r Human Brain Mapping 38:3210–3225 (2017) r

The Functional Architectures of Addition

and Subtraction: Network Discovery Using
fMRI and DCM

Yang Yang,1,2,6,7 Ning Zhong ,1,2,5,6,7* Karl Friston,3 Kazuyuki Imamura,4

Shengfu Lu,1,5,6,7 Mi Li,1,5,6,7 Haiyan Zhou,1,5,6,7 Haiyuan Wang,1,5,6,7
Kuncheng Li,7,8 and Bin Hu9*
1
Beijing Advanced Innovation Center for Future Internet Technology, Beijing University of
Technology, Beijing, China
2
Department of Life Science and Informatics, Maebashi Institute of Technology, Maebashi,
Japan
3
The Wellcome Trust Centre for Neuroimaging, University College London, London, UK
4
Department of Systems Life Engineering, Maebashi Institute of Technology, Maebashi,
Japan
5
International WIC Institute, Beijing University of Technology, Beijing, China
6
Beijing International Collaboration Base on Brain Informatics and Wisdom Services, Beijing,
China
7
Beijing Key Laboratory of MRI and Brain Informatics, Beijing, China
8
Department of Radiology, Xuanwu Hospital, Capital Medical University, Beijing, China
9
Ubiquitous Awareness and Intelligent Solutions Lab, Lanzhou University, Lanzhou,
China

r r

Abstract: The neuronal mechanisms underlying arithmetic calculations are not well understood but the
differences between mental addition and subtraction could be particularly revealing. Using fMRI and
dynamic causal modeling (DCM), this study aimed to identify the distinct neuronal architectures
engaged by the cognitive processes of simple addition and subtraction. Our results revealed signifi-
cantly greater activation during subtraction in regions along the dorsal pathway, including the left
inferior frontal gyrus (IFG), middle portion of dorsolateral prefrontal cortex (mDLPFC), and supple-
mentary motor area (SMA), compared with addition. Subsequent analysis of the underlying changes in
connectivity – with DCM – revealed a common circuit processing basic (numeric) attributes and the
retrieval of arithmetic facts. However, DCM showed that addition was more likely to engage (numeric)
retrieval-based circuits in the left hemisphere, while subtraction tended to draw on (magnitude) proc-
essing in bilateral parietal cortex, especially the right intraparietal sulcus (IPS). Our findings endorse
previous hypotheses about the differences in strategic implementation, dominant hemisphere, and the
neuronal circuits underlying addition and subtraction. Moreover, for simple arithmetic, our

Additional Supporting Information may be found in the online Received for publication 19 March 2016; Revised 25 January 2017;
version of this article. Accepted 15 March 2017.
Conflict of interest: All the authors have declared that they do not DOI: 10.1002/hbm.23585
have any competing interests. Published online 27 March 2017 in Wiley Online Library
*Corresponding author: Ning Zhong and Bin Hu, E-mail: zhong@ (wileyonlinelibrary.com).
maebashi-it.ac.jp, [email protected] Address: 460-1 Kamisadorima-
chi, Maebashi, Gunma Prefecture 371-0816, Japan, 222 Tianshui
South Road, Lanzhou, Gansu Province 730000, China

C 2017 Wiley Periodicals, Inc.

V
 r The Functional Architectures of Addition and Subtraction r

connectivity results suggest that subtraction calls on more complex processing than addition: auxiliary
phonological, visual, and motor processes, for representing numbers, were engaged by subtraction, rel-
ative to addition. Hum Brain Mapp 38:3210–3225, 2017. VC 2017 Wiley Periodicals, Inc.

Key words: arithmetic processing; fMRI; dynamic causal modeling; network discovery

r r

quantitative procedural strategies (i.e., an analogue mental

INTRODUCTION number line) that rely on the intraparietal sulci. This
Neuroimaging studies have shown that mental arithmet- hypothesized recruitment of distinct neural networks is
ic is not only a basic daily faculty but provides a powerful supported by (double dissociation) lesion studies: lesions
paradigm for characterizing fundamental cognitive pro- in the left perisylvian cortex result in impairments in mul-
cesses underlying abstract problem solving in a variety of tiplication but not subtraction, whereas lesions to regions
domains, such as magnitude representation, executive con- of the intraparietal cortex lead to difficulties with subtrac-
trol, verbal processes, and sensorimotor concepts [Ansari, tion but not multiplication [Dehaene and Cohen, 1997;
2008; Houde et al., 2010; Nieder and Dehaene, 2009, for a Dehaene et al., 2003].
review]. Moreover, pragmatically, an explicit understand- By contrast, neuroimaging studies focusing on the rela-
ing of the neural mechanisms of arithmetic calculations tionship between addition and subtraction are relatively
may also provide new pointers for education [De Smedt scarce. Behavioral studies have shown significantly longer
et al., 2011]. With an increasing knowledge of the human reaction times and lower accuracy in subtraction than in
brain’s functional anatomy, the core processes of arithmet- addition [Campbell et al., 2006; Klein et al., 2014]. Howev-
ic calculations that support mental manipulation of numer- er, it remains unclear how neuronal processes contributes
ical quantities (e.g., magnitude comparison) have been to these differences. Three kinds of hypotheses can be
associated with the bilateral intraparietal sulci (IPS) invoked to explain these differences: (1) strategic differ-
[Ansari, 2008; Dehaene et al., 2003]. Another breakthrough ences; (2) differences in hemispheric lateralization; and (3)
in mental arithmetic – in neuroimaging – is the finding of differential processing complexity. First, compared to the
the dissociation of cognitive processes that underlie mental consensus about the use of quantity-based (quantitative)
subtraction and multiplication [Yi-Rong et al., 2011; Zhou processes in subtraction, the computational strategies
et al., 2006]. Nevertheless, with respect to the two simplest involved in addition are still debated. Some studies have
operations we learn early in life, little is known about suggested that addition, like multiplication, depends on a
whether addition and subtraction employ shared or sepa- strategy based on retrieval of rote arithmetic facts, rather
rate cognitive procedures or neuronal processes. Against than a “subtraction strategy” [Schmithorst and Brown,
this background, we conducted a functional magnetic reso- 2004]. A recent study using fiber tractography supported
nance imaging (fMRI) study of mental arithmetic and the shared strategy of fact retrieval in both multiplication
examined the neuronal correlates of addition and subtrac- and addition, in view of the positive correlation observed
tion – and how they were related. in arithmetic competency (for both multiplication and
Converging evidence has identified the fronto-parietal addition) and the fractional anisotropy in the left anterior
network – including the parietal and prefrontal cortices – portion of the arcuate fasciculus – that was considered to
as the main cortical frame that supports mental arithmetic subserve phonological processing and link the frontal and
[Arsalidou and Taylor, 2011]. However, dissociations have inferior parietal cortices [Van Beek et al., 2014]. Further
been observed between arithmetic operations, especially evidence endorsing the retrieval hypothesis comes from
between subtraction and multiplication. Functional neuro- studies of children. De Smedt et al. [2011] reported a dis-
imaging data in healthy adults have revealed that the sociable pattern in children aged 10-12 years who engaged
intraparietal sulcus and the posterior superior parietal lobe the fronto-parietal network during subtraction but only
are more active during subtraction than during multiplica- activated the left hippocampus during addition. This sug-
tion; whereas the left angular and supramarginal gyri are gests reliance on the hippocampus for arithmetic fact
modulated to a greater degree by multiplication relative to retrieval during addition in the early stages of learning
subtraction [Chochon et al., 1999; Ischebeck et al., 2006]. arithmetic facts. Cho et al. [2012] reported a consistent acti-
The engagement of different strategies has been proposed vation of prefrontal and hippocampal regions during men-
to account for this dissociation [Barrouillet et al., 2008; tal addition in 7-9-year-old children. However, there are
Campbell and Xue, 2001; Dehaene et al., 2003]. Multiplica- some observations that cannot be explained by a common
tion is being attributed to rote arithmetical fact retrieval, strategy hypothesis for addition and multiplication, such
which engages the left perisylvian language areas and as common activation by both addition and subtraction in
modularizes phonological associations between a digit pair the parietal regions associated with procedural calculation
and their answer. In contrast, subtraction requires [Yi-Rong et al., 2011], and evidence suggesting distinct

