Stefanica 2019 - Volatility

October 15, 2019
A PDE METHOD FOR ESTIMATION OF IMPLIED VOLATILITY
IVAN MATIĆ, RADOŠ RADOIČIĆ, AND DAN STEFANICA
Abstract. In this paper it is proved that the Black-Scholes implied volatility

satisfies a second order non-linear partial differential equation. The obtained
PDE is then used to construct an algorithm for fast and accurate polyno-
mial approximation for Black-Scholes implied volatility that improves on the
existing numerical schemes from literature, both in speed and parallelizabil-
ity. We also show that the method is applicable to other problems, such as
approximation of implied Bachelier volatility.
1. Introduction and formulation of the problem

The price of an asset in the Black-Scholes model [4] is governed by several com-
ponents. Volatility is the parameter that controls the random component of the
dynamics. All other parameters can be obtained from historical data by analysis of
the risk-free interest rate, expected returns, and dividend payments. The market
prices of derived securities such as call and put options uniquely1 determine the
value for volatility for which the Black-Scholes equation is satisfied. This unique
value is called the Black-Scholes market implied volatility, or implied volatility for
short.
After a standard dimension √reduction, fully described in Section 4, implied
volatility is equal to ΣBS (k, c)/ T , where the integrated volatility Σ = ΣBS (k, c)
is a function of log-moneyness k and normalized price c of the underlying asset.2
We will prove that Σ solves the partial differential equations
4
3 2 Σ 2
Σ Σcc = Σc −k , (1)
4
2 2
Σ Σk Σ Σk
Σ3 Σkk = + Σ − kΣk − Σ + kΣk , and (2)
2 2
2
Σk Σ2

Σ
Σ3 Σkc = Σc −k + kΣk − Σ , (3)
2 2
∂ ∂
where subscripts denote partial derivatives Σc = ∂c Σ and Σk = ∂k Σ. Using (1),
(2), and (3) we design an algorithm to approximate ΣBS (k, c) with two-dimensional
polynomials in variables (k, c). Our implementation improves the existing methods
by increasing the speed and accuracy. We further demonstrate how (1–3) can lead
to convexity results for derivatives of Σ. The method described here can also be
applied to the Bachelier model.
1
The Black-Scholes price function is strictly monotone in volatility.
2As defined in Section 4, k = log Ke−rT denotes the log-moneyness, and c = C0
denotes
−qT
S0 e S0 e−qT
the normalized price.
1
Electronic copy available at: https://ssrn.com/abstract=3264356

2 I. MATIĆ, R. RADOIČIĆ, AND D. STEFANICA
2. Practitioner’s Corner
The partial differential equation (1–3) makes it possible to design an algorithm
for approximation of the implied volatility. The algorithm consists of two phases.
The first is “precomputation” phase in which we construct the partition of the
(k, c) plane and the polynomials. The second is “runtime” phase in which the
implied volatility is repeatedly calculated for different choices of parameters from
the market.3
In the precomputation phase we specify the precision ε, degree d for the polyno-
mials, and choose the desired range for parameters k, c, and σ. Small choices for ε
result in finer partitions and larger number of polynomials. Likewise, intervals for
k, c, and σ that include values close to 0 lead to more complex partitions.
We will first present the brief outlines of the two phases of the algorithm. The
full details are developed later in the text.
Algorithm 1 Precomputation Phase

Input
1: ε = desired precision
2: [kmin , kmax ] = domain for the log-moneyness k
3: [cmin , cmax ] = domain for the normalized option price c
4: [σmin , σmax ] = domain for integrated volatility
5: d = required degree for polynomials
Output
1: {κ0 , κ1 , . . . , κm } = partition of the segment [kmin , kmax ]
2: {ci,0 , ci,1 , . . . , ci,maxi } = partitions of the segment [cmin , cmax ] for each i ∈
{0, 1, . . . , m}
3: Pi,j (k, c) = polynomials of two variables of degree d that approximate integrated
volatility ΣBS (k, c) with precision ε on the interval [κi , κi+1 ] × [ci,j , ci,j+1 ] for
each i ∈ {0, 1, 2 . . . , m} and each j ∈ {0, 1, . . . , maxi }.
Code
1: function PrecomputationPhase
2: {κ0 , κ1 , . . . , κm } ← KAxisPartition (ε, d, kmin , kmax )
3: for i ∈ {0, 1, 2, . . . , m} do
4: {ci,0 , ci,1 , . . . , ci,maxi } ← CAxisPartition (ε, d, cmin , cmax , κi )
5: for j ∈ {0, 1, 2, . . . , maxi } do
6: Pi,j ← CreatePolynomial (d, κi , cj )
7: return {κi , ci,j , Pi,j }
The functions KAxisPartition, CAxisPartition, and CreatePolynomial are based

on the partial differential equation (1–3) and they are implemented in Section 7.
We will now outline the data structures that are constructed after Algorithm 1 is
executed. The numbers {κi , ci,j } determine the partition of the region [kmin , kmax ]×
[cmin , cmax ] of the (k, c) plane. A depiction of such partition is shown in Figure 1.
In addition, for each pair (i, j) we have a polynomial
Pi,j : [κi , κi+1 ] × [ci,j , ci,j+1 ] → R
3This structure is identical with the one recently proposed in [18].

A PDE METHOD FOR ESTIMATION OF IMPLIED VOLATILITY 3
that approximates the integrated volatility (k, c) 7→ ΣBS with precision ε.
Figure 1. Partition of the (k, c) plane.
The runtime phase of the algorithm (see Algorithm 2) starts by uploading the
partition and the polynomials that are generated by Algorithm 1. Then we can
repeatedly receive the input parameters from the market and approximate the
integrated volatility by identifying the appropriate polynomial and evaluating it.4
Algorithm 2 Runtime Phase

Input
1: (S0 , K, CBS , r, q, T ) = (asset spot price, strike, market price of the option, in-
terest rate, dividend rate, expiration)
Output
1: σBS = implied volatility
Code
1: function RuntimePhase
2: {κi , ci,j } ← LoadPartition(ε, kmin , kmax , cmin , cmax , σmin , σmax , d)
3: {Pi,j } ←LoadPolynomials(ε, kmin , kmax , cmin , cmax , σmin , σmax , d)
−rT
C0
4: (k, c) ← log SKe 0e
−qT , S e−qT
0
5: if k < kmin then
6: k ← 1.01 · 10−7
7: (i, j) ← the pair of indices such that (k, c) ∈ [κi , κi+1 ] × [ci,j , ci,j+1 ]
8: ΣBS ← Pi,j (k, c)
9: return Σ√BST
Our implementation of the above algorithm in C++ and OpenCL is available as

an open source [39].
4Because of the divergence of our approximation at k = 0 as explained later in the text, in

lines 5 and 6 of Algorithm 2 the small values of k are replaced by 1.01 · 10−7 . The chosen overall
precision error ε = 10−7 is unaffected.

