RADIOENGINEERING, VOL. 22, NO.

4, DECEMBER 2013 969

A Memristor Model with Piecewise Window Function

Juntang YU, Xiaomu MU, Xiangming XI, Shuning WANG

Dept. of Automation, TNList, Tsinghua University, Qinghuayuan 1, 100084 Beijing, People’s Republic of China

yjt11@mails.tsinghua.edu.cn, mxm09@mails.tsinghua.edu.cn,
xxm10@mails.tsinghua.edu.cn, swang@mail.tsinghua.edu.cn

Abstract. In this paper, we present a memristor model with cally, the film consists of two regions with low and high con-
piecewise window function, which is continuously differen- centrations of dopants, respectively. Giving the full length D
tiable and consists of three nonlinear pieces. By introduc- of the film, total resistance R of this device is
ing two parameters, the shape of this window function can
 
w(t) w(t)
be flexibly adjusted to model different types of memristors. R = RON + ROFF 1 − (2)
D D
Using this model, one can easily obtain an expression of
memristance depending on charge, from which the numer- where w(t) denotes the length of the high dopant concentra-
ical value of memristance can be readily calculated for any tion region, and RON and ROFF are resistances of the film
given charge, and eliminate the error occurring in the simu- when the dopant concentration of the entire film is on high
lation of some existing window function models. and low levels, respectively.
Under the assumption of linear dopant drift in a uni-
form field with average ion mobility µV , in [4] the HP phys-
ical model is characterized by
Keywords 
w(t)

w(t)

v(t) = RON + ROFF 1 − i(t), (3a)
Memristor, window function, mathematical model. D D
dw(t) RON
= µV i(t) (3b)
dt D

1. Introduction where v(t) denotes the external voltage applied to the de-
vice, and i(t) denotes the excited current through the device.
Memristor is a fundamental two-terminal passive cir- Although (3) can yield a linear equation between memristor
cuit element, which was firstly postulated in 1971 [1]. This M and charge q, this model doesn’t take into consideration
circuit element is characterized by a nonlinear relationship boundary nonlinear dopant drift.
between charge q and flux ϕ through it. Its resistance M,
called memristance, is defined by To overcome this drawback, a window function
w(D − w)/D2 is multiplied to the right side of (3b) [4]. Let
dϕ
M= . (1) state variable x = w(t)/D. The HP physical model with win-
dq dow function can be described as
It can be seen that memristance M depends on charge q,
v(t) = (RON x + ROFF (1 − x))i(t), (4a)
which is defined as the time integral of the memristor cur-
rent. Thus, the memristor can be regarded as a nonlinear dx
= αi(t) f (x) (4b)
resistor with memory. Later, the concept of memristor was dt
extended to memristive systems and other circuit elements
where window function f (x) = x(1 − x) and constant coeffi-
with memory [2], [3].
cient α = µV RON /D2 . Generally, a class of memristor mod-
A research from Hewlett-Packard labs indicates that the els can be obtained by using different window functions.
characteristics of a nanoscale thin-film device can be suc-
According to [4], [8], a sensible window function
cessfully interpreted by the memristor theory [4]. Since then,
model should be able to depict linear dopant drift when
this promising circuit element has been widely investigated
w(t) ∈ (0, D), and boundary nonlinear dopant drift when
in various areas, ranging from nonlinear oscillators [5], [6]
w(t) → 0+ or w(t) → D− . However, the model (4) cannot
to logic applications [7]. Particular attention is also devoted
eliminate nonlinear effects when w(t) is around D/2. Addi-
to modeling the memristor, which is of great significance for
tionally, it lacks the flexibility in adjusting the shape of the
designing memristor circuits and analyzing their nonlinear
window function f (x) to model different types of memris-
dynamics.
tors. To address these problems, Y. Joglekar and S. Wolf
In [4], the nanoscale device is a semiconductor thin- propose another window function f p (x) = 1 − (2x − 1)2p in
film sandwiched between two metal contacts. More specifi- [9], where p ∈ Z+ . This window function can satisfy the
970 J. YU, X. MU, X. XI, S. WANG, A MEMRISTOR MODEL WITH PIECEWISE WINDOW FUNCTION

previous demands well in the case of p  1. Besides, dif- b(1 − 4a2 )