r 3211 r
 r Yang et al. r

cerebral processes between addition and multiplication matrix, and C matrix) [Friston et al., 2003]. These parame-
[Rosenberg-Lee et al., 2011; Zhou et al., 2006; Zhou et al., ters correspond to the effective connectivity. The purpose
2007]. Furthermore, individual growth and functional mat- of DCM is to estimate the coupling parameters of a model
uration might result in a shift from a memory-based fron- and evaluate how well a particular model explains the
tal activation to the quantity-specific parietal activation observed data. In other words, one first identifies the best
during arithmetic processing [Rivera et al., 2005]. Second- model, in terms of which parameters exist – and then one
ly, hemispheric asymmetries in parietal were distinguished examines the parameter estimates to make inferences
by operation types: the intraparietal activation was more about directed coupling among regions. This involves
prominent in the right hemisphere during a quantity- inferring which nodes receive average (or fixed) connec-
related comparison task [Chochon et al., 1999]. A triple- tions from other nodes (encoded by the A matrix parame-
code model suggested quantity manipulation in bilateral ters), which specific connections are modulated (i.e.,
hemispheres during subtraction, relative to left-lateralized enhanced or reduced transmission between nodes) by
reliance of verbal arithmetic facts during addition and experimental conditions (encoded by the B matrix parame-
multiplication [Dehaene and Cohen, 1997]. This dissociable ters), and which nodes receive driving inputs (stimuli)
pattern was supported by laterality indices from a meta- from experimental manipulations (encoded by the C
analysis [Arsalidou and Taylor, 2011]. Nevertheless, recent matrix parameters). Typically, several pre-specified models
evidence has demonstrated that this lateralized dominance can be compared, and one model that represents the best
might not be fixed: children with developmental dyslexia explanation for the data among hypothesized models can
accompanied by reading disability and subtle deficits in be identified by Bayesian model selection (BMS) [Penny
retrieval-based arithmetic engaged the right parietal et al., 2004]. Alternatively, one can perform an exhaustive
regions for both addition and subtraction – and showed search over all possible combinations of connections to
no significant difference in performance compared with identify the best model – all average the parameter esti-
healthy controls [Evans et al., 2014]. Third, a new theory mates over models to produce the Bayesian parameter
challenging the former hypotheses has been proposed, average. In this work, we will use Bayesian model reduc-
claiming that differences between different mathematical tion to perform exhaustive searches over all combinations
operations, reported in prior studies, may have emerged of connections (and how they change between addition
due to invalid measures of complexity, rather than disso- and subtraction) to identify the underlying functional
ciable strategies or neuronal circuits [Tschentscher and architecture. This is known as structure learning or net-
Hauk, 2014]. Klein et al. [2010] investigated addition with work discovery.
different levels of difficulty (larger sum vs. smaller sum, In this study, we invited healthy subjects to solve simple
carry vs. non-carry) and proposed a flexible use of strate- addition and subtraction problems. Blood-oxygen-level-
gies based on the difficulty or complexity level. Easier dependent changes were measured during addition and
problems (with smaller sums) appear to engage the ventral subtraction. After establishing regionally specific activa-
pathway, associated with the temporo-parietal areas and tions using standard whole brain (SPM) analyses, DCM
subcortical-limbic areas, which are thought to subserve a network discovery was performed to identify the optimum
fact-retrieval strategy. More difficult problems (with larger operation-specific model, and the associated intrinsic
sums) activated the dorsal pathway associated with the (within-node) and extrinsic (between-node) connections.
fronto-parietal areas, which appears to subserve The resulting coupling parameters were used to interpret
magnitude-related (quantitative) procedural processing. A the context-sensitive changes in connectivity imposed by
recent fiber tracking study suggested a predominance of each arithmetic operation. We used simple (as opposed to
ventral fiber tracks for easy arithmetic, and dorsal as well complicated) arithmetic problems to increase the likeli-
as ventral streams for complex arithmetic tasks [Klein hood that the patterns of activations – and the implicit
et al., 2013]. connections between different brain systems – would dis-
To address the ambiguity about the neuronal processes sociate. In other words, by keeping the cognitive opera-
underlying addition and subtraction, effective connectivity tions simple, we hoped to engage the extra cognitive
analysis can be used to characterize distributed and processes required for subtraction that were not needed
context-sensitive neuronal processing by estimating the for simple addition.
influence that one neural system exerts over another at a
neuronal level [Friston, 2011]. In this study, we examined
the directed (effective) connectivity among brain regions MATERIALS AND METHODS
during addition and subtraction with dynamic causal Participants
modeling (DCM) [Friston et al., 2003]. This biophysical
model of the underlying neuronal interactions among an a Data from 20 Chinese college graduates (9 females) with
priori selected set of brain regions (nodes) is based on a a mean age of 25.76 (SD 5 3.76) years were included in the
system of bilinear differential state equations with cou- final analysis. Data from 3 participants were excluded due
pling parameters specified by three matrices (A matrix, B to unacceptable head movements within the scanning

r 3212 r
 r The Functional Architectures of Addition and Subtraction r

sessions. All of the participants were right-handed, educat- the bilateral frontal and parietal cortices. Consequently,
ed with identical schooling, had normal or corrected-to- when considering the parallel processing of units and tens
normal vision, and reported no history of neurological or separately for 2-digit numbers [Nuerk et al., 2001; Nuerk
psychiatric disorders. Prior to their participation in the et al., 2004], number size and the distance for each digit of
study, written informed consent was obtained from each a problem varied from 1 to 9 and 1 to 8, respectively. The
participant after the nature and possible consequences of frequency of occurrence of each number was balanced
the studies were explained. This study was approved by across the three conditions. Neither “tie” problems (e.g.,
the Ethics committee of Xuanwu Hospital, Capital Medical 32 1 32, 67 - 67) that involved different operand encoding
University, Beijing. [Blankenberger, 2001; Zhou et al., 2007], nor repeated
problems were used. False proposed solutions deviated
by 6 1 or 6 10 from the correct solutions in 50% of all the
Stimuli and Procedure trials. Each trial lasted for 6 s, resulting in an exposure
time of 250 ms for the first operand, 250 ms for the second
We used a block-designed fMRI experiment involving 2-
operand, 500 ms for operation sign, and 2000 ms for the
digit simple addition and subtraction problems. Partici-
proposed answer. Every two stimuli were separated by a
pants were asked to complete three types of tasks – in a
500-ms mask (background only). The inter-trial interval
sequential pattern – during scanning, including an addi-
was 1500 ms. Four successive trials of the same condition
tion task (AT), a subtraction task (ST), and a number
were presented in one block that lasted for 24 s. Blocks of
matching task (NT) that served as the baseline. Signs of three conditions were mixed and counterbalanced, and
“1”, “-”, and “#” denote the three conditions, respectively. every two task blocks were separated by a 24-second rest
All of the visual stimuli were presented in a partitioned block. Data were acquired in two functional runs with a
format and in the sequential order of “first operand”, total of 32 trials for each type of task.
“second operand”, “operation sign”, and “proposed sol-
ution” within each trial. Units and tens in the first operand
were always larger than the corresponding digits in the MR Data Acquisition
second operand, for all three conditions. This controlled The fMRI data were acquired with a 3.0 Tesla MRI scan-
for the “unit-decade compatibility effect” observed in ner (Siemens Trio Tim, Siemens Medical System, Erlanger,
number comparison tasks (e.g., faster and more accurate Germany) using a 12-channel phased array head coil.
response to 56 vs. 32, which is compatible because 5 > 3 Foam padding and headphones were used to limit head
and 6 > 2; slower and less accurate response to 43 vs. 34, motion and reduce scanning noise. One hundred ninety
which is incompatible because 4 > 3 but 3 < 4) [Nuerk two slices of anatomical images with a thickness of 1 mm
et al., 2001; Nuerk et al., 2004]. were obtained using a T1 weighted 3D magnetization pre-
Participants were required to compute “the first operand pared rapid gradient echo (MPRAGE) sequence (TR 5 1600
plus the second” in the AT, or “the first operand minus ms, TE 5 3.28 ms, TI 5 800 ms, FOV 5 256 mm, flip
the second” in the ST, or judge whether the “proposed sol- angle 5 98, voxel size 5 1 3 1 3 1 mm3). Functional images
ution” was the same as one of the two previous operands were collected using an echo-planar imaging (EPI)
in the NT. Subjects then pressed one of two response keys sequence (TR 5 2000 ms, TE 5 31 ms, flip angle 5 908,
using the left and right thumbs to report the validity of FOV 5 240 3 240 mm2, matrix size 5 64 3 64). Thirty axial
the “proposed solution” as quickly and accurately as pos- slices with a thickness of 4 mm and an inter-slice gap of
sible. Reaction times and accuracy were obtained for each 0.8 mm were acquired.
trial. The difficulty of the implicit mental arithmetic is
associated with carrying or borrowing manipulations.
Whenever such operations were required, response laten-
Data Preprocessing
cies, error rates, and functional brain activation increases The preprocessing of fMRI data was performed with
considerably [Imbo et al., 2007a; Imbo et al., 2007b]. There- SPM8 software (Wellcome Trust Centre for Neuroimaging,
fore, carrying and borrowing were excluded to ensure that London, UK, http://www.fil.ion.ucl.ac.uk). The first two
all of the problems were relatively simple. So far, few images were discarded to allow the magnetization to
studies have tried to examine the arithmetic processes approach dynamic equilibrium. The data were converted
with 2-digit simple problems. In fact, 2-digit stimuli pro- to make the fMRI data compatible with the SPM software.
vide some advantages: (1) they are more likely to activate Functional images were corrected for slice-timing differ-
the prefrontal cortex than small single-digit problems [De ences and realigned to the median image to correct rigid
Pisapia et al., 2007; Rosenberg-Lee et al., 2011; Zhou et al., body motion. Cases with head movement exceeding 2 mm
2007], and (2) they elicit neuronal responses in wider brain or 2 degrees were excluded from further analysis. The
areas, in both hemispheres [Ratinckx et al., 2006]. In this high resolution anatomical image was co-registered with
way, an unbiased comparison can be implemented by the mean image of the EPI series and then spatially nor-
using simple problems to test whether addition and sub- malized to the MNI template. After applying the spatial
traction differ, when both the operations elicit activation in normalization parameters to the EPI images, all volumes