3. Overview of the results from the literature

3.1. Existing methods. Originally suggested by [27], the implied volatility has
been omnipresent in financial industry as it is common practice on trading desks
for the implied volatility rather than the option price to be quoted. This practice
allows option market makers to compare option prices across different asset classes.
As stated in [2, 11], very accurate real-time calculations of implied volatility for
millions of option prices are often required for high-frequency trading. Moreover,
intraday recalibrations of more sophisticated stochastic volatility models that have
been introduced and studied since the Black-Scholes model (see [14], for example)
are typically performed by minimizing the difference of the corresponding implied
volatilities, rather than the difference of model and market prices, across a number
of model parameters. Since the inversion of the Black-Scholes formula cannot be
done in a closed form (see [17]), numerous numerical procedures, as well as analyti-
cal approximations have been devised in the past. First numerical procedures were
based on iterative root finding algorithms such as the bisection, Newton–Raphson,
or Dekker–Brent method. While Newton–Raphson method has quadratic conver-
gence (when it does converge, see [37]), it has pitfalls such as divisions by the option
vegas that can be extremely small for away-from-ATM options [22]. Most commer-
cial software (e.g. Matlab) uses the considerably slower Dekker–Brent algorithm
that is guaranteed to converge. Although these methods work well for single op-
tions, these numerical solvers are too slow for real-time applications. To overcome
this problem, Jäckel [22] exploits the limit behavior of the normalized option price
to provide a smarter initial guess, which reduces the number of iterations in the
Newton–Raphson algorithm, modified according to Halley’s method. In [23], this
approach is further improved using four rational function branches for the initial
guess and the order four convergence Householder method for the iteration. In
spite of being the state-of-the-art numerical solver for implied volatility applicable
to any domain of input variables (see [29]), Jäckel’s solver comes with the burden of
a relatively complex implementation that makes it slower than some other methods
reviewed below.
Non-iterative approximations to implied volatility are a popular alternative to
the above-mentioned iterative solvers, as they provide closed-form expressions that
are fast and easy to implement and maintain, and that can be eventually used
for other gradient-based optimization routines. In the early works of Brenner and
Subrahmanyan [5], Chance [7], Corrado and Miller [10], Li [31], and others (see
[8, 11, 41], for review, discussion, and comparison), as well as in the more recent
work of Lorig et al. [34] (that applies to more general local-stochastic volatility
models), the analytical approach was almost exclusively based on Taylor series ex-
pansions of the call price around ATM (ATM) to a certain order, and as such, it
suffers from a serious lack of precision for options that are more than 10% away
from ATM, yielding these methods impractical. Over the past two decades, there
has been a great amount of focus on obtaining asymptotic expansions of implied
volatilities in extreme regions where numerical schemes typically become less effi-
cient (see, for example, [15] for small maturity, [25] for large-time behavior, and
[30] for extreme strikes). In principle, one can derive an approximation, containing
no special functions and requiring no numerical integration, for the implied volatil-
ity in any stochastic volatility model whose characteristic function is available in
closed form [26], by truncating its expansion with respect to some small parameter.

However, the resulting volatility expansion is accurate only around ATM, and there
is no reason why it should be so in the tails.
With the work of Li [32], economized rational approximations for implied volatil-
ity have entered the picture, obtaining a better fit over a wider moneyness region,
while losing some of the goodness of fit near ATM. Li used a fine mesh of about one
million options to densely populate the 2-dimensional domain of approximation in
log-moneyness and integrated volatility, and subsequently employed the downhill
simplex search method to minimize an objective function, obtaining a single ratio-
nal function of log-moneyness k and normalized price c with about 30 parameters,
that can invert one million options within a few seconds. The domain D[32] covered
by this approximation is fairly large, with k ranging from -0.5 and 0.5, and inte-
grated volatility ranging from 0 to 1, while the uniform error of 10−3 order in the
integrated volatility is claimed numerically.5 For comparison, previously described
Jäckel’s solver [23] achieves the error of 10−14 order in the same domain, for a price
of additional time (slower by roughly a factor of 10), complexity and maintenance
required. One should keep in mind that Li’s approximation is tailored to a spe-
cific fine mesh of options that covers only about 80% of the whole range of traded
options in many cases (see [18]), excluding many options on underlyings with very
low or high volatilities, as well as deep in-the-money (ITM) and out-of-the-money
(OTM) options. Li’s rational approximation was improved on D[32] in [43] and [45].
Pistorius and Stolte [43] use an iterative method of Press et al. [44] (Section 5.13)
to √obtain, for 105 different values of k, (univariate) Chebyshev rational functions
of c with varying degrees (7 to 9), and eventually proceed with cubic splines to
interpolate in between, reducing the maximum absolute error to the order of 10−5 .
Although, much like with [32], all approximations are done once and stored to the
hard-drive, so few calculations are required during the runtime, this improvement in
error comes with increased storage and no significant improvement in speed, while
still struggling where most methods do: with deep ITM/OTM options as well as
options with short maturity. Salazar Celis in [45] works on a Chebyshev grid that
is structured in k-direction, scattered in c-direction, and most importantly denser
in the corners of D[32] , and uses singular value decomposition to solve the highly√
non-linear optimization problem of finding a univariate rational function of c,
with weights that are polynomial in k, hence obtaining a parametrized barycentric
approximation that is (unlike [43]) truly bivariate rational. With no improvement
in speed over [32] reported, the maximum absolute error is of the order of 10−4 on
D[32] , albeit requiring storage of only 70 coefficients.
Glau et al. [18] used Jäckel’s approximation [23] at the gridpoints of a large
1000 × 1000 Chebyshev grid in the (k, c)-space, and a chebfun Matlab package for
interpolation in between, thus obtaining a bivariate Chebyshev polynomial approx-
imation for implied volatility with subexponential rate of convergence. On domain
D[32] it achieves the error of order 10−5 /10−8 /10−11 . Their Matlab implementa-
tions are roughly 4/3/2 times faster (respectively) than the algorithm of Jäckel
[23], and 4/5/6 times slower than the algorithm of Li [32] (whose inferior error is
of order 10−3 on the same domain). However, like [23], and unlike [32], this ap-
proach is applicable on a much wider approximation domain D[18] , with k ranging
from -5 and 5, and integrated volatility ranging from 0 to 6, that completely covers
5It should be noted that one still needs to divide by

√
T to obtain the implied volatility, making
the actual error significantly bigger for short maturities.

the market data. In comparison to Jäckel’s method [23] whose error is of order
10−13 on D[18] , the approximation of [18] has the error of order 10−5 /10−8 /10−10
in roughly 4/3/2 times shorter running time, respectively. It is worth noting that
the Chebyshev method from [18] is the first to use polynomials to approximate
the implied volatility. Their method is similar to ours in that it uses the domain
partitioning in pre-computation phase leaving the evaluation of polynomials for the
runtime phase. There are several caveats with their method, though: their approxi-
mation polynomials have significantly larger degrees than ours, degrees vary across
different domains (unlike ours), and the data transformations before evaluations
of Chebyshev polynomials are computationally expensive as they feature multiple
logarithms, exponential functions, and square roots.
In [33], Li and Lee introduced a new successive over-relaxation (SOR) iterative
method to compute the implied volatility, that is applicable on D[18] , and has a
quadratic order of convergence of Newton-Raphson algorithm and the robustness
of the Dekker-Brent algorithm. With just five iterations, it achieves the uniform
error of order 10−13 in the integrated volatility, and manages to invert one million
options in about 10 seconds. Moreover, its accuracy is reported (see [47]) to be
better than that of [23], quite possibly due to the fact that already the generation
of the initial guess in Jäckel’s algorithm relies on rational cubic interpolations and
transformations that are highly sensitive in terms of the accuracy of the error
function and the inverse of the normal cumulative distribution function (CDF).
Finally, in a series of papers [38, 48, 49], the authors derived a simple closed-form
approximation formula for the implied volatility based on the Pólya-type approxi-
mations of the normal CDF. The relative error of this approximation is uniformly
bounded for options with any moneyness and with arbitrary large or small option
maturities and volatilities, including for long dated options and options on highly
volatile underlying assets. It is reported in [29] to be more robust than the rational
approximation of Li [33]. For ATM options (i.e. when k = 0), the approximation
is even more precise, having relative error of order 10−6 for all options with inte-
grated volatility less than 2. Furthermore, Gatheral and the authors [16] proved it
is actually a lower bound on the implied volatility over a large range of strikes and
option prices. Moreover, combined with the variational principle of Tehranchi [51],
they obtained tight uniform bounds on the implied volatility that can be used as
initial guesses to effectively speed up the bisection method, the most basic iterative
root finding algorithm still used by many practitioners. When the SOR method,
reviewed in the previous paragraph, uses this approximation as an initial guess, it
has an implementation that is faster than Jäckel’s method for most of the input
values [29, 47].
While we are fully aware that comparisons of approximation algorithms and
formulae, described above in some detail, may depend on the performance mea-
sures (e.g. maximum absolute error, relative error, speed, storage, maintenance),
implementations, frameworks, libraries, languages and platforms used, the general
strategy for inverting millions of option prices for Black-Scholes implied volatility
in practice seems to be to choose the SOR method of Li and Lee [33], if one opts
for optimal speed, the solver of Jäckel [23] if one opts for optimal accuracy, or the
closed-form approximation of [49], if one opts for simplicity. Therefore, it is safe to
conclude that the new algorithm for approximating the implied volatility derived
in this paper should be compared to the two state-of-the-art numerical solvers for

implied volatility of [23] and [33] (the latter implemented with the approximation
of [49] as the initial guess), that are available as open source.
3.2. PDE-based method. Our new numerical procedure is relatively simple in