> 2. (7)
ferent realizations of window function also have been dis- 4a2
cussed in literature, such as 1 − (x − sgn(−i))2p in [10] and
1 − ((x − 0.5)2 + 0.75)q in [11], where sgn(·) is sign func- This can be shown as follows. Let h(z) = (1 + b)zb −
b b−1
tion, p ∈ Z+ , and q ∈ R+ . 4a2
z + 1. For any a and b satisfying (7), it can be seen
that h(0) > 0 and h(1) < 0. Hence, there exists a solution
Replacing f (x) with these window functions in (4b), z0 of (6b) such that 0 < z0 < 1, in view of the continuity of
the explicit functional relation between memristance M and √
h(z). Then x0 = 0.5 − z0 a and k = 1/(x0 (1 − x0 )(1 + zb0 ))
charge q cannot be derived from these models (from a flux- are a pair of solution for (6). In general, we prefer to the
controlled memristor model with window function f p (x), smallest possible x0 to enhance the shape controllability of
only a quite complicated relationship between memristance this window function. Thus, the window function fPW (x)
M and flux ϕ can be obtained in the case of p > 1 [12]). is determined by the two parameters a and b. It should be
Thus, in the simulation of these window function models, the remarked that the condition (7) can be easily satisfied by
numerical methods for solving differential equations have to choosing sufficiently large b for any given a ∈ (0, 0.5).
be applied. There exists the non-negligible simulation error
in some cases. This limits the application of these memristor With proper parameters a and b, i.e. a → 0.5− and
models to design and analyze memristor circuits. b  1, the window function fPW (x) is close to 1 when
0 < x < 1, which can model the linear dopant drift of mem-
Except window function model, there are many other ristor. It can also be seen that fPW (0) = 0 and fPW (1) = 0
memristor models [13], [14], [15], which introduce the non- from (5), which can model the boundary nonlinear dopant
linear dependence of the state variable derivation and the drift of memristor. A family of window functions fPW (x)
memristance on the current. These models also suffer from are illustrated in Fig. 1.
the same problem.
(a)
In this paper, a memristor model with piecewise win- 1
a=0.35
dow function is presented to solve this problem. The pro- 0.9
a=0.4
0.8 a=0.45
posed window function is continuously differentiable and
0.7
consists of three nonlinear pieces. The single-value func- 0.6
fP W (x)

tion between memristance M and charge q can be obtained 0.5

from this model, and the numerical value of M(q) can be 0.4

easily computed for any given q at a considerable precision 0.3

level. Thus, the simulation error can be eliminated. In ad- 0.2

0.1
dition, the model shows more flexibility than other window 0
function models. 0 0.2 0.4 0.6
State variable x
0.8 1

(b)
1
b=4
0.9
b=8

2. Memristor Model with Piecewise 0.8 b=12

0.7

Window Function 0.6

fP W (x)

0.5
The piecewise window function proposed in this paper 0.4
is 0.3

(1 + ( x−0.5 2b −1

a ) ) , for x0 ≤ x ≤ 1 − x0
0.2
fPW (x) = (5)
kx(1 − x), otherwise 0.1

where a ∈ (0, 0.5), b ∈ Z+ , and x0 , k ∈ R+ . In order to en-

0 0.2 0.4 0.6 0.8 1
State variable x
sure the continuous differentiability of this window function Fig. 1. A family of piecewise window functions fPW (x).
for given a and b, x0 and k should satisfy (a) Given b = 10, when a = 0.35, 0.4, 0.45, the corre-
sponding values of x0 are 0.0233, 0.0239, 0.0335 and
1 those of k are 0.0909, 1.2771, 10.1102. (b) Given a =
kx0 (1 − x0 ) − = 0, (6a)
1 + zb0 0.4, when b = 4, 8, 12, the corresponding values of x0
are 0.0806, 0.0306, 0.0198 and those of k are 5.4842,
b
(1 + b)zb0 − 2 zb−1 + 1 = 0, (6b) 2.4194, 0.6338. Actually, x0 and 1 − x0 are smooth con-
4a 0 nection points of two fragments of (5).
x0 − 0.5 2
 
z0 = , (6c) Using this window function, the memristor model can
a
be described by
1
x0 < . (6d)
2 v(t) = (RON x + ROFF (1 − x))i(t), (8a)
There exists the solution of (6) if the given a and b sat- dx
= αi(t) fPW (x). (8b)
isfy that dt
RADIOENGINEERING, VOL. 22, NO. 4, DECEMBER 2013 971