r 3213 r
 r Yang et al. r

Figure 1.
Regions with significant activation revealed by comparing contrasts are set at a threshold of p < 0.05 voxel-wise with FDR
AT > NT, ST > NT, and ST >AT. (A) Regions responding specifi- correction and cluster size >10 voxels. Warm color bars indicate
cally to addition processing were revealed by comparing AT > NT. the t-scores specific to activation, whereas cold color bars indi-
(B) Regions specific to subtraction processing were revealed by cate the t-scores specific to deactivation. Abbreviation: AT, addi-
comparing ST > NT. (C) Compared with addition, greater activa- tion task; ST, subtraction task; NT, number matching task. [Color
tion during subtraction was revealed by comparing ST >AT. No figure can be viewed at wileyonlinelibrary.com]
significant difference resulted from the contrast of AT > ST. All

were re-sampled into 3 3 3 3 3 mm3 and smoothed with to the contrast images with the effects of sex and age mod-
an 8-mm FWHM isotropic Gaussian kernel. eled as confounds. To test for common cognitive processes
across the three conditions (i.e., basic visual encoding, dig-
it maintenance, matching judgment, and button pressing,
Statistical Parametric Mapping as shown in Figure 1), we assessed the conjunction of addi-
tion and subtraction using (group level) contrasts of
Regional responses were analyzed using SPM8. After AT > NT and ST > NT. To test whether subtraction differed
specifying the design matrix, each participant’s hemody- from addition we tested for ST > AT. All the resulting statis-
namic responses – induced by trial blocks – were modeled tical parametric maps were presented using a FDR-corrected
with a box-car function convolved with a hemodynamic threshold of p < 0.05 using a cluster extent threshold of 10
response function. The parameters for the effects of the contiguous voxels.
addition task (AT), subtraction task (ST), and number
matching task (NT) were estimated – including head Dynamic Causal Modeling and Network
motion as confounding regressors of no interest. Contrast
Discovery
images were constructed for each individual based on the
ensuing parameter estimates. At the group-level (i.e., In this analysis, we examined the directive connectivity
random-effects analysis), one-sample t-tests were applied among the brain regions identified by the whole brain

r 3214 r
 r The Functional Architectures of Addition and Subtraction r

analysis above using dynamic causal modeling (DCM) centers of VOIs, summarizing both addition and subtrac-
[Friston et al., 2003]. DCM analyses generally focus on a tion effects for dynamic causal modelling. Other common
subgraph or limited number of brain regions to reduce regions with deactivation overlapped the default mode
the number of extrinsic connections and their conditional network (DMN) [Fox et al., 2005], and thus were consid-
dependencies [Daunizeau et al., 2011]. Studying a few ered as task-negative areas and not included in DCM anal-
key regions also reduces computational load [Ma et al., ysis. Subtraction-specific regions – with activation that
2015]. Although, we used a relatively small number of could only be observed when comparing subtraction to
regions, the number of potential architectures or models the control task – were selected to represent subtraction-
entailed by each DCM is potentially enormous. However, specific areas. Most of these overlapped the dorsal path-
recent advances in DCM (Bayesian model reduction) way associated with magnitude or quantitative processing
have been introduced that can be used for structure [Klein et al., 2013]. With the help of the VOI tool in SPM8,
learning or network discovery. This approach searches time series were taken from 8 mm spheres centered on
for an optimum model by identifying the sparsity struc- each region, using the principal eigenvariate across all
ture (i.e., absence of connections, or modulatory effects) suprathreshold voxels for each participant (uncorrected
using an efficient form of model comparison [Friston P < 0.05), with the mean and other confounds removed
et al., 2011]. In what follows, we described the specifica- from the measured time-series [Stephan et al., 2010]. Please
tion of the regions or nodes and how regional activities see the results section and Figure 2 for the (average) loca-
were summarized. We then describe the connectivity archi- tion of the ensuing regions.
tectures of two (full) models that were subsequently used
to explore (reduced) variants with one or more connections DCMs specification and network discovery
(changes in connections) removed.
As implemented in DCM 12 in SPM12 (http://www.fil.
ion.ucl.ac.uk/spm), we then specified full connectivity
Regions of interest and time series extraction models: all within-node (intrinsic) and between-node
Three components are necessary for specifying a DCM: (extrinsic) connections for both the endogenous (fixed)
(i) a design matrix of external or experimental inputs and connectivity and modulation effects were considered. The
(ii) the time series stored in volume-of-interest (VOI) files term “modulation effect” denotes bilinear modulation
and (iii) on an adjacency matrix or graph specifying which effects in the context of a deterministic DCM. These bilin-
ear effects model the change in connectivity attributable to
connections at present, and how they are affected by
addition or subtraction. The number processing experi-
experimental condition. The design matrix that is optimal
mental input entered at the hierarchically lowest region
for a given DCM is sometimes subtly different than the
above the fusiform gyrus [Dehaene and Cohen, 1997;
one used for fMRI univariate analysis, since the regressors
Schmithorst and Brown, 2004]. This driving input was
of the design matrix need to be recombined to define the
modeled in terms of the C matrix. The remaining (addition
explanatory variables and inputs for the DCM. However, and subtraction) inputs exerted only modulatory effects on
this re-combination does not change the underlying gener- intrinsic and extrinsic connections. These modulatory
al model, simply the way in which the explanatory varia- effects were modeled using two B matrices.
bles are combined to explain neuronal responses in the We investigated two distinct networks comprising six
DCM. and eight nodes respectively. The first network comprised
Here, we used an experimental input that covered all nodes that showed both addition and subtraction effects,
number processing and (as above) two distinct inputs for while the second network focused on subtraction – com-
addition and subtraction: in detail, we used three experi- prising nodes that showed significant subtraction effects,
mental inputs (1) “Number processing” that comprised all relative to the number task. These two networks corre-
conditions with visual number input, i.e., AT, ST, and NT; spond to the ventral and dorsal pathway regions implicat-
(2) “Addition” that was exclusively involved in addition- ed in arithmetic computations. We will refer to them as
specific blocks and (3) “Subtraction” that modelled sub- the ventral pathway (first) and dorsal pathway (second)
traction blocks. Regions or volumes of interest (VOIs) DCMs respectively.
showing addition and subtraction effects were identified Finally, the resulting models were pruned to discover
using the group-level random effects in the aforemen- the most likely model of connections and their modulation
tioned whole brain analysis (see Figure 2). To characterize by addition and subtraction. These procedures rely upon
the differences between addition and subtraction within Bayesian model reduction [Friston et al., 2011]. This reduc-
the same distributed network, it was necessary to define tion searched over all possible reduced models of our fully
nodes that were engaged by both mathematical operations. connected models and returned a discovered network that
On the basis of the separate analysis of addition and sub- best explained the observed fMRI data. Changes in connec-
traction, the ensuing overlap motivated a conjunction anal- tivity were quantified using Bayesian model averages over
ysis to identify regions involved in both operations. The the reduced model space. These changes are expressed in
peaks of the conjunction analysis were then used as hertz – because connection strengths in DCM can be