comparison with previous methods. It is concise and analytic, being based on
partial differential equations (PDEs) for the implied volatility and appropriate re-
cursions derived to calculate their power series solutions. Although using power
series to solve differential equations is a classical method that can be traced back
to Frobenius [13] or even Euler (see also [40, 42]), the literature is scarce when
it comes to differential equations satisfied by the implied volatility; the singular
occurrence being the result of Berestycki et al. [3] showing that, in a rather general
class of stochastic volatility models, implied volatility in the small-maturity limit
can be characterized as a viscosity solution to a first order non-linear eikonal PDE.
A method similar to ours was used in [50] for establishing ordinary differential equa-
tions (ODEs) for quantile functions of common probability distributions, which in
turn yields immediate applications in modern methods of Monte–Carlo simulation,
techniques based on low-discrepancy sequences, as well as copula methods. In par-
ticular, the ODE for the quantile of the normal CDF is a special case of the PDE
for the implied volatility in the Black–Scholes model, namely, equation (4) from
[50] turns out to be a special case of equation (1) in Theorem 1 when k = 0. It
is worth noting that a recent result by Xia and Cui [54] provides one-dimensional
power series expansion for the implied volatility for a fixed strike. In fact, their
Theorem 4 after proper transformations is actually equivalent to our power series
(27) in variable c (with k constant).
After obtaining the PDEs for the implied volatility, a bird’s–eye view of our
numerical scheme sees the entire (k, c)-domain partitioned into rectangular cells and
on each cell uses a (two-dimensional) polynomial approximation around its bottom-
left corner, with coefficients obtained recursively from the PDEs.6 Our partition
is specifically tailored to the application at hand, yielding additional accuracy over
methods that use uniform, or even Chebyshev grids. Namely, the precision of the
polynomials depends on the point the expansion is made around and, consequently,
the partition of the domain is driven to be non-uniform, as it can clearly be seen in
Figure 1, where the grid cells are much smaller and abundant for small values of k
and/or when c is close to 0 or 1; these values corresponding to the ATM options, and
deep OTM/ITM options, respectively. The precomputation phase of the algorithm
consists of preparing the partition and the polynomials for given precision ε. After
this preparation is completed, the evaluation of the polynomials is fast for any given
set of inputs. The full details of our implementation are provided in Section 7.
Our procedure for approximating the implied volatility provides a combination
of efficiency, accuracy, and robustness (in a sense that the precision and the domain
of input variables can be adjusted for the application at hand). It covers a large
domain of input variables, as it is able to deal with small and large (absolute) values
of log-moneyness k (that is, deep ITM/OTM options), short times to maturity,
large values of underlying, as well as options with very high/low volatilities. We
will illustrate
our method on the domain where log-strike k and relative price c
satisfy k ∈ 10−6 , 5 and c ∈ 10−6 , 0.997 . The range for the integrated volatility

6The use of local polynomial expansion around many points yields the apparent precision of
our algorithm, and is reminiscent of their use in Hairer’s theory of regularity structures [21].

√
that is covered by the method is σ T ∈ 10−6 , 7.3 . For comparison, the domain

k
D[32] , covered by some of the existing methods, required that σ√ T
is smaller than
2. Our method allows the quantity σ√T to range over the interval 2.6 · 10−6 , 4.9 .
k

For the chosen precision level of 10−7 , it is sufficiently accurate for practical
purposes and highly efficient in terms of absolute error and speed with large data
sets, delivering precise real-time evaluations of implied volatility. For example, the
running time of our implementation to invert one million randomly generated op-
tion prices for implied volatility is about three times shorter than that of [23] and
[33] on a single processor. Our algorithm is easy to implement and maintain, and,
unlike the competing routines of [18, 23, 33, 43, 45], it is completely independent of
more involved numerical recipes, such as splines, rational approximation packages,
singular value decomposition, or transformations/objective functions that involve
the error function or the inverse of the normal CDF. Indeed, in a commercial envi-
ronment, it is of little significance what may be implemented in, say Matlab, while
our implementation can be easily extracted and ported to the respectively relevant
industrial system or analytics library. With maximum absolute error of 10−7 , the
number of calculated constants is 2.5 million in the precomputation phase that
needs to be executed only once. It is also very robust, as both precision and the
domain of input variables can be easily improved/extended, alas for the price of
slower running time and additional memory to store partitions and polynomials.
Another main advantage of our numerical scheme lies in a fewer number of condi-
tional statements and branchings/domain splittings. Although our approximation
diverges for k = 0 (that corresponds to the ATM options), simply replacing k = 0
with k = 1.01 · 10−7 handles the ATM case and maintains the same absolute error
of 10−7 .
The gain in speed can be made to be even more substantial. Our OpenCL im-
plementation on p3.2xlarge instance on Amazon Web Services to invert one million
randomly generated option prices for implied volatility is five times faster than
when the algorithm is executed on an eight-core CPU. Since graphic cards consist
of numerous processing elements that are performing optimally if all pipelines are
equally loaded, numerous branchings/domain splittings in the competing routines
of [23, 33] make them less suitable for parallelization. In addition, note that our
partition of the domain is done with line segments parallel to the coordinate axes
(see Figure 1). This feature allowed us to use a simple indexing algorithm of O(1)
time complexity and avoid look-up bottleneck for parallelization purposes. If one
were to circumvent the coordinate transformations performed before evaluations of
Chebyshev polynomials (see φi,x (c) on pages 10–13 of [18]), the Chebyshev method
would also be amenable to an efficient indexing algorithm of O(1) time complexity.
3.3. Organization of the paper. Section 4 introduces the necessary notation and
defines the inverse problem of finding the implied volatility in the Black–Scholes
model after the standard dimension reduction. Section 5 contains the proof of
equations (1–3), stated as Theorem 1. In Section 6 we use equations (1–3) to
obtain recursive formulas for two-dimensional polynomial approximation for the
implied volatility. The details of the implementation are presented in Section 7.
The speed and accuracy of our method is compared to that of other numerical
schemes in Section 8. Our method of finding a differential equation satisfied by
the implied volatility (after an appropriate dimension reduction) appears to be

general enough to be applicable in models other than the Black–Scholes model. In

Section 9, we derive an efficient and accurate numerical scheme for the calculation
of implied Bachelier volatility that surpasses the rational approximation of [9] and
provides a more transparent, albeit less accurate, alternative to recent machine-
precision approximations of [24, 28]. Since inverting for Bachelier volatility is a one-
dimensional problem, it may be helpful for the reader to first go through Section
9 before delving into more technical implementations for Black–Scholes implied
volatility presented in Sections 6 and 7. Section 10 discusses the drawbacks of our
method and provides examples of how it can be of further use for approximations
and derivations of theoretical results. We conclude the paper with remarks in
Section 11.
4. Implied volatility in Black-Scholes model

In what follows, we adopt the approach and notation of [16] (see also [32, 45]).
Let S0 be the current stock price, σBS the volatility of the stock, r the risk-free
interest rate and q the dividend rate. Let C0 be the price of a plain vanilla European
call option at time 0 with expiration date T and strike price K. Then the Black–
Scholes formula gives C0 in closed form as follows:
Ke−rT √

−qT
C0 (S0 , K, σBS , r, q, T ) = S0 e CBS log , σBS T ,
S0 e−qT
where the function CBS : R × [0, +∞) → [0, 1) is defined as

k σ k k σ
CBS (k, σ) = N − + −e N − − . (4)
σ 2 σ 2
Finding the implied volatility reduces to the following problem
For given c and k find σ that solves CBS (k, σ) = c.
Definition 1. The function ΣBS (k, c) is the inverse of CBS (k, σ). In other words
h +
ΣBS (k, ·) : 1 − ek , 1 → [0, +∞)
is given by
σ = ΣBS (k, c) ⇐⇒ c = CBS (k, σ). (5)
In practice one observes the values C0 , K, r, q, and T , and evaluates
C0
the normalized call option price c = , (6)
S0 e−qT
Ke−rT
the log-moneyness k = log . (7)
S0 e−qT
After obtaining the integrated volatility σ = ΣBS (k, c), the volatility σBS can be
recovered from σ using
σ
σBS = √ . (8)
T
It is known that one merely needs to focus on call prices due to the put-call
parity
CBS (k, σ) = PBS (k, σ) + 1 − ek , (9)
where the normalized put price PBS (k, σ) is defined in analogy to (4).

