2. H. Shinzawa, I. Noda / Vibrational Spectroscopy 60 (2012) 180–184 181
molecular units into a certain kind of ordered structure. This
process includes the formation of dimer based on hydrogen
bonding of carboxyl group and aggregation of dimer units
into tightly packed smectic liquid crystals. Such phenomenon
is closely related to the degree of hydrogen bonding of the
carboxyl group and segmental movements of carbon atoms
along the alkyl chain [9–13]. Consequently, the variation of the
time-dependent ATR-IR spectra of the binary mixture solution sub-
stantially reflects such transition of OA molecules. Thus, in turn,
the detailed analysis of the dynamic behavior of the spectra pro-
vides useful background information on how the OA molecules
aggregate into the smectic liquid crystals during the evaporation
of the CCl4.
2. Background of EMT
Assume a spectral data matrix A of m by n dimension, where
m is the number of spectral traces and n is the number of data
points per spectrum. Reconstructed matrix A* by singular value
decomposition (SVD) is given as follows
A∗
= USVt
(1)
The superscript t indicates the transpose of matrix. The m by r
so-called left singular matrix U contains the first r eigenvectors of
the matrix AAt, and the n by r right singular matrix V contains the
first r eigenvectors of AtA. The r by r matrix S is a relatively small
diagonal matrix. The diagonal elements of S are the first r singular
values of A, which are the positive square roots of the eigenvalues
of either AAt or AtA, arranged in the decreasing order.
The new EMT-reconstructed data matrix A** is obtained by
manipulating and replacing eigenvalues of A*. The reconstructed
data matrix A** based on general form of PC attenuating EMT is
given by
A∗∗
= A∗
−
i
kisiUiVt
i (2)
This form indicates that the data matrix A** is reconstructed by
attenuating the contribution from ith singular values. The attenua-
tion parameter ki can be chosen individually for different PCs. The
value of individual ki can be set to 0 (no attenuation), 1 (full elim-
ination) and so on, depending on the specific strategy of EMT. For
example, it is possible to suppress the effect of PC1 by setting i = 1.
The reconstructed data are then analyzed with conventional 2D cor-
relation spectroscopy scheme to elucidate the dynamic behavior of
the components presented in the data [1,14,15].
3. Experimental
Time-dependent ATR-IR spectra of a binary mixture solution of
OA and CCl4, undergoing a spontaneous evaporation process were
measured by a NEXS 870 FT-IR spectrometer equipped with a MCT
detector (Thermo Nicolet). The initial mole fraction of the OA in the
mixture solution was 0.02. The 20 L sample solution was analyzed
by depositing it on a horizontal ZnSe ATR plate. The sample was
exposed to open atmosphere at room temperature (24 ◦C), and sets
of IR spectra were collected at intervals of 4 s, with each set consist-
ing of eight coadded scans at a 4 cm−1 resolution. Once the solution
mixture was exposed to air, CCl4 started evaporating. Eventually,
CCl4 was completely removed from the system and only oleic acid
remained behind.
Fig. 1. Time-dependent ATR-IR spectra of binary mixture solution of OA and CCl4.
4. Results and discussion
4.1. 2D correlation analysis of original data matrix
Fig. 1 represents the time-dependent ATR-IR spectra of the
binary mixture solution. Peaks observed in this region are spe-
cific to vibrational modes of carboxyl group of OA. For example,
a minor peak observed at 1740 cm−1 is assignable to the monomer
of OA [16]. A major peak observed at around 1710 cm−1 exhibits
gradual increase in the spectral intensity and shift in position from
1714 to 1708 cm−1 [16]. Such variation of the spectral feature in
this region may be explained as the co-existence of the contribu-
tions from the dimers with disordered orientation and the dimers
forming quasismectic liquid crystals in which the dimers are tightly
packed together and have only short-range positional order. Iwa-
hashi et al. reported that OA tends to form a specific self-assembled
model, which provides most condensed packing form of the dimers
of OA due to the segmental movements of carbon atoms along
the alkyl chain [10]. The variation of the spectral intensity here
substantially reflects the structural alternation of OA induced by
the change in the concentration. Thus, the detailed analysis of the
change in the spectral feature may, in turn, provide useful intima-
tion on how OA molecules undergo the self-assembly. However, it
is not straightforward to investigate these spectral changes from
the conventional one-dimensional stack of the spectra. The appli-
cation of 2D correlation analysis becomes useful to elucidate such
subtle but pertinent information.