Let x(0) = x(t) |t=0 and q(0) = q(t) |t=0 . Without loss of 15

generality, assume that x(0) < x0 and q(0) = 0, in view of

the fact that q(t) can be simply replaced by q0 (t) = q(t) + c
10

in other cases where c is a constant. Integrating (8b), it fol- 5

lows that

g(y)
0

x(t) = 1/(1 + exp(−αkq(t) − c1 )) (9) −5

−10
for q(t) < q1 , where c1 = ln x(0) − ln(1 − x(0)) and q1 =
(ln x0 −ln(1−x0 )−c1 )/αk. Letting auxiliary variable y(t) = −15
−1.5 −1 −0.5 0 0.5 1 1.5
1 y
(x(t) − 0.5)/a and g(y) = y + 2b+1 y2b+1 (see Fig. 2), the dif-
1
ferential equation (8b) yields the following equation Fig. 2. Illustration of the function g(y) = y + 2b+1 y2b+1 , b = 10.

g(y) = α0 q(t) − c2 (10) (a)

1

0.9

for q1 ≤ q(t) ≤ q2 , where α0 = α/a, c2 = α0 q1 − g((x0 − 0.8

0.5)/a) and q2 = (g((0.5 − x0 )/a) + c2 )/α0 . Let g−1 de- 0.7

State variable x
note the inverse function of the one-to-one mapping g(y). 0.6

It’s worth noting that the numerical value of g−1 (u) for any 0.5

0.4
u ∈ R can be readily calculated by Newton Iterative Method 0.3

in several iterations at a considerable precision level. In par- 0.2

ticular, the initial value of Newton Iterative Method can be 0.1

always taken as y = 0. And this algorithm converges for any 0

−0.06 −0.04 −0.02 0 0.02 0.04 0.06 0.08 0.1

given u in this problem. Then from (10), the variable y has Charge q / C
(b)
the form of 140

y = g−1 (α0 q − c2 ). (11)

120
Memristance M(q) / Ω

It follows that for q1 ≤ q(t) ≤ q2 , 100

80
x(t) = ag−1 (α0 q(t) − c2 ) + 0.5. (12)
60

40
Similarly, from (8b) it can be obtained that
20

x(t) = 1/(1 + exp(−αkq(t) − c3 )) (13)

0
−0.06 −0.04 −0.02 0 0.02 0.04 0.06 0.08 0.1
Charge q / C
for q > q2 , where c3 = ln(1 − x0 ) − ln x0 − αkq2 . Fig. 3. Illustrations of the state variable x(q) and the memris-
tance M(q). a = 0.4, b = 10, x(0) = 0.01, RON = 1 Ω,
Therefore, combining (9), (12) and (13), the solution of
ROFF = 125 Ω, D = 10 nm, µV = 10−10 cm2 s−1 V−1 (the
(8b) is corresponding values of x0 and k are 0.0239 and 1.2771,
respectively).
 (1 + exp(−αkq(t) − c1 ))−1 , for q(t) < q1 ,


x(t) = ag−1 (α0 q(t) − c2 ) + 0.5, for q1 ≤ q(t) ≤ q2 , as shown in Fig. 4. It is consistent with the experimen-
(1 + exp(−αkq(t) − c3 ))−1 , for q(t) > q2 .