Our domain of approximation can be restricted to k ≥ 0, due to the following

duality
CBS (−k, σ) = e−k CBS (k, σ) + 1 − e−k , (10)
which implies
ΣBS (−k, c) = ΣBS (k, e−k c + 1 − e−k ). (11)
5. Piecewise approximation of the implied volatility

In this section we often omit the subscript in ΣBS (k, c) and write only Σ(k, c)
instead. For clarity of exposition, we again write equations (1), (2), and (3), now
stated as our main theorem.
Theorem 1. The implied volatility Σ(k, c) satisfies the following three partial dif-
ferential equations
4
3 2 Σ 2
Σ Σcc = Σc −k , (1)
4
2 2
Σ Σk Σ Σk
Σ3 Σkk = + Σ − kΣk − Σ + kΣk , and (2)
2 2
2
Σk Σ2

Σ
Σ3 Σkc = Σc −k + kΣk − Σ . (3)
2 2
Proof. We will start from the relation CBS (k, ΣBS (k, c)) = c that can be written
in an equivalent form

k Σ(k, c) k k Σ(k, c)
c = N − + −e N − − . (12)
Σ(k, c) 2 Σ(k, c) 2
z2
The derivative of N (z) satisfies N 0 (z) = √12π e− 2 . Therefore we obtain the
following two relations

0 k y k 0 k y
N − + = e N − − (13)
y 2 y 2
00 0
N (z) = −zN (z). (14)
Taking partial derivatives with respect to c of both sides of (12) gives us

k Σ(k, c) ∂ k Σ(k, c)
1 = ek N 0 − − − +
Σ(k, c) 2 ∂c Σ(k, c) 2

∂ k Σ(k, c)
− − − .
∂c Σ(k, c) 2
After simplifying the right-hand side we obtain

0 k Σ(k, c)
1 = N − + · Σc (k, c). (15)
Σ(k, c) 2
The last equation can be re-written as
√ Σ(k,c) 2
2πe 2 (− Σ(k,c) + 2 ) .
1 k
Σc (k, c) = (16)
Once we differentiate both sides of (15) with respect to c and use (14) we obtain
equation (1).

We can obtain the derivative Σk by differentiating both sides of (12) with respect
to k. Using (13) and (14) we obtain

k
N − Σ(k,c) − Σ(k,c)
2
Σk (k, c) = . (17)
k Σ(k,c)
N 0 − Σ(k,c) − 2
We can re-write (17) as

0 k Σ(k, c) k Σ(k, c)
N − − Σk (k, c) = N − − . (18)
Σ(k, c) 2 Σ(k, c) 2
Differentiating both sides with respect to k and using (14) we obtain (2).
Once we take the derivative of both sides of (15) with respect to k we obtain

00 k Σ(k, c) ∂ k Σ(k, c)
0 = N − + · Σc (k, c) · − +
Σ(k, c) 2 ∂k Σ(k, c) 2

k Σ(k, c)
+N 0 − + · Σkc (k, c).
Σ(k, c) 2
Using (14) we can further simplify the last equation and obtain (3).

6. Power series representation

In this section we will do the same as in the previous one – omit the subscript
in ΣBS (k, c) and work with Σ(k, c). We will use equations (1–3) to obtain the
recursive relations allowing us to determine the two dimensional sequence αm,n =
αm,n (k0 , c0 ) of real numbers such that
∞
X m n
Σ(k, c) = αm,n (c − c0 ) (k − k0 ) . (19)
m,n=0
∞
We will first use (1) to construct a recursion for the sequence (αm,0 )m=0 . Then
∞
we will use (2) to create a recursion for (α0,n )n=0 . After finding the sequences
∞ ∞
(αm,0 )m=0 and (α0,n )n=0 we will use (3) to get a recursive relation for αu,v in
terms of αu0 ,v0 with u0 ≤ u and v 0 ≤ v. In other words, we will prove that αu,v is
a polynomial function of variables αm,n whose indices are pairs of numbers (m, n)
that belong to the set
Su,v = {(u0 , v 0 ) : 0 ≤ u0 ≤ u, 0 ≤ v 0 ≤ v, (u0 , v 0 ) 6= (u, v)} . (20)
Let us first observe that the following identities hold for partial derivatives of
Σ(k, c)
∞
X m n
Σc (k, c) = (m + 1)αm+1,n (c − c0 ) (k − k0 ) , (21)
m,n=0
X∞
m n
Σk (k, c) = (n + 1)αm,n+1 (c − c0 ) (k − k0 ) , (22)
m,n=0
X∞
m n
Σkc (k, c) = (m + 1)(n + 1)αm+1,n+1 (c − c0 ) (k − k0 ) . (23)
m,n=0

6.1. Calculation of αm,0 . Assume that we are given c0 , Σ0 , and k0 such that
Σ0 = ΣBS (k0 , c0 ). In this section the value k = k0 is fixed and the function
c 7→ Σ (k0 , c) depends only on one variable c. The coefficients in power series
expansion of such function will form the sequence (αm,0 ). In order to simplify
notation and emphasize that we are dealing with functions of a single variable we
will use αC to label this sequence. In other words, we define
C
αm = αm,0 .
The first term satisfies α0C = Σ0 . Together with (16) and (1) we obtain the
C ∞

equations for the first three terms of the sequence αm m=0
α0C = Σ0 , (24)
√ 1

k0 Σ0 2

2 − Σ0 + 2
α1C = 2π · e , (25)
C 2 C 4
!
1 α1 α0
α2C = · · − k02 . (26)
2 αC 3 4
0
Placing k = k0 in equation (19) and differentiating twice implies

∞
X
C m
Σ (k0 , c) = αm (c − c0 ) (27)
m=0
X∞
C m
Σc (k0 , c) = (m + 1)αm+1 (c − c0 ) (28)
m=0
X∞
C m
Σcc (k0 , c) = (m + 2)(m + 1)αm+2 (c − c0 ) . (29)
m=0
CD ∞ CDD ∞

We will consider the sequences αm m=0
and αm m=0
that correspond to the
expansions of the first and second derivatives of Σ (k0 , ·) in (28) and (29). The
sequences are defined as
CD C
αm = (m + 1)αm+1 , (30)
CDD C
αm = (m + 2)(m + 1)αm+2 . (31)
Assume now that u ≥ 2 is fixed and that we know all of the terms αiC for
i ∈ {0, 1, 2, . . . , u}. The goal is to find a procedure to calculate αu+1 . The idea is
u−1
to compare the coefficients of (c − c0 ) on both sides of (1). Let us recall the
∞
definition of the discrete convolution of sequences. For two sequences (µm )m=0 and
∞ ∞
(νm )m=0 their convolution µ ∗ ν is defined as the sequence (µ ∗ ν)m=0 that satisfies
m
X
(µ ∗ ν)m = µi νm−i .
i=0
P∞ i
If the sequences µ and ν are generating the power series M (x) = i=0 µi (x − x0 )
P∞ i
and N (x) = i=0 νi (x − x0 ) then µ ∗ ν is generating the power series of the
product M (x) · N (x).
Let us denote by LC u−1 the coefficient on the left-hand side of (1) corresponding
u−1
to (c − c0 ) . Since the power series is the product of four power series Σ·Σ·Σ·Σcc

the sequence of coefficients LC is the convolution of four sequences. Three of them

are equal to αC while the fourth sequence is αCDD . In other words,
LC C C C CDD

u−1 = α ∗ α ∗ α ∗ α u−1
. (32)
The convolution is commutative and associative and there is no need for brackets
to specify the order of operations.
Only one term in LC C
C depends on αu+1 , while all other summands depend on
u−1
elements from the set αi : 0 ≤ i ≤ u . According to (31) this particular term in
LC C
u−1 that depends on αu+1 is obtained by multiplying lowest order terms from the
three sequences α and the highest order term from the sequence αCDD . Hence,
C
it must be equal to
3 CDD
α0C · αu−1 = Σ30 · (u + 1) · u · αu+1
C
.
Let us denote by LC,lower

u−1 the portion of the summation in LC
u−1 that is obtained
using only elements of the sequence αC whose indices are smaller than or equal to
u. Using this notation we can write
LC
u−1 = Σ30 · (u + 1) · u · αu+1
C
+ LC,lower
u−1 .
u−1
The coefficient of (c − c0 ) on the right-hand side of (1) satisfies