Fig. 2(A) represents synchronous correlation spectrum cal-
culated from the set of IR spectra shown in Fig. 1. The
synchronous correlation spectrum provides only one specific auto-
peak at around 1710 cm−1, indicating the gradual increase in the
concentration of the OA during the evaporation. It is important
to point out that some minor synchronous correlation peaks, for
example an auto-peak and cross peaks arising from the monomer,
do not show up in this synchronous correlation spectrum. Having
a band with a large magnitude of intensity variations creates the
problem in simultaneously displaying the fine features of bands
exhibiting only small amount of intensity variations. Consequently,
the generation of such seemingly simple correlation pattern sug-
gests that large portion of the variation of the spectral feature is
dominated with the intensity change of the dimer peaks.
The corresponding asynchronous correlation spectrum is shown
in Fig. 2(B). The generation of a negative cross peak between
1708 and 1714 cm−1 reveals that the change in the dimer occurs
before the change in the dimer cluster. Thus, it is likely that for-
mation of the dimer is followed by the aggregation of the dimer
units into a certain cluster structure. Such sequential order of
the events agrees well with the development of smectic liquid
3. 182 H. Shinzawa, I. Noda / Vibrational Spectroscopy 60 (2012) 180–184
Fig. 2. (A) Synchronous and (B) asynchronous correlation spectra calculated from
time-dependent ATR-IR spectra of binary mixture solution of OA and CCl4.
crystals via the packing of dimer units reported by Iwahashi et al.
[10]. On the other hand, one can see the development of correla-
tion feature along the 1740 cm−1 coordinate of the asynchronous
correlation spectrum, suggesting the dissimilar behavior between
the monomer and dimer components. However, the demarcation
between the monomer and dimer bands is somewhat unclear in the
asynchronous correlation spectrum. With the specific knowledge
of the spectral coordinate of the monomer peak, one may expect to
deduce the sequential order between the monomer and dimer (or
dimer cluster) from the positive asynchronous correlation intensi-
ties at 1708 and 1714 cm−1 along the 1740 cm−1 coordinate of the
asynchronous correlation spectrum. Yet the absence of the signif-
icant peak center makes the interpretation somewhat ambiguous,
and it may run a risk of overinterpretation.
4.2. 2D correlation analysis of reconstructed data matrix
So far, the main feature of the 2D synchronous and asynchronous
correlation spectrum is dominated with the major factor arising
from the dimer components, making the identification of other
minor component difficult. Thus, it becomes useful to selectively
diminish the contribution from the dominant peaks which obscure
other small but significant spectral features. Fig. 3 shows (A) score
and (B) loading vectors of the first PC (PC1) derived from PCA of the
set of IR spectra. The score value gradually decreases with the time
and the entire feature of the corresponding loading vector results in
a very similar manner with the IR spectra of OA. Linear combination
of the score and loading vectors describes the major variation of the
spectral intensity induced by the change in the concentration of OA.
Thus, the attenuation of the PC1 may in turn provide an opportunity
Fig. 3. (A) Score and (B) loading vectors of PC1.
to elucidate the subtle but important variation associated with
minor components in the system.
Fig. 4 represents (A) synchronous and (B) asynchronous correla-
tion spectra obtained from the reconstructed data, which is devoid
of the dominant contribution by the PC1 by setting the tunable
parameters in Eq. (2) as k1 = 1 and ki = 0 for i > 1. It is noted that
the reconstructed data matrix resulted in the marked enhance-
ment of selectivity in the 2D correlation spectrum. For example,
one can find that the synchronous correlation pattern in Fig. 4(A) is
now quite different from that of the conventional synchronous cor-
relation spectrum. Many correlation peaks, which are not readily
noticeable the conventional 2D correlation spectrum, become visi-
ble in Fig. 4(A). Autocorrelation peaks are observable at 1708, 1714
and 1740 cm−1, respectively. Cross correlation peaks are also obvi-
ously identifiable in the synchronous correlation spectrum. Such
enhancement of the correlation feature is mostly due to the atten-
uation of the PC1. Thus, it is demonstrated that the attenuation of
major PC is useful in simultaneously displaying the fine features of
bands exhibiting large and small amount of intensity variations.
It is noted that much more detailed features are also visible
in the asynchronous correlation spectrum shown in Fig. 4(B). The
demarcation between overlapped asynchronous correlation peak
became much sharper compared to the original asynchronous cor-
relation spectrum. Thus, cluster of peaks connected by ridges are
fragmented into distinct peaks separated by clear boundaries. For
example, one can find the development of cross peaks at 1708 and
1714 cm−1 along the 1740 cm−1 coordinate of the asynchronous
correlation spectrum. Highly overlapped correlation features are
now well resolved into the monomer, dimer and cluster compo-
nents, respectively, suggesting dissimilar behavior among these
components. Consequently, it reveals that there are three distinct
populations, each having different absorption band and unique
variation rate during the evaporation experiment.
4. H. Shinzawa, I. Noda / Vibrational Spectroscopy 60 (2012) 180–184 183
Fig. 4. (A) Synchronous and (B) asynchronous correlation spectra calculated from
PC1-attenuated data.