r 3215 r
 r Yang et al. r

Figure 2.
Volumes-of-interest (VOIs) selected for the DCM analysis. mDLPFC, middle portion of dorsolateral frontal cortex;
Orange spots denote brain regions with common activation dur- pDLPFC, posterior portion of dorsolateral prefrontal cortex;
ing both addition and subtraction processes; purple spots INS, insula; IPS, intraparietal sulcus; FFG, fusiform gyrus; CAU,
denote brain regions with activation during subtraction only; the caudate; SMA, supplementary motor area; FEF, frontal eye field.
blue spot denotes where the number processing experimental [Color figure can be viewed at wileyonlinelibrary.com]
input entered the brain. Abbreviation: IFG, inferior frontal gyrus;

regarded as rate constants, i.e., the rate of change in a target ACC was 94.47 6 4.59% (mean 6 SD) for the addition task,
area produced by unit activity in another. 90.67 6 6.33% for the subtraction task, and 95.51 6 5.98%
for the number matching task. The main effect of condition
was significant, F (2, 57) 5 4.009, p 5 0.024. Post-hoc paired
RESULTS t-tests indicated that the accuracy of AT and NT were sig-
nificantly higher than that of ST after a Bonferroni correc-
Behavioral Results
tion (p 5 0.039, and p 5 0.009, respectively). The difference
The behavioral results are shown in Table I. We per- between AT and NT did not reach significance.
formed one-way repeated-measures analyses of variance The average RT was 1216.04 6 105.34 ms for the addition
on the accuracy (ACC) and reaction time (RT) using AT, task, 1259.51 6 140.17 ms for the subtraction task, and
ST and NT as the experimental conditions. The average 1244.99 6 115.21 ms for the number matching task. No

TABLE I. In-scanner behavioral results

Accuracy (% correct) Reaction Time (ms)

M SD F (2,57) P M SD F (2,57) p

Addition 94.47 4.59 4.009 0.024 1216.04 105.34 0.667 0.517

Subtraction 90.67 6.33 1259.51 140.17
Number Matching 95.51 5.98 1244.99 115.21

r 3216 r
 r The Functional Architectures of Addition and Subtraction r

TABLE II. Regions with significant activation and deactivation elicited by contrasts of AT > NT and ST > NT. Threshold
was set at p < 0.05 voxel-wise with FDR correction and cluster size >10 voxels. Abbreviation: AT, addition task; ST,
subtraction task; NT, number matching task; IPS, intraparietal sulcus; pDLPFC, posterior portion of dorsolateral pre-
frontal cortex; FFG, fusiform gyrus; FEF, frontal eye field; SMA, supplementary motor area; AG, angular gyrus; ITG,
inferior temporal gyrus; PCC, posterior cingulate cortex; mPFC, medial portions of the prefrontal cortex; L, left; R,
right; BA, Brodmann area

MNI Coordinates

Contrast Region BA Cluster x y z T-score

Activation
AT > NT L. IPS 7 28 227 254 45 4.72
R. IPS 7 33 27 254 45 5.45
L. pDLPFC 9 11 245 9 30 5.15
L. Caudate 19 29 3 18 5.94
R. Caudate 32 9 9 15 4.81
L. Insula 13 35 230 24 6 5.90
ST > NT L. FFG 37 18 248 251 218 4.04
L. IPS 40 144 227 254 48 6.76
R. IPS 7 151 27 263 45 6.05
L. pDLPFC 9 354 242 9 30 6.66
R. pDLPFC 46 225 48 3 30 6.23
R. FEF 6 26 33 0 54 4.75
L. SMA 6/32 106 23 12 54 5.82
L. Caudate 115 26 9 12 4.58
R. Caudate 151 15 215 21 4.83
L. Insula 13 85 230 24 3 7.11
R. Insula 13 47 33 21 3 6.31
Deactivation
AT > NT L. AG 39 409 254 269 21 8.85
R. AG 39 360 51 266 24 6.06
L. ITG 20 181 254 3 224 7.74
R. ITG 20 246 45 3 245 7.05
L. PCC 31 495 23 233 42 7.15
L. mPFC 11 389 29 54 0 6.68
ST > NT L. AG 39 238 251 266 30 6.83
R. AG 39 186 45 260 24 6.73
L. ITG 21 75 257 23 224 5.36
R. ITG 20 83 51 29 233 6.68
R. PCC 31 535 12 245 36 7.36
L. mPFC 11 363 26 39 218 5.88

significant differences were found for the RT among the bilateral caudate nuclei, left insula, as well as posterior
three conditions (F (2, 57) 5 0.667, p 5 0.517). portion of dorsolateral prefrontal cortex (pDLPFC, BA9) in
To ensure that behavioral results were not influenced by the left hemisphere. Regions that showed activation during
age or sex effects, we examined the Pearson correlations subtraction were detected in the left frontal cortex cover-
among age, response speed, and accuracy for all subjects, ing dorsolateral and inferior portions, left supplementary
and then implemented independent-samples t-tests on the motor area (SMA, BA 6/32), left fusiform gyrus (FFG,
behavioral performance, comparing male and female sub- BA37), right pDLPFC, right precentral gyrus/middle fron-
groups. No significant effect was found. tal gyrus (frontal eye field, FEF, BA 6), and bilateral insula,
in addition to the common areas shared with addition in
Results of Whole Brain Analyses the bilateral IPS, and bilateral caudate nuclei. Regions with
decreased activation were also revealed by tests for both
Specific neuronal correlates of addition and subtraction addition and subtraction effects, which were located in the
processes were revealed by the contrasts of AT > NT and areas that overlapped with the DMN [Fox et al., 2005],
ST > NT, respectively (see Table II). As shown in Figure including the bilateral inferior temporal gyri (ITG), bilater-
1A and 1B, regions that showed activation during addition al angular gyri (AG), medial portions of the prefrontal cor-
were found in the bilateral intraparietal sulci (IPS), tex (mPFC), and posterior cingulate cortex (PCC).

r 3217 r
 r Yang et al. r

TABLE III. Regions with significant activation detected connectivity matrix (A matrix) exhibited a highly intercon-
by contrasts of ST >AT. Thresholds were set at p < 0.05 nected (dense) infrastructure (see Figure 3A). All nodes
voxel-wise with FDR correction and cluster size >10 were linked by bidirectional connections with the excep-
voxels. Abbreviation: AT, addition task; ST, subtraction tion of connections from the left IPS to the left caudate,
task; SMA, supplementary motor area; IFG, inferior from the right IPS to the right caudate, from the left IPS to
frontal gyrus; mDLPFC, middle portion of dorsolateral the right caudate, and from the left pDLPFC to the left
frontal cortex; L, left; R, right; BA, Brodmann area caudate.
In terms of the modulatory effects of addition and sub-
MNI
traction, several connections showed enhanced or inhibito-
Coordinates
ry effects to facilitate arithmetic processing. For addition
Contrast Region BA Cluster x y z T-score (see Figure 3C), enhanced effects were imposed on connec-
tions from the left caudate to the left pDLPFC, from the
ST > AT L. SMA 8 51 26 18 48 3.78 left caudate to the left IPS, from the left insula to the left
L. IFG 45/44 28 257 24 21 4.82 pDLPFC, and the self-connection in the left caudate. Inhib-
L. mDLPFC 8/9 21 251 21 42 4.49 itory effects were inferred on connections from the left
insula to the left IPS, from the left insula to the right IPS,
from the left caudate to the left insula, from the left cau-
Differences between addition and subtraction were test- date to the right IPS, from the right caudate to the right
ed with a contrast of ST > AT, as shown in Figure 1C. IPS, and the self-connection in the right IPS. For subtrac-
Compared with addition, subtraction induced greater acti- tion (see Figure 3D), enhanced connections were found
vation in the left SMA and lower frontal area that con- from the left caudate to the left pDLPFC, from the left IPS
sisted of inferior frontal gyrus (IFG, BA45/44) and the to the right IPS, from the left caudate to the left IPS, from
middle portion of dorsolateral frontal cortex (mDLPFC) the left caudate to the right IPS, from the left insula to the
(see Table III). No significant differences resulted from the left pDLPFC, from the right caudate to the left insula, as
contrast of AT > ST. well as from the left caudate and right IPS to themselves.
The axial view of brain activation revealed by contrasts Inhibitory effects were exerted on connections from the
of AT > NT, ST > NT, and ST > AT are provided in Figure left caudate to the left insula, from the right caudate to the
S1, S2, and S3, respectively.
TABLE IV. VOIs selected for DCM analysis. Abbrevia-
Results of DCM Analysis tion: IPS, intraparietal sulcus; pDLPFC, posterior por-
tion of dorsolateral prefrontal cortex; FEF, frontal eye
Six regions corresponding to areas activated by both field; SMA, supplementary motor area; IFG, inferior
addition and subtraction were included in the first DCM. frontal gyrus; mDLPFC, middle portion of dorsolateral
These regions were identified as maxima in the corre- frontal cortex; FFG, fusiform gyrus; L, left; R, right; BA,
sponding SPM testing for the joint activation under addi- Brodmann area
tion and subtraction. The regions selected comprised the
bilateral IPS, bilateral caudate nuclei, left pDLPFC, and left MNI Coordinates
insula. The second DCM corresponded to the dorsal path-
Region BA x y z
way network for subtraction and included eight regions
showing subtraction effects, including the right pDLPFC, Common
right FEF, left SMA, left IFG, left mDLPFC, in addition to L. IPS 40 227 254 42
the bilateral IPS and left pDLPFC that were also included R. IPS 7 33 257 48
in the common network. The left FFG was chosen as the L. pDLPFC 9 245 6 30
entrance where the driving inputs were allowed to elicit L. Caudate 212 3 18
responses initially. The location and coordinates of the R. Caudate 15 23 21
ensuing nodes comprising the two networks are shown in L. Insula 13 230 21 3
Subtraction-Specific
Figure 2 and Table IV.
L. IPS 40 227 254 42
The network discovery identified optimal sparse model R. IPS 7 33 257 48
structures from initial full models by pooling model evi- L. pDLPFC 9 245 6 30
dence over subjects. Redundant connections (and changes) R. pDLPFC 9 48 3 30
with low posterior probability are effectively eliminated. R. FEF 6 33 0 54
The remaining connectivity and modulatory effects with L. SMA 6/32 23 12 54
nonzero posterior probability (and a mean of 91.02%) con- L. IFG 45/44 257 24 21
stitute the key structure of the discovered networks. L.mDLPFC 8/9 251 21 42
Driving Input
With respect to the six-node (ventral pathway) DCM
L. FFG 37 248 251 218
comprising addition and subtraction nodes, the