C 1 C
Ru−1 = αCD ∗ αCD ∗ α ∗ αC ∗ αC ∗ αC − k02 .
4 u−1
C

Notice that Ru−1 is calculated using only the terms from the set αiC : 0 ≤ i ≤ u .
Therefore we obtain the recursion
C
C
Ru−1 − LC,lower
u−1
αu+1 = . (33)
(u + 1) · u · Σ30
6.2. Calculation of α0,n . Assume that we are given c0 , Σ0 , and k0 such that Σ0 =
ΣBS (k0 , c0 ). We will find the expansion of the single-variable function k 7→ Σ(k, c0 )
around k0 . Let us denote αnK = α0,n . We proceed in a similar way as in Subsection
6.1. The equation (19) with c = c0 becomes
∞
X n
Σ (k, c0 ) = αnK (k − k0 ) (34)
n=0
X∞
K n
Σk (k, c0 ) = (n + 1)αn+1 (k − k0 ) (35)
n=0
X∞
K n
Σkk (k, c0 ) = (n + 2)(n + 1)αn+2 (k − k0 ) . (36)
n=0
One has to take care of the appearance of the variable k on the right-hand side
of (2) which is no longer a constant. We re-write (2) as
2
3 Σ Σk
Σ Σkk = + Σ − k0 Σk − (k − k0 ) Σk
2
2
Σ Σk
· − Σ + k0 Σk + (k − k0 ) Σk . (37)
2

As in Subsection 6.1 we introduce the sequences αnKD = (n + 1)αn+1 K

and αnKDD =
K
(n+2)(n+1)αn+2 . These two sequences are generating the power series expansions
for the functions Σk (·, c0 ) and Σkk (·, c0 ), respectively. We also need the sequence
that generates the power series for k 7→ (k − k0 ) Σk (k, c0 ). This sequence is ob-
tained as a shift to the right of the sequence αKD . We will denote it by αKD→ .
The terms of this sequence must satisfy

KD→ 0, if n = 0
αn = KD (38)
αn−1 , if n > 0.
The formulas for the initial members α0K , α1K , and α2K are calculated using (17)
and (2). They are given by
α0K = Σ0 , (39)

N − Σk00 − Σ20
α1K = , (40)
N 0 − Σk00 − Σ20
2 K 2 K
Σ0 α1 K Σ0 α1 K
2 + Σ 0 − k 0 α1 · 2 − Σ 0 + k0 α1
α2K = . (41)
2Σ30
Having calculated α0K , α1K , . . . , αvK for some v ≥ 2 we obtain
K
K
Rv−1 − LK,lower
v−1
αv+1 = , where (42)
v(v + 1)Σ30

K 1 K
Rv−1 = α ∗ αK ∗ αKD + αK − k0 αKD − αKD→
2

1 K
∗ α ∗ αK ∗ αKD − αK + k0 αKD + αKD→ , (43)
2 v−1
LK,lower αK ∗ αK ∗ αK ∗ αKDD v−1 − Σ30 · (v + 1) · v · αv+1

K

v−1 = . (44)
K
Here Rv−1 and LK,lower
v−1
C
are analogous to the corresponding Ru−1 and LC,lower
u−1 in
Subsection 6.1, and a reasoning similar to the one in the paragraph after (32)
K K,lower
implies
K that Rv−1
and Lv−1 are calculated using only the terms from the set
αi : 0 ≤ i ≤ v .
6.3. Calculation of αm,n . As before, we assume that we are given k0 , c0 , and

Σ0 that already satisfy Σ0 = ΣBS (k0 , c0 ). We will first recall the basics of two-
∞ ∞
dimensional convolution. Assume that (µm,n )m,n=0 and (νm,n )m,n=0 are two two-
dimensional sequences. Their convolution η = µ ∗ ν is defined as a new two-
∞
dimensional sequence (ηm,n )m,n=0 in the following way
X
ηm,n = µi,p νj,q , (45)
i+j=m,p+q=n
where the indices i, j, p, and q are non-negative integers. If the functions A and B
have power series representations
X m n
A(x, y) = µm,n (x − x0 ) (y − y0 ) and
X m n
B(x, y) = νm,n (x − x0 ) (y − y0 ) ,

then their product satisfies

∞
X m n
A(x, y)B(x, y) = (µ ∗ ν)m,n (x − x0 ) (y − y0 ) .
m,n=0
Equation (3) contains a variable k. Unlike k0 , the quantity k is not a constant,

so we will re-write the equation in the form
Σ2

3
Σ Σkc = Σc −k0 − (k − k0 ) + · (46)
2
Σk Σ2

· −Σ + k0 Σk + (k − k0 ) Σk + .
2
Let us define the sequences α∂C , α∂K , α∂C∂K , α∂K→ to correspond to the func-
tions Σc , Σk , Σck , and (k − k0 ) Σk , respectively. As it can be seen from the equa-
tions (21), (22), and (23), the formal definitions of these sequences can be written
as
∂C
αm,n = (m + 1)αm+1,n ,
∂K
αm,n = (n + 1)αm,n+1 ,
∂C∂K
αm,n = (m + 1)(n + 1)αm+1,n+1 ,

∂K→ 0, if n = 0
αm,n = ∂K
αm,n−1 , if n > 0.
Assume that u ≥ 1 and v ≥ 1 and that we know all the terms of αm,n for which
0 ≤ m ≤ u, 0 ≤ n ≤ v and (m, n) 6= (u, v). In (20) we introduced the notation Su,v
for the set of such pairs of integers (m, n). We will express αu,v in terms of αu0 ,v0
for (u0 , v 0 ) ∈ Su,v which are previously calculated. We do this by comparing the
u−1 v−1
coefficients of (c − c0 ) (k − k0 ) on both sides of (46).
∞ ∞ ∞
Let us introduce the sequences (βm,n )m,n=0 , (γm,n )m,n=0 , and (δm,n )m,n=0 for
which the following hold
∞
Σ2 X m n
−k0 − (k − k0 ) + = βm,n (c − c0 ) (k − k0 ) , (47)
2 m,n=0
X∞
m n
Σ3 = γm,n (c − c0 ) (k − k0 ) , (48)
m,n=0
2 ∞
Σk Σ X m n
−Σ + k0 Σk + (k − k0 )Σk + = δm,n (c − c0 ) (k − k0 ) . (49)
2 m,n=0
We have that

Σ20
2 − k0 , if m = n = 0,


βm,n = 1 (50)
2 (α ∗ α)0,1 − 1, if m = 0, n = 1,
1
2 (α ∗ α)m,n ,

 otherwise,
γ = α ∗ α ∗ α, (51)
1
α ∗ α ∗ α∂K + α∂K→ + k0 α∂K − α.

δ = (52)
2

Let us denote by Lu−1,v−1 and Ru−1,v−1 the coefficients on the left and right-
u−1 v−1
hand side of (46) that correspond to the term (c − c0 ) (k − k0 ) . These coeffi-
cients must be equal. Since both sides of equations are products of two-dimensional
polynomials the coefficients Lu−1,v−1 and Ru−1,v−1 can be obtained using convo-
lutions. The coefficient Ru−1,v−1 satisfies
Ru−1,v−1 = α∂C ∗ β ∗ δ u−1,v−1 ,

(53)
hence it can be calculated using only the terms αu0 ,v0 for (u0 , v 0 ) ∈ Su,v . Similarly,
Lu−1,v−1 satisfies
Lu−1,v−1 = α∂C∂K ∗ γ u−1,v−1 .