tal results in [4]. The v − i characteristic curve “shrinks to
(14) a straight line” when the power frequency increases to 10
Using (1), (8a) and (14), the memristance M(q) is found to times larger, as illustrated in Fig. 5. Thus, the simulation
be results accord with the memristor fingerprint in [16].
M(q) = ROFF − (ROFF − RON )x(t). (15)
There are several benefits of the piecewise window
So far, the solution of (8) has been obtained. function model. One advantage is that the expression of
memristance M(q) can be easily obtained and the numerical
value of M(q) can be readily calculated from this expression.
Due to the simplicity of the function g−1 , this memristance
3. Simulation and Analysis expression can be applied as conveniently as the analytic so-
In this section, we do some experiments on the piece- lution to design and analyze memristor circuits. Here we
wise window function model with different driving sources give an example to show the superiority of using the mem-
to test its performances, and indicate main advantages of this ristance expression.
model.
When the analytic solution of the memristor model de-
When driven by a sinusoidal voltage source, the scribed by (4) (of course, f (x) should be replaced by the
“pinched hysteresis loop” in the v − i plane can be observed, corresponding window function) cannot be obtained, the
972 J. YU, X. MU, X. XI, S. WANG, A MEMRISTOR MODEL WITH PIECEWISE WINDOW FUNCTION

(a) (a)
1 0.013

0.9
0.0125
0.8

State variable x(t)

State variable x(t)

0.7 0.012

0.6
0.0115
0.5
0.011
0.4

0.3 0.0105

0.2
0.01
0.1

0 0.0095
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Time / s Time / s
(b) (b)
2 0.04

1.5 0.03

1 0.02
Current / A

Current / A
0.5 0.01

0 0

−0.5 −0.01

−1 −0.02

−1.5 −0.03

−2 −0.04
−4 −3 −2 −1 0 1 2 3 4 −4 −3 −2 −1 0 1 2 3 4
Voltage / V Voltage / V

Fig. 4. Simulation results of the memristor model with piece- Fig. 5. Simulation results of the memriestor model with piece-
wise window function when driving source V (t) = wise window function when driving source V (t) =
4 sin(πt). The values of parameters are the same as 4 sin(10πt). The values of parameters are the same as
Fig. 3. Fig. 3.
(a)
1
numerical methods for solving differential equations such
0.9
as the Runge-Kutta method usually have to be applied in 0.8
the simulation of the memristor model. This brings non-
State variable x(t)

0.7

negligible error when the excitation source has large ampli- 0.6

tude or low frequency. Taking the simulation of the memris- 0.5

tor model with window function f p (x) as an example, it can 0.4

be shown in detail as follows. 0.3

0.2

Fig. 6 illustrates the simulation results of the memristor 0.1

model with window function f p (x) when using the Runge- 0

0 0.5 1 1.5 2 2.5 3 3.5 4
Time / s
Kutta method. From (4), it can be obtained that (b)
4
dx v(t)
=α f p (x). (16) 3.5

dt RON x + ROFF (1 − x) 3

2.5
Current / A

It follows that 2
Z t Z x(t)
RON x + ROFF (1 − x) 1.5
v(t)dt = dx. (17)
0 x(0) α f p (x) 1

0.5

Considering v(t) = 4 sin(πt), the above equation (17) yields 0

Z x(t) −0.5
4 RON x + ROFF (1 − x) −4 −3 −2 −1 0
Voltage / V
1 2 3 4

(1 − cos(πt)) = dx. (18)

π x(0) α f p (x) Fig. 6. Simulation results of the memristor model with win-
R x(t)
Letting H(x(t)) = dow function f p (x) when driving source V (t) = 4 sin(πt).
x(0) (RON x + ROFF (1 − x))/α f p (x)dx, it f p = 1−(2x −1)2p , p = 3, the values of other parameters
can be seen that are the same as Fig. 3.

H(x(t)) = H(x(t + T )) (19) Obviously, the simulation results in Fig. 6 violate the rela-
tionship x(t) = x(t +T ). This is due to the simulation numer-
where T = 2s. In view of RON x + ROFF (1 − x) > 0 and ical error. Actually, when H(x) is sufficiently large, x → 1
f p (x) > 0 for any x ∈ (0, 1), H(x) is a strictly monotone in- and f p (x) → 0. Thus, a small error of f p (x) can cause a sig-
creasing function of x. Therefore, it holds that nificant difference of 1/ f p (x). Therefore, there is a signifi-
cant difference between the state variable curves of the two
x(t) = x(t + T ). (20) excitation cycles in Fig. 6.
RADIOENGINEERING, VOL. 22, NO. 4, DECEMBER 2013 973