r 3218 r
 r The Functional Architectures of Addition and Subtraction r

Figure 3.
Networks discovered for arithmetic operations based on six subtraction) was identified; (C) modulatory effects elicited by
(ventral pathway) regions showing addition and subtraction addition processing; (D) modulatory effects elicited by subtrac-
effects. (A) Addition and subtraction processes shared a net- tion processing; Abbreviation: FFG, fusiform gyrus; IPS, intrapar-
work with common endogenous (or fixed) connectivity revealed ietal sulcus; pDLPFC, posterior portion of dorsolateral
by non-redundant parameters in the A matrix; (B) a core net- prefrontal cortex; INS, insula; CAU, caudate; L, left; R, right.
work with identical nodes, connections, and modulatory direc- [Color figure can be viewed at wileyonlinelibrary.com]
tions (enhancement or suppression with addition and

left pDLPFC, and from the right caudate to the right IPS. and subtraction processes was identified. Posterior proba-
As shown in Figure 3B, a core component of the network bilities and connection strengths (or coupling parameters
with overlapping nodes, connections, and modulatory in units of Hz) of remaining bilinear modulation effects
effects (enhancement or suppression) between addition are listed in Supplementary Table S1.

r 3219 r
 r Yang et al. r

Figure 4.
Optimal model for the extended subtraction (dorsal pathway) pDLPFC, posterior portion of dorsolateral prefrontal cortex;
network. (A) The endogenous (or fixed) connectivity was IFG, inferior frontal gyrus; FEF, frontal eye field; SMA, supple-
revealed by non-redundant entries in the A matrix; (B) modula- mentary motor area; mDLPFC, middle portion of dorsolateral
tory effects of subtraction encoded by the B matrix were frontal cortex; L, left; R, right. [Color figure can be viewed at
shown with corresponding changes in connectivity (in hertz). wileyonlinelibrary.com]
Abbreviation: FFG, fusiform gyrus; IPS, intraparietal sulcus;

With respect to the extended subtraction network used differences between addition and subtraction. To examine
to characterize effective connectivity in the magnitude- distributed processing during addition and subtraction,
related dorsal pathway, most of the eight nodes were bidi- we conducted effective connectivity analysis using DCM,
rectionally connected (see Figure 4A). Absences of fixed in a structured learning or discovery mode by identifying
connections were all afferent connections associated with optimum architectures. Our results disclosed a ventral-
the left IFG, from the left mDLPFC, right FEF, and right pathway-dependent addition network and a subtraction
IPS, respectively. A few of the connections were enhanced network comprising both ventral and dorsal pathways.
by subtraction (see Figure 4B), including those from the Moreover, the distributed nature of differences between
left IFG to the left SMA, left pDLPFC, and right IPS, addition and subtraction – in terms of strategies and later-
respectively, in addition to the connection from the left ality were also revealed by examining the connectivity
IPS to the right IPS and self-connections in the left SMA between brain regions.
and right IPS. Inhibitory effects of subtraction were seen Simple mental arithmetic problems induced activation
on connections from the left pDLPFC to the right IPS, in a wide range of brain regions that were consistent with
from the right FEF to the right IPS, and from the left most previous studies on mental arithmetic [Arsalidou
mDLPFC to itself. Parameters of modulations in the and Taylor, 2011], such as the bilateral IPS and insula, the
extended subtraction network are shown in Supplementa- left DLPFC, IFG, SMA, FFG, and caudate nuclei, and
ry Table S2. induced decreased activation located in the regions of the
default mode network (DMN) [Fox et al., 2005], including
the bilateral ITG, AG, mPFC, and PCC. As Klein et al.
DISCUSSION [2013] proposed, the bilateral IPS, posterior IPS, SMA, FEF,
and the left IFG (BA44, 45, 47) and the right IFG (BA45)
In the present study, we compared the neuronal pro- constituted the dorsal pathway; whereas the perisylvian,
cesses involved in mental arithmetic, specifically addition occipito-temporal, and subcortical areas appeared to be
and subtraction, in 20 normal healthy subjects by using connected entirely via the ventral route. Thus, many of
simple calculation problems – with a special focus on the resulting regions were specific to subtraction calculation

r 3220 r
 r The Functional Architectures of Addition and Subtraction r

(see Figure 2), and demonstrated a recruitment of a results do not allow us to ask whether the SMA is
network comprising ventral and dorsal areas during sub- involved in goal decomposing/sequencing or finger count-
traction – and only ventral areas during addition. ing, it is plausible that subtraction calls on an auxiliary
Subtraction produced significantly greater activation in process of motor representation, compared to addition.
the inferior frontal areas covering the left IFG (BA45/44) Collectively, significantly greater activation observed in
and middle part of DLPFC (mDLPFC), as well as the left the frontal areas in subtraction, relative to addition pro-
SMA than addition (as shown in Figure 1C). The left IFG cesses, is consistent with their specialization for procedural
is well known as one part of Broca’s area that is an impor- processing that implicates working memory and numeric
tant locus for linguistic processing and working memory representations in multiple forms.
[D’Esposito et al., 1999; Gabrieli et al., 1998; Mainy et al., To address the question of whether calculation differ-
2007; Poldrack et al., 1999]. Subdivisions of Broca’s area ences exist in strategic application and cerebral laterality,
have been specified and implicated in disparate functions. we performed a DCM analysis of networks based on six
It has been proposed that the superior and posterior por- regions engaged by addition and subtraction. Our ratio-
tions of the IFG (BA45/44) are related to phonological and nale was that addition and subtraction entail differential
syntactic processing, while the ventral and anterior por- switching on or off (or modulating) connections within a
tions of the IFG (BA47/45) are more dedicated to semantic distributed system. Effective connectivity describes the
processing [Hagoort, 2005; Poldrack et al., 1999]. Studies of directed propagation of activity from one neuronal system
the relationship between working memory and arithmetic to another in a context-sensitive fashion.
(with dual tasks) have suggested the involvement of pho- As suggested in many studies, the bilateral IPS may be
nological processes in procedural strategy-based calcula- the core neural substrate underlying magnitude represen-
tions as phonological load interfered with performance on tation [Ansari, 2008; Dehaene et al., 2003]. The left posteri-
arithmetic trials [Hecht, 2002; Imbo and Vandierendonck, or portion of DLPFC (BA 9) is specialized for working
2007c]. Hence, the activation in the left IFG during sub- memory [Baddeley, 2003; Smith et al., 1998], and related to
traction, especially in the superior and posterior part information segregation and integration [De Pisapia et al.,
detected in this study, is likely to correspond to extra pho- 2007]. We speculate that the left pDLPFC subserved the
nological processes that contribute to a temporary storage integration of sporadic information received from other
of verbal information, such as subvocal rehearsal of oper- nodes within the network, given the multiple afferent con-
ands or intermediate results. Another key part of the fron- nections observed in our data. Compared with the fronto-
tal area is the mDLPFC. A large body of evidence parietal areas, the left caudate nucleus and left insula –
associates this area with visual working memory [Petrides, which were categorized as fact-retrieval related ventral
2000; Stern et al., 2000]. It has been suggested that the areas [Arsalidou and Taylor, 2011; Klein et al., 2013] – also
mDLPFC is related to the active monitoring of multiple played a crucial part in both addition and subtraction. As
events in working memory, such as the executive process part of the basal ganglia, the caudate is implicated in
of monitoring and manipulation of spatial information higher-order motor control [Menon et al., 1998] as well as
[Owen et al., 1996], self-ordered processes of visual stimuli learning and memory [Graybiel, 2005; Nomura and Reber,
[Petrides et al., 1993], and so forth. In the current study, 2008]. The triple-code model suggests that the left caudate
therefore, a putative role of the mDLPFC in active moni- is involved in the retrieval of rote arithmetic facts since
toring of operands displayed sequentially can be plausibly lesions to this area impair arithmetic fact retrieval from
inferred, providing a visual support to subserve working memory, irrespective of the particular arithmetic operation
memory together with the phonological support contribut- [Dehaene and Cohen, 1995, 1997]. The insula may act as a
ed by the left IFG. With regard to the left SMA, this area toggle system that switches between outward processes
has been associated with movement sequencing for a (i.e., integration) and inward processes (i.e., mental repre-
series of mental manipulations during calculation [Clower sentation) in line with its role as a network hub (i.e.,
and Alexander, 1998; Nieder and Dehaene, 2009]. Howev- salience network), responsible for switching between the
er, recent experiments have demonstrated the great reli- executive control network and the default mode network
ance on finger-counting-related perception during mental during information processing [Arsalidou and Taylor,
calculation in 8-13-year-old children, particularly during 2011; Sridharan et al., 2008].
larger subtraction problems requiring numerical quantity Addition and subtraction specific networks were con-
processing [Berteletti and Booth, 2015]. Studies of adult structed on basis of the regions identified above (see Fig-
subjects also provide evidence for overlapping brain cir- ure 3C and D), in which differential addition and
cuits that are shared between arithmetic and finger repre- subtraction modulation effects were quantified. Within
sentation [Andres et al., 2012]. Given the crucial role of both networks, connections for propagating signals (i.e.,
SMA in representations for finger movements [Diedrichsen retrieved arithmetic facts) from the left caudate nucleus to
et al., 2013], the converging results indicate the possibility the left IPS and pDLPFC were increased, indicating the
that the SMA underlies finger perception when subjects contribution of fact retrieval to numeric processes and the
engage in arithmetic problem solving. Although our final stage of information integration, not only in addition