It is important to point out here that the simple correlation
peak-sign rules used for identifying the sequence of events in the
conventional 2D spectrum are no longer reliable, since the quanti-
tative features, especially the sequential order information in the
asynchronous spectrum, will be inevitably corrupted by the EMT
operation. The practical utility of the EMT-enhanced 2D spectrum
will be found in the area of identifying overlapped bands exhibit-
ing similar or dissimilar patterns of intensity variations. Even a very
small amount of variation will be readily detected since such infor-
mation is often carried more heavily in the secondary PC factors.
Accordingly, it becomes possible to go back to the interpreta-
tion of the conventional 2D correlation spectrum with the specific
knowledge of spectral coordinates for correlation peaks in the 2D
correlation spectrum generated from the PC1-free reconstructed
data. For example, now it seems safe to determine the sequential
order between the monomer and dimer from the positive asyn-
chronous correlations at 1708 and 1714 cm−1 along the 1740 cm−1
coordinate of the asynchronous correlation. All put together, the
simple linear sequence relationship can be derived. The first event
in the intensity variations occurs for the band due to the monomer,
followed by the intensity changes in the band assigned to the dimer.
The intensity variation arising from cluster formation based on the
dimers takes place next. Consequently, it indicates that the iso-
lated monomer substantially undergoes the self-assembly process
into the development of smectic liquid crystalline structure via the
formation of the dimer.
4.3. Comparison with other techniques
It is useful to compare the effect on the correlation feature
enhancement by EMT with other techniques, e.g. normalization
Fig. 5. (A) Synchronous and (B) asynchronous correlation spectra calculated from
normalized spectra.
[17,18] and Pareto scaling [19]. The use of normalization of raw
spectral data is a popular method to reduce various interfering
effects that can later obscure the information extracted by 2D cor-
relation analysis. For example, it can be effectively used to suppress
the overwhelmingly strong nonselective effect of the linear spec-
tral response proportional to the concentration (i.e., Beer–Lambert
Law), which may obscure much more subtle but interesting spectral
responses associated with specific molecular interactions [17,18].
Fig. 5 shows (A) synchronous and (B) asynchronous correlation
spectra calculated from the normalized spectra. One can find that
correlation features appearing in the synchronous correlation spec-
trum becomes similar to that derived from the EMT reconstructed
spectra. Many correlation peaks, which are not readily noticeable in
the conventional 2D correlation spectra, become visible in Fig. 5(A).
The negative correlation between the monomer and dimer (or
dimer cluster) reveals that changes in the monomer and dimer
concentrations occur in the opposite directions. Thus, the over-
whelmingly strong spectral intensity change proportional to the
concentration can effectively be suppressed by the normalization.
On the other hand, one can find that the asynchronous correlation
spectrum is dominated by numerical artifacts mostly due to the
unwanted effect of the noise amplification by the normalization.
The development of such numerical artifacts obviously makes the
identification of meaningful correlation peak difficult.
Normalized spectra are shown in Fig. 6. It is clear that the effect
of the noise amplification by normalization is especially acute for
the IR spectra measured at the early stage of the evaporation mea-
surement due to weak IR signals. In Fig. 6, one can find that the
exaggeration of the noise becomes obvious when the concentra-
tion of the OA is low. Consequently, the normalization of the spectra
5. 184 H. Shinzawa, I. Noda / Vibrational Spectroscopy 60 (2012) 180–184
Fig. 6. Normalized IR spectra.
with little signals exaggerates the contribution of noise rather than
enhancing minor but meaningful variation of the spectra.
Proper scaling of spectral datasets prior to 2D correlation anal-
ysis may accentuate certain useful features of the resulting 2D
spectra. For example, Pareto scaling, i.e., dividing of dataset by the
square root of its standard deviation, often tends to sharpen demar-
cation between overlapped peaks. The concept of Pareto scaling
was first introduced by Wold et al. [20] and latter expended to the
generalized scaling form by Noda [19].
Fig. 7 shows (A) synchronous and (B) asynchronous correlation
spectra calculated from the spectra subjected to the generalized
scaling. The scaled IR spectra were obtained by two scaling con-
stants set to ˛ = 0.5 (i.e. Pareto scaling) and ˇ = 1 [19]. The main fea-
ture of the synchronous correlation is dominated by the autopeak
at around 1710 cm−1, indicating the gradual increase in the
Fig. 7. (A) Synchronous and (B) asynchronous correlation spectra calculated from
scaled spectra.
concentration of OA during the evaporation. Development of cross
peaks is not identified in the synchronous correlation spectra. On
the other hand, it is interesting to point out that, in Fig. 7(B),
the asynchronous correlation intensity arising from the monomer
becomes greater by the scaling, reveling the correlation enhance-
ment of the monomer component. However, demarcation between
the monomer and dimer bands is still unclear in the asynchronous
correlation spectrum based on the scaling. In practice, an advan-
tage of the generalized Pareto scaling lies in the fact that the Noda’s
rules to interpret sign relations remain applicable to determine the
sequential order of events encoded within the set of the data, since
the signs of the cross peaks do not change by the scaling. Thus,
practice of the scaling technique will often be a reasonable starting
point to explore the even more detailed correlation relationship in
synchronous and asynchronous spectra.