(54)
∂C∂K 3
One of the terms in Lu−1,v−1 is γ0,0 · αu−1,v−1 = α0,0 · uv · αu,v while the remaining
terms can be calculated using only the variables αu0 ,v0 for (u0 , v 0 ) ∈ Su,v . Let us
denote by Lloweru−1,v−1 the portion of Lu−1,v−1 that is expressed in terms of αu0 ,v 0 for
(u0 , v 0 ) ∈ Su,v . In other words,
Llower 3
u−1,v−1 = Lu−1,v−1 − α0,0 · uv · αu,v . (55)
Then we have
Ru−1,v−1 − Llower
u−1,v−1
αu,v = . (56)
Σ30 · u · v
The last equation gives us the recursive procedure for calculating the terms of the
∞
sequence α. The initial data are the one-dimensional sequences (αm,0 )m=0 and
∞
(α0,n )n=0 that are calculated in Subsections 6.1 and 6.2.
7. Implementation
7.1. Partition of the interval for k. First, the program partitions the k-axis
using the code in Algorithm 3, with numbers κ0 , κ1 , . . . , κm , as a part of precom-
putation phase. Higher precision results in larger values of m and finer partitions
−7
of the k-axis. In our implementation, where the required precision is ε = 10 , the
−6 −6
degree of the polynomials is 8, and (k, c) ∈ 10 , 5 × 10 , 0.997 , the number m
turns out to be 149. In order to build the partition of the k-axis we need to work
with one-dimensional polynomials in which the c coordinate is fixed. We can choose
an arbitrary c? , and in our case we had c? = 0.017. Then we select the smallest
k for which we want to approximate the implied volatility. This is the given left
endpoint kmin , hence we make the choice κ0 = kmin . Then we build the first poly-
nomial approximation P0 in variable k around (κ0 , c? ) using the PDE method as
shown in Subsection 6.2. Then we determine κ1 as a number for which the error
of the approximation P0 (κ1 , c? ) is ε/40. After this we build our second polynomial
P1 around (κ1 , c? ) and obtain the point κ2 as the one in which the approximation
error for P1 (κ2 , c? ) is ε/40. The procedure continues until we reach kmax .
Thus, we partitioned the interval [kmin , kmax ] with the sequence → −κ = (κ0 , κ1 ,
. . . , κm ) and constructed one-dimensional polynomials k 7→ Pi (k, c? ) for i ∈
{1, 2, . . . , m − 1} such that
• Pi (k, c? ) is a polynomial of degree d;
ε
• |Pi (k, c? ) − ΣBS (k, c? )| ≤ 40 for k ∈ [κi , κi+1 ].
In this step we are calling the bisection method to calculate ΣBS . Bisection
method is only used in precomputation phase.

Algorithm 3 Partition of the k axis

Input
2: [kmin , kmax ] = domain for the log-moneyness k
Output
Code
1: function KAxisPartition
2: c? ← 0.017
3: i←0
4: κ0 ← kmin
5: while κi < kmax do
−→ d
6: αK = αnK n=0 ← Recursion (42) with initial conditions (39), (40),
and (41). The center of the expansion k0 is replaced by κi while
Σ0 = ΣBS (κi , c? ) is obtained using bisection method.
Pd n
7: P (k, c? ) ← n=0 αnK · (k − κi )
ε
8: κi+1 ← Maximal k for which |P (k, c? )) − ΣBS (k, c? )| < 40
9: i←i+1
10: m←i
return the sequence → − m
11: κ = (κi )i=0
7.2. Partitions of the interval for c. In the previous step we have created the
partition → −
κ . The next task is to create the partitions of the interval for c. The
summary is shown in Algorithm 4.
The goal is to iterate over each value i ∈ {0, 1, 2, . . . , m}, fix κi , and divide
[cmin , cmax ] into subintervals on which the polynomials of degree d in c-variable
provide an approximation to ΣBS (κi , ·) with an error smaller than ε/40. This
task is performed by induction. We start with σi0 = σmin (in our case we chose
σmin = 10−6 ), construct ci0 = CBS (κi , σi0 ) and then find a polynomial Qi0 of
degree d around (κi , ci0 ) as described in Subsection 6.1. We then start with value
0
σi1 = σi0 + 1.6, calculate c0i1 = CBS (κi , σi1 0
) and check whether Qi0 (κi , c0i1 ) is
within ε/40 from σi1 . If it is, then we select ci1 = c0i1 , otherwise we repeat the
0
00
procedure with σi1 = σi0 + 0.8. Eventually, we arrive to ci1 and repeat algorithm
to get ci2 , ci3 , . . . until we reach σmax .
Observe that the number of cells along the c-axis is not the same for each κi ,
rather it varies significantly and takes values from 119 to 298. When ε = 10−7 the
total number of rectangular cells in the partition of D is 29474.
Since this is a part of the algorithm that is executed only once, it is not essential
to fully optimize it and perform it in parallel. However, it is possible to perform the
treatment of different values of κi in parallel. We saw a small gain in performance
by doing so.
7.3. Creation of polynomials for each cell in partition. For each κi and each
ci,j we now create a two-dimensional polynomial Pi,j around (κi , ci,j ) as described
in Subsection 6.3 and outlined in Algorithm 5. This task also belongs to the pre-
computation phase of the algorithm.

Algorithm 4 Partition of the c axis

Input
3: [kmin , kmax ] = domain for the normalized option price c
4: [σmin , σmax ] = domain for integrated volatility
Output
1: {ci,0 , ci,1 , . . . , ci,maxi } = partitions of the segment [cmin , cmax ] for each i ∈
{0, 1, . . . , m}
Code
1: function CAxisPartition
2: for i ∈ {0, 1, . . . , m} do
3: σnext ← σmin , j ← 0, ci,0 ← CBS (κi , σ)
4: while ci,j < cmax and σ < σmax do
−→ d
5: αC = αC n=0 ← Recursion (33) where initial conditions are (24),
(25), and (26). The quantities Σ0 , k0 , and c0 are replaced by σnext ,
κi , and ci.j , respectively.
Pd C m
6: P (κi , c) ← m=0 αm (c − ci,j )
7: σnext ← σ + 1.6, cnext ← CBS (ki , σnext )
ε
8: while |P (κi , cnext ) − σnext | > 40 do
σ+σnext
9: σnext ← 2 , cnext ← C BS (κi , σnext )
10: ci,j+1 ← cnext , j ← j + 1
11: maxi ← j
return the sequences → − max
12: ci = (ci,j )j=0 i for i ∈ {0, 1, . . . , m}
7.4. Evaluation of implied volatility. This is the runtime phase of the algo-
rithm. At this point we assume that the partitions and polynomials are created
and pre-loaded in the memory. For a given pair (k, c) we first need to find the
indices i and j that determine the polynomial Pi,j (k, c) to be evaluated.
This task can be done in O(1) time and without any branching. We first locate
the index i ∈ {0, 1, 2, . . . , m − 1} such that k ∈ (κi , κi+1 ) and for that particular i
we locate the index j and the interval (ci,j , ci,j+1 ) to which c belongs.
We explain the procedure to find the index i (same applies to j). Rather than
providing a formal description, we proceed with an example. Assume that interval
A = [0, 100) is split into three sets
A0 = [0, 2.8), A1 = [2.8, 33.5), and A2 = [33.5, 100].
For every x ∈ [0, 100) we need to find the index i(x) ∈ {0, 1, 2} such that x ∈ Di(x) .
We define the sequence (I` )999
`=0 of 1000 integers in the following way:
I0 = · · · = I27 = 0, I28 = · · · = I334 = 1, I335 = · · · = I999 = 2.

Given x ∈ [0, 100), we first calculate the integral part b10xc of 10x. The value
i(x) is defined as i(x) = Ib10xc .

Algorithm 5 Creation of the polynomial

Input
1: (c0 , k0 ) = the center of the expansion
2: d = required degree
Output
d
1: (αm,n )m,n=0 = coefficients of the polynomial
Code
1: function CreatePolynomial
2: Σ0 ← ΣBS (k0 , c0 ). The calculation is performed using bisection method.
−→ d
3: αK = αnK n=0 ← Recursion (42) with initial conditions (39), (40),
and (41).
−→ d
4: C
α = αC n=0 ← Recursion (33) where initial conditions are (24),
(25), and (26).
5: for i ∈ {0, 1, . . . , d} do
6: αi,0 ← αiC
7: α0,i ← αiK
d
8: (αu,v )u,v=1 ← Recursion (56).
d
9: return the sequence (αm,n )m,n=0
7.5. Approximation to volatility surface. For our chosen level of precision

ε = 10−7 the actual approximation surface is quite smooth and virtually indistin-
guishable from the true ΣBS (k, c). In order to demonstrate its rough piecewise
structure, we implemented our algorithm with precision ε = 10−1 and polynomi-
als of degree 5. Figure 2 shows five cross sections corresponding to the values
k ∈ {0.1, 0.3, 0.5, 0.7, 0.9}. Notice that different values for k generate different in-
terval partitions on the c-axis, as already depicted in Figure 1.
8. Analysis of performance
The precomputation phase of the algorithm consists of creating the partition and
polynomials for a given precision ε and domain D for pairs (k, c). This step has
to be performed only once and the results
can be stored
for future use. When the
choices are ε = 10−7 and D = 10−6 , 5 × 10−6 , 0.997 it takes less than a minute
to create the partition and polynomials. Once the set-up phase is completed, the
evaluation of implied volatility is faster than when using other methods, as it can
be seen in the table below.
Method Time (in seconds) Domain Error