r 3221 r
 r Yang et al. r

but also subtraction processes. However, modulated effer- specialized for basic numeric representation, fact retrieval,
ent connections from the left insula to the bilateral IPS dis- and information integration, mental addition is more likely
tinguished the two networks. In terms of addition effects, to employ a retrieval-based approach, while mental sub-
connections from the left insula to the bilateral IPS were traction is likely to draw on the magnitude processing
inhibited. In view of the established co-activation associat- capabilities in the parietal cortex, especially the right IPS.
ed with the anterior portion of insula (as activated in this With respect to the operation-specific laterality, our ventral
study) and fronto-parietal cognitive processing areas and pathway DCM analysis emphasizes the role of the right
the causality of salience detection and executive processing IPS and right caudate nucleus. It has been proposed that
[Uddin, 2015], it is possible that observed inhibitory con- the right caudate is involved in assigning priority values
nections suppressed the functioning of the bilateral IPS or sequence to information that needs to be processed in
during addition. Conversely, such inhibition was not seen number tasks [Arsalidou and Taylor, 2011]. Enhanced cou-
in the effects of subtraction on network connectivity, sug- pling of the right caudate and the left insula during sub-
gesting more reliance on magnitude-related processing in traction suggests that this region might take a similar role
subtraction compared to prior selection of retrieved facts to the insula in coordinating competing neuronal systems.
in addition. Another apparent difference between subtrac- More active participations of the right IPS and caudate in
tion and addition comes from the distinct changes in affer- subtraction network suggest that, unlike the addition pro-
ent connections to the right IPS. Dissociations of the cesses relying primarily on the left hemisphere, mental
modulation effects on self-connections of the right IPS subtraction calls on the bilateral neural systems.
were revealed, reflecting suppression for addition and To substantiate our hypothesis that subtraction engages
enhancement for subtraction processes, respectively. Due bilateral cerebral regions, we optimized an extended fully-
to the inhibitory properties of the self-connection, the sup- connected network based on dorsal pathway (see Figure
pression (decreased self-connection) indicates unlocking of 1B and 4). A bilateral architecture is clearly seen. Howev-
neuronal excitability of sensitivity, while the enhancement er, only a few fixed connections were modulated, which
(increased self-connection) suggests more rigorous limita- were largely associated with the right IPS. In line with
tion of the neural fluctuations. The former is more likely aforementioned roles, the left IFG may support subtraction
to associate with regions showing reduced activity, e.g., by transmitting phonological information to the left SMA,
the right IPS in addition network, and the later might be pDLPFC, and the right IPS. Enhancement under subtrac-
related to regions showing greater activity, e.g., the right tion was seen for these connections. Connections from the
IPS in subtraction network and the left caudate in both left IPS to its right counterpart were also increased, sug-
networks. In addition to the inhibitory effects received gesting a greater role of the right IPS than the left IPS in
from the left insula, the afferent signal from the left cau- selecting the final answers in subtraction tasks. Inhibited
date to the right IPS for transferring arithmetic facts was afferent connections to the right IPS came from the left
also suppressed in addition network, whereas the same pDLPFC and right FEF. The former reduction is plausibly
connection was intensified in subtraction network. associated with restraints on spontaneous impulses for
The right IPS is considered to play a particularly impor- checking computations undertaken by the left pDLPFC –
tant role in the quantitative processing of numbers. to ensure efficiency. The latter inhibition is likely to result
Although bilateral parietal areas are involved in manipu- from the particular saccade direction during multi-digit
lating magnitude information, number comparison-related subtraction processing against the normal rightward read-
task performance has been found to rely more on the right ing habits of Chinese subjects. Since the right FEF was
parietal lobule, while numerical processing, requiring responsible for eye movement, and subtraction lead to left-
access to linguistic code, is more strongly associated with ward and downward shifts of spatial attention, which are
the left hemisphere [Chochon et al., 1999]. A right intra- opposite in addition [Fischer and Shaki, 2014]. Regarding
parietal lesion creates difficulties in quantitative processing modulations of self-connections, the left SMA and right
of numbers, even when knowledge about arithmetic facts IPS were inhibited, whereas activity in the visual
is unimpaired [Dehaene and Cohen, 1997]. Evidence from monitoring-related left mDLPFC was disinhibited. Taken
intraoperative cortical electro-stimulation also demon- together, the right IPS showed responses to mental sub-
strates impaired performance on simple subtraction prob- traction across the ventral and dorsal networks. In short,
lems, compared to multiplication problems when receiving compared with the left-lateralized addition, subtraction is
stimulation at right parietal areas [Yu et al., 2011]. characterized by recruitment of bilateral circuits.
Cues from dynamic interactions within the networks As for simple multi-digit arithmetic problems, our DCM
indicate how the nodes were differentially mobilized in results disclosed key differences between mental addition
addition and subtraction. We tentatively propose that sub- and subtraction in strategic application, laterality, and
tle differences in use of computational strategies, rather engagement of neuronal circuits. However, we speculate
than a double dissociation between simple multi-digit that such differences are not absolute. As the co-activation
addition and subtraction processes. Based on a common in both ventral and dorsal regions (that overlap with the
frontal-subcortical-parietal circuit (as shown in Figure 3B) subtraction pattern in current study) are also observed in

