksensa.
»Kunnoton! Vieläkö uskallat sinä huonettani häväistä! Tuossa
lahjasi, tyttäreni ei siitä huoli!» huusi hän ja vetäysi kiireesti saliin
takaisin, temmaten oven jäljessänsä kiinni.

Molemmat, sekä Albert että Ellen Burg — lukiani on jo luultavasti
neitosen arvannut häneksi — pitivät kauppaneuvoksen lauseen heille
itsellensä sanotuksi. Ellen ei tätä lausetta ymmärtänyt. Albert sen
ymmärsi. — Hän ei ollut muistanut, että hänen omasta käskystään
vakkaan oli silkkihameen päälle pantu hänen oma kirjoittamansa
suljettu kirje, johon seuraava runo oli kirjoitettu:
    »Jos sa kuljet repaleissa
    Tahi silkkiin puettuna,
    Mulle rakas olet aina.
    Onneni ja kunniani
    Päivänä sä tätä kanna;
    Niinpä tietää antajaki,
    Ettei ole halpa sulle».
Tämän runon oli kauppaneuvos lukenut ja siitä nähnyt, keneltä
lahja oli. Hän ei ruvennut vakan sisustaa sen enemmän
tarkastamaan. Kaikkien salissa olevain kummaksi heitti hän sen ulos
takaisin.
»Ottakaa vakkanne, se ei enää tähän taloon ole tuleva!» sanoi
Albert yhä vielä kummastelevalle neitolle. »Lähtekäämme; minä
tahdon teille selvittää asian, sillä lahja, vakassa oleva silkkihame, on
minulta».
Ellen siinä vähän aikaa seisoi tietämättä mitä tehdä. Vihdoin, kun
nuori herra, joka Ellenin kummaksi tiesi, mitä vakassa oli, kertoi
lähtökehoituksensa, ei nuorella tytöllä ollut muuta neuvoa kuin
totella.
Kun Albert kadulle tuli, heitti hän kapusäkkinsä rekeen ja käski
rengin viedä se kestikievariin. Itse otti hän seuratakseen Elleniä.

Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.
More than just a book-buying platform, we strive to be a bridge
connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.
Join us on a journey of knowledge exploration, passion nurturing, and
personal growth every day!
ebookbell.com