Consequently, it is likely that the attenuation of the spec-
tral intensity variation associated with the concentration becomes
important to enhance the minor correlation features encoded in the
IR spectral data. The above results reveal that the selective removal
of the PC1 (i.e. concentration) by the EMT works well to enhance
the minor correlation feature in this case.
5. Conclusion
Dynamic behavior of self-assembly system of OA was studied by
2D IR correlation spectroscopy in conjunction with EMT technique.
A spontaneous evaporation process of a binary mixture solution
of OA and CCl4 was monitored by ATR-IR spectroscopy. The cor-
relation features of the conventional 2D correlation spectra are
dominated by the contribution from the major monomer compo-
nents of OA. Consequently, sufficient information is not provided
on the dynamic behavior of the minor monomer component. In
contrast, a newly reconstructed PC1-free data generated distinct
correlation features of the minor monomer component by attenu-
ating the contribution of the major dimer component contributions.
This study effectively demonstrated that 2D IR correlation anal-
ysis combined with EMT technique is a powerful tool to unravel
dynamic behavior of minor component in a complex system. Such
partial attenuation of PC may potentially become a very useful tool
in the analysis of highly congested 2D correlation spectrum often
encountered in practice.
References
[1] I. Noda, Y. Ozaki, Two-Dimensional Correlation Spectroscopy, John Willey &
Sons, Chichester, West Sussex, 2004.
[2] Y.M. Jung, S.H. Shin, S.B. Kim, I. Noda, Appl. Spectrosc. 56 (2002) 1562.
[3] Y.M. Jung, S.B. Kim, I. Noda, Appl. Spectrosc. 57 (2003) 564.
[4] Y.M. Jung, S.B. Kim, I. Noda, Appl. Spectrosc. 57 (2003) 557.
[5] Y.M. Jung, S.H. Shin, S.B. Kim, I. Noda, Appl. Spectrosc. 57 (2003) 850.
[6] D.M. Small, The Physical Chemistry of Lipids: From Alkanes to Phospholipids,
Plenum Press, New York, 1986.
[7] R.J. Hamilton, A. Bhati, Recent Advances in Chemistry and Technology of Fats
and Oils, Elsevier Applied Science Publishers, London, 1987.
[8] R.B. Gennis, Biomenbranes: Molecular Structure and Function, Springer-Verlag,
New York, 1989.
[9] M. Iwahashi, Y. Kasahara, H. Minami, H. Matsuzawa, M. Suzuki, Y. Ozaki, J. Oleo
Sci. 51 (2002) 157.
[10] M. Iwahashi, Y. Kasahara, H. Matsuzawa, K. Yagi, K. Nomura, H. Terauchi, Y.
Ozaki, M. Suzuki, J. Phys. Chem. B 104 (2000) 6186.
[11] F. Pi, H. Shinzawa, M.A. Czarnecki, M. Iwahashi, M. Suzuki, Y. Ozaki, J. Mol.
Struct. 974 (2010) 40.
[12] H. Shinzawa, M. Iwahashi, I. Noda, Y. Ozaki, J. Mol. Struct. 883–884 (2008) 27.
[13] H. Shinzawa, M. Iwahashi, I. Noda, Y. Ozaki, J. Mol. Struct. 883–884 (2008) 73.
[14] I. Noda, Appl. Spectrosc. 47 (1993) 1329.
[15] I. Noda, Appl. Spectrosc. 54 (2000) 994.
[16] M. Iwahashi, A. Umehara, K. Wakisaka, Y. Kasahara, H. Minami, H. Matsuzawa,
H. Shinzawa, Y. Ozaki, M. Suzuki, J. Phys. Chem. B 111 (2007) 740.
[17] M.A. Czarnecki, Appl. Spectrosc. 57 (2003) 107.
[18] I. Noda, Y. Ozaki, Appl. Spectrosc. 57 (2003) 110.
[19] I. Noda, J. Mol. Struct. 883–884 (2008) 216.
[20] S. Wold, E. Johansson, M. Cocchi, in: H. Kubinyi (Ed.), 3D QSAR in Drug Design:
Theory, Methods and Applications, ESCOM Science Publishers, Leiden, 1993.