r 3222 r
 r The Functional Architectures of Addition and Subtraction r

complex addition [Klein et al., 2013; Stanescu-Cosson implicated in mental arithmetic, although its precise func-
et al., 2000]. It has been suggested that brain activation tional role has not been well established. Possible func-
induced by addition shifts from subcortical and perisyl- tions of the left AG have been linked with task-specific
vian language areas into dorsal regions, with the increas- processes, such as arithmetic fact retrieval [Dehaene et al.,
ing difficulty of the problem [Klein et al., 2013]; and 2003; Jost et al., 2011] and automatic mapping between
previously reported differences between arithmetic opera- mathematical symbols and their semantic referents
tions probably emerged due to differential task complexity [Ansari, 2008; Grabner et al., 2013]. However, no evidence
but not different types per se [Tschentscher and Hauk, has demonstrated that the left AG plays a distinct role
2014]. Our data also support the possibility of context- from other components of the DMN or that it facilitated
sensitive neuronal couplings and available circuits in addi- calculation in this study. Further study will be necessary
tion, given that addition and subtraction processes shared to explore whether operation effects exist between the left
the same (fixed) network infrastructures. Nevertheless, AG and other calculation-related regions.
when solving simple multi-digit arithmetic, we propose
that addition presents easy problems that only engage the
ventral pathway; whereas subtraction requires difficult CONCLUSION
mental arithmetic that calls on hybrid pathways. Our find-
ings suggest that, on this occasion, mental subtraction is Differential cognitive processing during simple multi-
inherently more complex than mental addition. This per- digit mental addition and subtraction were examined by
spective can also explain our behavioral results, in which means of univariate and multivariate DCM analyses. By
subtraction was performed with significantly lower accura- exploring information propagation among neuronal sys-
cy than addition (p < 0.05). To investigate whether subtrac- tems, our findings endorse earlier hypotheses about differ-
tion is always more difficult than addition, further ences between mental addition and subtraction in strategic
experiments may be necessary. application, laterality, and engagement of neural circuits.
In our study, significantly greater activation in the fron- Specifically, mental addition mainly relied on the so-called
tal cortex (e.g., left IFG) that was related to more general ventral circuit that includes temporo-parietal and
(non-numerical) cognitive processes reflected the higher subcortical-limbic areas; whereas subtraction depended on
task demands of subtraction. Nevertheless, Katzev et al. ventral as well as dorsal pathways, including extra fronto-
[2013] suggest that activation in the left IFG is mediated parietal regions. Although a common frontal-subcortical-
not only by task demands but also by individual ability. In parietal network was recruited by both operations for
this study, all of the subjects can be considered as skilled processing basic numeric quantity and retrieval of arith-
in solving simple mental arithmetic – due to their identical metic facts, addition appears to employ a retrieval-based
experience of at least 16 years of education in China. Fur- approach based on the left hemisphere, while subtraction
ther research is necessary to test whether the differences shows tendency to draw on magnitude or quantitative
we found can be observed in individuals with low arith- processing in bilateral parietal cortex, especially the right
metic ability. In addition, considering that the effect of IPS. At the easy level of difficulty, mental subtraction is
abacus-based mental calculation training might be to inherently more complex than mental addition. Auxiliary
enhance motor and visuospatial processes [Hu et al., phonological, visual, and motor processes for representing
2011], we excluded subjects who had experience with an numbers are also needed to complete the calculation of
abacus using questionnaires. subtraction problems.

Limitations ACKNOWLEDGEMENTS
Several limitations of the present study deserve com- This work was supported by grants from the National Basic
ment. First, we only focused on the neural basis of differ- Research Program of China (2014CB744600), the National
ences between simple addition and subtraction processes. Natural Science Foundation of China (61420106005 and
Therefore, we failed to provide direct evidence for the 61272345), the International Science & Technology Coopera-
orthogonal effects of difficulty. Secondly, to ensure com- tion Program of China (2013DFA32180), and the JSPS
putability under the computational loads of network dis- Grants-in-Aid for Scientific Research of Japan (26350994).
covery procedures on data from 20 subjects, regions KJF is funded by a Wellcome Trust Principal Research
showing deactivation during both addition and subtraction Fellowship (Ref: 088130/Z/09/Z).
calculations were excluded from DCM analyses. These
included the bilateral AG, ITG, the left mPFC, and the
right PCC, which responded in line with task-unrelated REFERENCES
difficulty effects associated with the default mode network Andres M, Michaux N, Pesenti M (2012): Common substrate for
(DMN) [Buckner et al., 2008; McKiernan et al., 2003]. mental arithmetic and finger representation in the parietal cor-
Among these regions, the left AG has been repeatedly tex. Neuroimage 62:1520–1528.

r 3223 r
 r Yang et al. r

Ansari D (2008): Effects of development and enculturation on Fischer MH, Shaki S (2014): Spatial associations in numerical cog-
number representation in the brain. Nat Rev Neurosci 9: nition–from single digits to arithmetic. Q J Exp Psychol (Hove)
278–291. 67:1461–1483.
Arsalidou M, Taylor MJ (2011): Is 2 1 254? Meta-analyses of brain Fox MD, Snyder AZ, Vincent JL, Corbetta M, Van Essen DC,
areas needed for numbers and calculations. Neuroimage 54: Raichle ME (2005): The human brain is intrinsically organized
2382–2393. into dynamic, anticorrelated functional networks. Proc Natl
Baddeley A (2003): Working memory: looking back and looking Acad Sci U S A 102:9673–9678.
forward. Nat Rev Neurosci 4:829–839. Friston KJ (2011): Functional and effective connectivity: a review.
Barrouillet P, Mignon M, Thevenot C (2008): Strategies in subtrac- Brain Connect 1:13–36.
tion problem solving in children. J Exp Child Psychol 99: Friston KJ, Harrison L, Penny W (2003): Dynamic causal model-
233–251. ling. Neuroimage 19:1273–1302.
Berteletti I, Booth JR (2015): Perceiving fingers in single-digit arith- Friston KJ, Li B, Daunizeau J, Stephan KE (2011): Network discov-
metic problems. Front Psychol 6:226. ery with DCM. Neuroimage 56:1202–1221.
Blankenberger S (2001): The arithmetic tie effect is mainly encod- Gabrieli JD, Poldrack RA, Desmond JE (1998): The role of left pre-
ing-based. Cognition 82:B15–B24. frontal cortex in language and memory. Proc Natl Acad Sci U
Buckner RL, Andrews-Hanna JR, Schacter DL (2008): The brain’s S A 95:906–913.
default network: anatomy, function, and relevance to disease. Grabner RH, Ansari D, Koschutnig K, Reishofer G, Ebner F
Ann N Y Acad Sci 1124:1–38. (2013): The function of the left angular gyrus in mental arith-
Campbell JI, Fuchs-Lacelle S, Phenix TL (2006): Identical elements metic: evidence from the associative confusion effect. Hum
model of arithmetic memory: extension to addition and sub- Brain Mapp 34:1013–1024.
traction. Mem Cognit 34:633–647. Graybiel AM (2005): The basal ganglia: learning new tricks and
Campbell JI, Xue Q (2001): Cognitive arithmetic across cultures. loving it. Curr Opin Neurobiol 15:638–644.
J Exp Psychol Gen 130:299–315. Hagoort P (2005): On Broca, brain, and binding: a new frame-
Cho S, Metcalfe AW, Young CB, Ryali S, Geary DC, Menon V work. Trends Cogn Sci 9:416–423.
(2012): Hippocampal-prefrontal engagement and dynamic Hecht SA (2002): Counting on working memory in simple arith-
causal interactions in the maturation of children’s fact retriev- metic when counting is used for problem solving. Mem Cognit
al. J Cogn Neurosci 24:1849–1866. 30:447–455.
Chochon F, Cohen L, van de Moortele PF, Dehaene S (1999): Dif- Houde O, Rossi S, Lubin A, Joliot M (2010): Mapping numerical
ferential contributions of the left and right inferior parietal processing, reading, and executive functions in the developing
lobules to number processing. J Cogn Neurosci 11:617–630. brain: an fMRI meta-analysis of 52 studies including 842 chil-
Clower WT, Alexander GE (1998): Movement sequence-related dren. Dev Sci 13:876–885.
activity reflecting numerical order of components in supple- Hu Y, Geng F, Tao L, Hu N, Du F, Fu K, Chen F (2011): Enhanced
mentary and presupplementary motor areas. J Neurophysiol white matter tracts integrity in children with abacus training.
80:1562–1566. Hum Brain Mapp 32:10–21.
D’Esposito M, Postle BR, Ballard D, Lease J (1999): Maintenance Imbo I, Vandierendonck A, De Rammelaere S (2007a): The role of
versus manipulation of information held in working memory: working memory in the carry operation of mental arithmetic: num-
an event-related fMRI study. Brain Cogn 41:66–86. ber and value of the carry. Q J Exp Psychol (Hove) 60:708–731.
Daunizeau J, David O, Stephan KE (2011): Dynamic causal model- Imbo I, Vandierendonck A, Vergauwe E (2007b): The role of
ling: a critical review of the biophysical and statistical founda- working memory in carrying and borrowing. Psychol Res 71:
tions. Neuroimage 58:312–322. 467–483.
De Pisapia N, Slomski JA, Braver TS (2007): Functional specializa- Imbo I, Vandierendonck A (2007c): The role of phonological and
tions in lateral prefrontal cortex associated with the integration executive working memory resources in simple arithmetic
and segregation of information in working memory. Cereb strategies. eur J Cogn PSYCHOL 19:910–933.
Cortex 17:993–1006. Ischebeck A, Zamarian L, Siedentopf C, Koppelstatter F, Benke T,
De Smedt B, Holloway ID, Ansari D (2011): Effects of problem Felber S, Delazer M (2006): How specifically do we learn?
size and arithmetic operation on brain activation during calcu- Imaging the learning of multiplication and subtraction. Neuro-
lation in children with varying levels of arithmetical fluency. image 30:1365–1375.
Neuroimage 57:771–781. Jost K, Khader PH, Burke M, Bien S, Rosler F (2011): Frontal and
Dehaene S, Cohen L (1995): Towards an anatomical and functional parietal contributions to arithmetic fact retrieval: a parametric
model of number processing. Math Cogn 1:82–120. analysis of the problem-size effect. Hum Brain Mapp 32:51–59.
Dehaene S, Cohen L (1997): Cerebral pathways for calculation: Katzev M, Tuscher O, Hennig J, Weiller C, Kaller CP (2013):
double dissociation between rote verbal and quantitative Revisiting the functional specialization of left inferior frontal
knowledge of arithmetic. Cortex 33:219–250. gyrus in phonological and semantic fluency: the crucial role of
Dehaene S, Piazza M, Pinel P, Cohen L (2003): Three parietal cir- task demands and individual ability. J Neurosci 33:7837–7845.
cuits for number processing. Cogn Neuropsychol 20:487–506. Klein E, Huber S, Nuerk HC, Moeller K (2014): Operational
Diedrichsen J, Wiestler T, Krakauer JW (2013): Two distinct ipsi- momentum affects eye fixation behaviour. Q J Exp Psychol
lateral cortical representations for individuated finger move- (Hove) 67:1614–1625.
ments. Cereb Cortex 23:1362–1377. Klein E, Moeller K, Glauche V, Weiller C, Willmes K (2013): Proc-
Evans TM, Flowers DL, Napoliello EM, Olulade OA, Eden GF essing pathways in mental arithmetic–evidence from probabil-
(2014): The functional anatomy of single-digit arithmetic in istic fiber tracking. PLoS One 8:e55455.
children with developmental dyslexia. Neuroimage 101: Klein E, Willmes K, Dressel K, Domahs F, Wood G, Nuerk HC,
644–652. Moeller K (2010): Categorical and continuous–disentangling