When using the piecewise window function model, the the shape of this window function can be flexibly adjusted
numerical value of M(q) can be directly obtained from the to model different types of memristors. It is significant that
memristance expression at a considerable precision level, the expression of memristance M depending on charge q can
which avoids numerically solving the differential equations. be derived from this model and the numerical value of M(q)
Hence, the simulation error can be eliminated (see Fig. 4). can be easily computed for any given q. Using the memristor
model, the error occurring in the simulation of some existing
A second advantage is that the memristor model with
window function models can be eliminated.
piecewise window function possesses strong flexibility, due
to its parameters a and b. The linear dopant drift inter-
val of state variable x(t) can be adjusted approximately
to (0.5 − a, 0.5 + a), and the slope of fPW (x) on bound-
aries is determined by the value of b. It can be seen that
Acknowledgements
fPW (x) → 1 for x ∈ (0.5 − a, 0.5 + a) and fPW (x) → 0 for This project is jointly supported by the National Nat-
x ∈ [0, 0.5 − a) ∪ (0.5 + a, 1] with sufficiently large b (see ural Science Foundation of China (61074118, 61134012)
Fig. 7). However, the window function f p (x) with only one and the National Basic Research Program of China
controlling parameter p cannot take values close to 0 for (2012CB720505).
a given boundary interval of state variable x(t). Therefore,
the shape of the window function fPW (x) can be more flexi-
bly adjusted to model different types of memrisotrs. References
1
[1] CHUA, L. O. Memristor-the missing circuit element. IEEE Transac-
0.9
tions on Circuit Theory, 1971, vol. 18, no. 5, p. 507 - 519.
0.8

0.7 [2] CHUA, L. O., KANG, S. M. Memristive devices and systems. Pro-
0.6 ceedings of the IEEE, 1976, vol. 64, no. 2, p. 209 - 223.
fP W (x)

0.5
[3] DI VENTRA, M., PERSHIN, Y. V., CHUA, L. O. Circuit elements
0.4
with memory: memristors, memcapacitors, and meminductors. Pro-
0.3
ceedings of the IEEE, 2009, vol. 97, no. 10, p. 1717 - 1724.
0.2

0.1 [4] STRUKOV, D. B., SNIDER, G. S., STEWART, D. R., WILLIAMS,

0
0 0.2 0.4 0.6 0.8 1
R. S. The missing memristor found. Nature, 2008, vol. 453, no. 7191,
State variable x p. 80 - 83.
Fig. 7. Illustration of the flexible of the piecewise window func- [5] ITOH, M., CHUA, L. O. Memristor oscillators. International Jour-
tion fPW (x). a = 0.35, b = 20 and the corresponding nal of Bifurcation and Chaos, 2008, vol. 18, no. 11, p. 3183 - 3206.
values of x0 and k are 0.0120 and 1.4190 × 10−4 , respec-
tively. [6] CORINTO, F., ASCOLI, A., GILLI, M. Nonlinear dynamics of
memristor oscillators. IEEE Transactions on Circuits and Systems
Introducing the nonlinear dependance of the state vari- I: Regular Papers, 2011, vol. 58, no. 6, p. 1323 - 1336.
able derivative and the memristance on the current, this [7] BORGHETTI, J., SNIDER, G. S., KUEKES, P. J., YANG, J. J.,
model can be developed to characterize memristive systems. STEWART, D. R., WILLIAMS, R. S. Memristive switches enable
In some special cases, the memristance expression depend- stateful logic operations via material implication. Nature, 2010,
ing on charge may also be obtained. For example, when the vol. 464, no. 7290, p. 873 - 876.
derivative of the state variable is [8] KAVEHEI, O., IQBAL, A., KIM, Y. S., ESHRAGHIAN, K., AL-
SARAWI, S. F., ABBOTT, D. The fourth element: characteristics,
dx
= α fPW L (i(t)) fPW (x) (21) modelling and electromagnetic theory of the memristor. Proceedings
dt of the Royal Society A: Mathematical, Physical and Engineering Sci-
ence, 2010, vol. 466, no. 2120, p. 2175 - 2202.
where fPW L (·) denotes a piecewise linear function, the func-
[9] JOGLEKAR, Y. N., WOLF, S. J. The elusive memristor: proper-
tional relation between x(t) and q(t) can still be derived.
ties of basic electrical circuits. European Journal of Physics, 2009,
Thus, it can yield the memristance of the memristive sys- vol. 30, no. 4, p. 661 - 675.
tems depending on q(t) and i(t). We will make an in depth
[10] BIOLEK, Z., BIOLEK, D., BIOLKOVÁ, V. SPICE model of mem-
study of this issue in future.
ristor with nonlinear dopant drift. Radioengineering, 2009, vol. 18,
no. 2, p. 210 - 214.