Jäckel [23] 0.13498 D 5.35 · 10−13
Li SOR [33] 0.16054 D 3.2 · 10−6
PDE on CPU 0.05276 D 10−7
PDE on GPU 0.00925 D 10−7
The tests consist of one million randomly generated option prices and were run
on Amazon Web Services using an instance of the type p3.2xlarge. The instance
is powered by an Intel Xeon E5-2686 v4 (Broadwell) processor with 8 virtual cores
and has an NVIDIA Tesla V100 GPU with 5120 CUDA Cores. The first three rows

Figure 2. Graphs of c 7→ Σ(k, c) for k ∈ {0.1, 0.3, 0.5, 0.7, 0.9}

and their approximations with tolerance 10−1 .
of the table correspond to the programs that were executed on CPU. All 8 cores
were utilized using the openMP framework. As the table shows, the PDE method
implemented on the graphic card with openCL framework further speeds up our
execution by a factor of 5.
Keep in mind that our power series expansions are divergent for k = 0. Thus,
the implied volatilities cannot be calculated directly for ATM options. However,
our algorithm deals with this case by simply replacing k = 0 with k = 1.01 · 10−7 .
The obtained implied volatility is still within ε = 10−7 from its true value.
There are alternative methods available for the ATM case that can be combined
with our PDE method while still maintaining a small overall absolute error. One
such method that could show promising performance is to use the expansion of
Grunspan (Proposition 6 from [19]). Another approach is discussed in [38]. A
simple closed-form approximation function ΣSR (k, c) for ΣBS (k, c) was derived in
[49], based on the Pólya approximation function Ñ (·) of the normal CDF N (·).
The function ΣSR (k, c) is cheap to compute, and moreover it was shown in [16]
that it is a lower bound for ΣBS (k, c) for most pairs (k, c) that are relevant in
practice. The approximation ΣSR is actually very accurate for small values of c
which makes it convenient to combine with our PDE method whose cells are very
dense and numerous for such c. More specifically, we tried using ΣSR instead of
our PDE method when k < 0.0016 and c < 0.063. The approximation happens to
be 20% faster and the space for storing the partition and polynomials is reduced by

the factor of 3. However, the error of such approximation turns out to be 5 · 10−5 .
Tests and results reported in the table above do not use any of these methods.
Our method also yields an improved repricing error, which was used as one of
the benchmarks in [45] and is defined as
εr = max |CBS (k, Σ(k, c)) − c| ,
(k,c)
where CBS is defined in (4). The corresponding repricing error for the PDE method
with parameters given above is 10−6 . It is better than 10−4 error achieved by [32]
and 10−5 error achieved by [45]. Note that the repricing error of 10−4 may be
substantial once it is multiplied by an underlying of high magnitude.
The algorithm was also tested on a data sample of options provided to us by
Wharton Research Data Services. We chose February 18, 2016 as an arbitrary day
in history for which option data was available. We pruned the data for meaningful
quotes that satisfy put-call parity, non-zero bid-ask spreads and a reasonable trading
volume. The corresponding re-scaled values
k and c for all
of the non-ATM options
were within the domain D = 10−6 , 5 × 10−6 , 0.997 , hence the PDE method

achieved the precision 10−7 . In fact the smallest c that occurred in the dataset was
2 · 10−5 . On this real-world data our PDE method was about three times faster
than other methods. As in [45] we did not observe any bid-ask range violations once
our implied volatility approximation is plugged into the Black-Scholes call option
price formula.
9. Implied volatility in Bachelier model

9.1. Introduction. Bachelier [1] created a model for prices of securities decades
before Black-Scholes. The underlying dynamics follows arithmetic Brownian motion
with volatility σBac .
dSt = σBac dWt . (57)
The price of the call option with strike K is given by
√
C = (S0 − K) N (d) + σ T · N 0 (d) where (58)
S0 − K
d = √ . (59)
σBac T
We only consider call options here, as put-call parity for Bachelier model makes it
easy to relate the two. For the sake of simplicity we assume that the stock pays no
dividends and that the interest rate is zero. The general case can be handled in a
very similar way.
The volatility parameter σBac is often called Bachelier volatility, normal volatil-
ity, or basis point volatility. One characteristic of this model is that the underlying
asset can get a negative value. Nevertheless, it is still appropriate for short matu-
rities and interest rate derivative markets [28]. An interesting comparison between
Bachelier and Black-Scholes model is done by Schachermayer and Teichmann [46]
who showed that for short maturities the former offers a good approximation for
the latter. Choi, Kim, and Kwak [9] nicely summarized several quantities that are
commonly expressed in terms of the Bachelier implied volatility: spread options,
“break-even move” of delta-hedge portfolio, and swaption prices where skewness
gets reduced when the Bachelier volatility is used. The Bachelier implied volatility
is related to the incomplete gamma function, as observed by Grunspan [20].

Recently Le Floc’h [28] and Jäckel [24] provide almost machine precision approxi-
mation for the Bachelier implied volatility based on the rational approximation from
[9].
We will illustrate how the idea of our PDE method can be used to obtain very
precise approximations to the Bachelier implied volatility. To reduce the problem
to one dimension, we divide both sides of (58) by S0 − K. Substituting c = S0C−K
gives us the equation c = CBac (d), where7
N 0 (d)
CBac (d) = N (d) + . (60)
d
Therefore we can formulate the problem of finding the implied volatility in Bachelier
model as
For given c find a positive real number d that solves CBac (d) = c. (61)
Once d is found, then σBac can be recovered from (59).
The problem (61) is equivalent to finding the function ψ(x) such that
N 0 (ψ(x))
x = N (ψ(x)) + . (62)
ψ(x)
Notice that x > 1 and that the right-hand side of (62), as a function of ψ, is decreas-
ing and converges to 1 as ψ → +∞. The implicit function ψ depends on only one
variable, see Figure 3. Recall that in Black-Scholes model, the integrated volatility
Σ was a function of two variables k and c. Therefore instead of constructing PDEs
we will be able to create an approximation using just ODEs.
Figure 3. Graph of function ψ(x) for 1 < x ≤ 8

7From this point on we will assume that d > 0 which corresponds to ITM call options. The
OTM case is treated by “in-out-duality” [33] when one replaces d by −d in (60) and instead of
c= SC−K
C
defines c = K−S + 1.
0 0

It is straightforward to derive the following identity from (14)

0
[N 0 (ψ(x))] = −N 0 (ψ(x)) ψ 0 (x)ψ(x). (63)
Using (63) and taking derivatives of both sides of (62) implies
−N 0 (ψ(x)) ψ 0 (x)ψ 2 (x) − N 0 (ψ(x)) ψ 0 (x)
1 = N 0 (ψ(x)) ψ 0 (x) + . (64)
ψ 2 (x)
The last equation is equivalent to
N 0 (ψ(x)) ψ 0 (x)
−1 = . (65)
ψ 2 (x)
Taking derivative of both sides of (65) gives us
1 nh
2
i
2
0 = 4
· −N 0 (ψ(x)) ψ(x) (ψ 0 (x)) + N 0 (ψ(x)) ψ 00 (x) · (ψ(x))
ψ (x)
o
2
−2N 0 (ψ(x)) ψ(x) (ψ 0 (x)) .
Multiplying both sides by ψ 3 (x) and dividing by N 0 (ψ(x)) gives us

2
2 + ψ 2 (x) · (ψ 0 (x)) = ψ 00 (x) · ψ(x).

(66)
Notice that equations (65) and (66) allow us to immediately conclude that ψ is
a decreasing and convex function.
9.2. Power series representation of ψ. Assume that x0 and ψ0 are fixed and
∞
that they satisfy ψ0 = ψ (x0 ).8 Let us denote by (αk )k=0 the real numbers such
that
∞
X k
ψ(x) = αk (x − x0 ) . (67)
k=0
Then we have α0 = ψ0 . We place ψ0 instead of ψ(x) in (65), solve for ψ 0 (x) and
∞
call α1 the obtained solution. The remaining terms of the sequence (αk )k=0 are
determined recursively.
Let us denote by Pm the set of all pairs of non-negative integers that sum up to
m. In other words
(p1 , p2 ) ∈ N20 : p1 + p2 = m .