r 3224 r
 r The Functional Architectures of Addition and Subtraction r

the neural correlates of the carry effect in multi-digit addition. Rosenberg-Lee M, Chang TT, Young CB, Wu S, Menon V (2011):
Behav Brain Funct 6:70. Functional dissociations between four basic arithmetic opera-
Ma L, Steinberg JL, Cunningham KA, Lane SD, Kramer LA, tions in the human posterior parietal cortex: a cytoarchitectonic
Narayana PA, Kosten TR, Bechara A, Moeller FG (2015): Inhib- mapping study. Neuropsychologia 49:2592–2608.
itory behavioral control: a stochastic dynamic causal modeling Schmithorst VJ, Brown RD (2004): Empirical validation of the
study using network discovery analysis. Brain Connect 5: triple-code model of numerical processing for complex math
177–186. operations using functional MRI and group Independent Com-
Mainy N, Kahane P, Minotti L, Hoffmann D, Bertrand O, Lachaux ponent Analysis of the mental addition and subtraction of frac-
JP (2007): Neural correlates of consolidation in working memory. tions. Neuroimage 22:1414–1420.
Hum Brain Mapp 28:183–193. Smith EE, Jonides J, Marshuetz C, Koeppe RA (1998): Components
McKiernan KA, Kaufman JN, Kucera-Thompson J, Binder JR of verbal working memory: evidence from neuroimaging. Proc
(2003): A parametric manipulation of factors affecting task- Natl Acad Sci U S A 95:876–882.
induced deactivation in functional neuroimaging. J Cogn Neu- Sridharan D, Levitin DJ, Menon V (2008): A critical role for the
rosci 15:394–408. right fronto-insular cortex in switching between central-
Menon V, Glover GH, Pfefferbaum A (1998): Differential activa- executive and default-mode networks. Proc Natl Acad Sci U S
tion of dorsal basal ganglia during externally and self paced A 105:12569–12574.
sequences of arm movements. Neuroreport 9:1567–1573. Stanescu-Cosson R, Pinel P, van De Moortele PF, Le Bihan D,
Nieder A, Dehaene S (2009): Representation of number in the Cohen L, Dehaene S (2000): Understanding dissociations in
brain. Annu Rev Neurosci 32:185–208. dyscalculia: a brain imaging study of the impact of number
Nomura EM, Reber PJ (2008): A review of medial temporal lobe size on the cerebral networks for exact and approximate calcu-
and caudate contributions to visual category learning. Neuro- lation. Brain 123: 2240–2255.
Stephan KE, Penny WD, Moran RJ, den Ouden HE, Daunizeau J,
sci Biobehav Rev 32:279–291.
Friston KJ (2010): Ten simple rules for dynamic causal model-
Nuerk HC, Weger U, Willmes K (2001): Decade breaks in the
ing. Neuroimage 49:3099–3109.
mental number line? Putting the tens and units back in differ-
Stern CE, Owen AM, Tracey I, Look RB, Rosen BR, Petrides M
ent bins. Cognition 82:B25–B33.
(2000): Activity in ventrolateral and mid-dorsolateral prefrontal
Nuerk HC, Weger U, Willmes K (2004): On the perceptual gener-
cortex during nonspatial visual working memory processing:
ality of the unit-decade compatibility effect. Exp Psychol 51:
evidence from functional magnetic resonance imaging. Neuro-
72–79.
image 11:392–399.
Owen AM, Evans AC, Petrides M (1996): Evidence for a two-stage
Tschentscher N, Hauk O (2014): How are things adding up? Neu-
model of spatial working memory processing within the lateral
ral differences between arithmetic operations are due to gener-
frontal cortex: a positron emission tomography study. Cereb
al problem solving strategies. Neuroimage 92:369–380.
Cortex 6:31–38. Uddin LQ (2015): Salience processing and insular cortical function
Penny WD, Stephan KE, Mechelli A, Friston KJ (2004): Comparing and dysfunction. Nat Rev Neurosci 16:55–61.
dynamic causal models. Neuroimage 22:1157–1172. Van Beek L, Ghesquiere P, Lagae L, De Smedt B (2014): Left
Petrides M (2000): Dissociable roles of mid-dorsolateral prefrontal fronto-parietal white matter correlates with individual differ-
and anterior inferotemporal cortex in visual working memory. ences in children’s ability to solve additions and multiplica-
J Neurosci 20:7496–7503. tions: a tractography study. Neuroimage 90:117–127.
Petrides M, Alivisatos B, Evans AC, Meyer E (1993): Dissociation Yi-Rong N, Si-Yun S, Zhou-Yi G, Si-Run L, Yun B, Song-Hao L,
of human mid-dorsolateral from posterior dorsolateral frontal Chan WY (2011): Dissociated brain organization for two-digit
cortex in memory processing. Proc Natl Acad Sci U S A 90: addition and subtraction: an fMRI investigation. Brain Res Bull
873–877. 86:395–402.
Poldrack RA, Wagner AD, Prull MW, Desmond JE, Glover GH, Yu X, Chen C, Pu S, Wu C, Li Y, Jiang T, Zhou X (2011): Dissocia-
Gabrieli JD (1999): Functional specialization for semantic and tion of subtraction and multiplication in the right parietal cor-
phonological processing in the left inferior prefrontal cortex. tex: evidence from intraoperative cortical electrostimulation.
Neuroimage 10:15–35. Neuropsychologia 49:2889–2895.
Ratinckx E, Nuerk HC, van Dijck JP, Klaus W (2006): Effects of Zhou X, Chen C, Dong Q, Zhang H, Zhou R, Zhao H, Qiao S,
interhemispheric communication on two-digit arabic number Jiang T, Guo Y (2006): Event-related potentials of single-digit
processing. Cortex 42:1128–1137. addition, subtraction, and multiplication. Neuropsychologia 44:
Rivera SM, Reiss AL, Eckert MA, Menon V (2005): Developmental 2500–2507.
changes in mental arithmetic: evidence for increased functional Zhou X, Chen C, Zang Y, Dong Q, Qiao S, Gong Q (2007): Dissoci-
specialization in the left inferior parietal cortex. Cereb Cortex ated brain organization for single-digit addition and multipli-
15:1779–1790. cation. Neuroimage 35:871–880.

r 3225 r