[11] PRODROMAKIS, T., PEH, B. P., PAPAVASSILIOU, C., TOUMA-

4. Conclusion ZOU, C. A versatile memristor model with nonlinear dopant kinet-
ics. IEEE Transactions on Electron Devices, 2011, vol. 58, no. 9,
In this paper, a model with piecewise window function p. 3099 - 3105.
is proposed as a proper choice to characterize the memris-
[12] CORINTO, F., ASCOLI, A. A boundary condition-based approach
tor. The window function fPW (x) with three nonlinear pieces to the modeling of memristor nanostructures. IEEE Transactions on
is exploited to depict linear and boundary nonlinear dopant Circuits and Systems I: Regular Papers, 2012, vol. 59, issue. 11,
drifts of memristor. By introducing two parameters a and b, p. 2713 - 2726.
974 J. YU, X. MU, X. XI, S. WANG, A MEMRISTOR MODEL WITH PIECEWISE WINDOW FUNCTION

[13] PICKETT, M. D., STRUKOV, D. B., BORGHETTI, J. L., YANG, J. His current research areas include nonlinear dynamic sys-
J., SNIDER, G. S., STEWART, D. R., WILLIAMS, R. S. Switching tem control, piecewise linear identification and memristor
dynamics in titanium dioxide memristive devices. Journal of Applied
Physics, 2009, vol. 106, no. 7, p. 1 - 6.
device.
Xiangming XI received the B.S. degree in Control Sci-
[14] KVATINSKY, S., FRIEDMAN, E. G., KOLODNY, A., WEISER,
U. C. TEAM: threshold adaptive memristor model. IEEE Transac- ence and Engineering from Harbin Institute of Technology,
tions on Circuits and Systems I: Regular Papers, 2013, vol. 60, no. 1, Harbin, China in 2010. He is currently a Ph.D. candidate
p. 211 - 221. in the Department of Automation, Tsinghua University, Bei-
jing, China. His research areas includes continuous piece-
[15] ASCOLI, A., CORINTO, F., SENGER, V., TETZLAFF, R. Memris-
tor model comparison. IEEE Circuits and Systems Magazine, 2013, wise linear representation models and identification meth-
vol. 13, no. 2, p. 89 - 105. ods, and global optimization algorithms for CPWL program-
ming.
[16] CHUA, L. O. The fourth element. Proceedings of the IEEE, 2012,
vol. 100, no. 6, p. 1920 - 1927. Shuning WANG received the B.S. degree in Electrical En-
gineering from Hunan University, China in 1982, the M.S.
degree and the Ph.D. degree both in the system engineer-
About Authors . . . ing from Huazhong University of Science and Technology,
China in 1984 and 1998, respectively. He was an Associate
Professor from 1992 to 1993 and a Full Professor from 1994
Juntang YU received the B.S. degree in Control Science and
to 1995, at the Institute of Systems Engineering, Huazhong
Engineering from Harbin Institute of Technology, Harbin, University of Science and Technology. He joined Tsinghua
China in 2011. He is currently pursuing the Ph.D. degree University, Beijing, China in 1996. Since then, he has been
from the Department of Automation, Tsinghua University, a Full Professor in Department of Automation, Tsinghua
Beijing, China. His current research interests lie in the
University. He was a Visiting Scholar in the College of
area of modeling and analysis of memristor circuits and sys-
Engineering, University of California at Riverside in 1994,
tems.
and a Visiting Fellow in Department of Electrical Engineer-
Xiaomu MU received the B.S. degree in control science ing, Yale University from 2001 to 2002. His current re-
and engineering from Tsinghua University, Beijing, China search interests are mainly in developing practical methods
in 2009. He is currently a Ph.D. candidate in the Depart- for nonlinear system identification, control and optimization
ment of Automation, Tsinghua University, Beijing, China. via piecewise-linear approximation.