Pm = (68)
∞ ∞
Define the sequences (βk )k=0 and (γk )k=0 as
( P
αp1 αp2 , if m ≥ 1,
βm = Pm (69)
α02 + 2, if m = 0;
X
γm = (p1 + 1)(p2 + 1)αp1 +1 αp2 +1 . (70)
Pm
8One way to get this pair (x , ψ ) is to start from x and use bisection method to solve x =
0 0 0 0
CBac (ψ0 ) for ψ0 . Alternatively, one can start from ψ0 and evaluate x0 explicitly as CBac (ψ0 ).

∞
From equation (66) we obtain the recursion for the sequence (αk )k=0 :

m
1 X
αm+2 =  βj γm−j (71)
(m + 2)(m + 1)α0 j=0
m−1
!
X
− (p + 1)(p + 2)αp+2 αm−p .
p=0
9.3. Simulations. We analyze the performance of the algorithm when the desired
error is ε = 10−6 with polynomials of degree 10. When restricted to x ∈ [1.04, 8],
we need 16 polynomials to achieve the error 10−6 , each of them providing an ap-
proximation on a sub-interval of [1.04, 8]. The endpoints of the subintervals are
1.037649, 1.050247, 1.067051, 1.089598, 1.120133, 1.162058, 1.22078,

1.30545, 1.432885, 1.637783, 2.005211, 2.317769, 2.814489, 3.71644,
4.509353, 5.834184, 8.488817.
Observe that the intervals of the partition are denser near x = 1 which is remi-
niscent of the behavior of the two-dimensional partition of the (k, c) plane in Figure
1. The error of the above approximation is shown in Figure 4.
Figure 4. Error for Bachelier approximation with polynomials of

degree 10 with 16 intervals.
10. Further directions

10.1. Drawbacks of the PDE method. Our choices for parameters c? , ε, thresh-
old ε/40 (in Algorithms 3 and 4) and the degree of polynomials used in approxima-
tions were obtained by trial and error in our experiments. We are well aware that

there is a possibility that more optimal choices could have been made. However,
that will likely require more time and space complexity.
Our implementation uses polynomials of degree 8 and achieves the 10−7 preci-
sion. The polynomials of lower degree are easier to handle by the graphic card, but
they require more partition cells. This results in bigger data structures that have
to be stored in memory. The polynomials of higher degree require more computa-
tion but smaller files. We have observed that degree 8 polynomials offer the best
compromise and result in fastest execution. In the near future memory capacities
of graphic cards may allow more efficient implementations and better precision.
The power series (19) is divergent at k = 0 which corresponds to ATM options.
We overcame this difficulty by replacing k = 0 with k = 1.01 · 10−7 and maintained
the same absolute error. It remains to be seen whether there are superior ways of
handling this case.
The existence of the solution to the system of partial differential equations (1),
(2), and (3) is established in Theorem 1. This solution is the implied volatility
ΣBS and our numerical simulations show that low degree truncations of the power
series (19) are precise approximations of ΣBS . The uniqueness of the solutions
to the system of partial differential equations (1), (2), and (3) remains an open
problem. If for a fixed pair (k0 , c0 ) the power series (19) is convergent, then it is
the unique solution to the system. The question of uniqueness is thus equivalent
to the convergence of the power series (19) in a neighborhood of (k0 , c0 ). We
conjecture that the system of partial differential equations (1), (2), and (3) has a
unique solution.
Conjecture 1. For every pair (k0 , c0 ) and every Σ0 > 0 such that k0 = 6 0 and
c0 > 0, the system of partial differential equations (1), (2), and (3) has a unique
solution Σ in a neighborhood of (k0 , c0 ) that satisfies Σ (k0 , c0 ) = Σ0 .
The recursive definition of the coefficients αm,n makes it difficult to study the
convergence of the series. One possible approach is to adapt the Lagrange inversion
methodology Xia and Cui used in [54] to obtain the one-dimensional power series
representation equivalent to (27) for fixed k.
10.2. Applications to other problems. Our PDE method can be used to ap-
proximate the Margrabe implied correlation as suggested in [33]. If two assets have
dynamics that is governed with ρ-correlated Wiener processes W1 and W2 then the
Margrabe formula gives the time 0 price C0 of the European option to exchange
one stock for the other at time T as follows
C0 = S1 (0)e−q1 T N (d+ ) − S2 (0)e−q2 T N (d− ), where (72)
S1 (0)e−q1 T σ2 T
log S2 (0)e−q2 T
± 2
d± = √ , and
σ T
q
σM = 2
σBS,1 2
+ σBS,2 − 2ρσBS,1 σBS,2 . (73)
The quantities q1 and q2 are the dividend rates for the underlying assets while σBS,1
and σBS,2 are their volatilities. After the substitution
S2 e−q2 T C0 √
k = log , c= , σ = σM T ,
S1 e−q1 T S1 e−q1 T

the formula (72) becomes

c = CBS (k, σ), (74)
where the function CBS is defined in (4).
In practice, one can use the PDE method to first find σBS,1 and σBS,2 . Then
we use the same method to invert the equation (??) for σ. Finally the implied
correlation ρ can be obtained from (73).
Similarly, one can use our PDE method to obtain the critical stock price in a
compound option as shown in [33].
10.3. Extensions of theoretical results. It is expected that Theorem 1 can be

used to obtain further theoretical results about implied volatility. For example, one
such result is
Theorem 2. The implied volatility Σ(k, c) satisfies the following equation

k 1 ∂ Σk
− = . (75)
Σ2 2 ∂c Σc
Proof. Using elementary algebraic transformations we can re-write equation (3) as
4 2
Σ Σ
Σ3 Σkc = Σc Σk − k 2 − ΣΣc −k . (76)
4 2
Using the relation (1) we obtain that the first term on the right-hand side of (76)
satisfies
4
Σ Σk
Σc Σk − k 2 = Σ3 Σcc .
4 Σc
1
Multiplying both sides of (76) by Σc Σ3 and using the previous equation gives us
Σkc Σc − Σcc Σk k 1
2
= 2− . (77)
Σc Σ 2
Applying the quotient rule to the left-hand side of (77) implies (75).
An easy consequence of Theorem 2 is the following curious convexity result about

Black-Scholes implied volatility.
Σk (k,c)
Corollary 1. For every fixed k the function f (c) = Σc (k,c) is concave.
Proof. Differentiating both sides of (75) with respect to c implies

2kΣΣc
f 00 (c) = − .
Σ4
Since Σ and Σc are positive, we obtain that f 00 (c) is negative, which implies the
required result.
Tehranchi in [52] established similar convexity results for the integrated volatility
Σ (see Corollary 1.2).
We also note that Theorem 1 can be concisely stated in terms of the Hessian
D2 Σ as follows.

Theorem 3. The Hessian D2 Σ of the function Σ satisfies the following partial

differential equation

3 2 A(Σ) + B(Σ) 0 A(Σ) − B(Σ) C(Σ)
Σ D Σ = · ,
0 C(Σ) A(Σ) + B(Σ) D(Σ)

where A, B, C, D : C 1 R2 → C R2 are defined as
1 ∂ ∂Σ
A(Σ) = Σ2 Σ, B(Σ) = k − Σ,
2 2 ∂k ∂k2
Σ ∂ Σ ∂
C(Σ) = −k Σ, D(Σ) = +k Σ.
2 ∂c 2 ∂c
It can be easily verified from here that the determinant of the Hessian satisfies
2 !
2
1 2 ∂ k Σ
det D Σ = Σ · − +
2 c ∂k Σ 2
which may be useful for gradient-based optimization procedures.
The system of partial differential equations (1), (2), and (3) might be numeri-
cally solvable using a finite difference method. It remains to be seen whether this
approach would yield effective algorithms that provide accuracy and speed desired
in practice.
11. Concluding remarks

We have proved that the integrated volatility satisfies a system of nonlinear
second order partial differential equations. The obtained equations are used to
construct a recursive algorithm for determining the power series approximation of
the integrated volatility. We demonstrated how our approximation can be used to
obtain faster algorithms for implied volatility and presented the detailed results of
the comparison we made between this new approximation and existing methods.
The gain in speed is substantial, especially when the algorithm is parallelized. We
also demonstrated that the partial differential equations offer additional insights
into the behavior of implied volatility.
Acknowledgements
This work was supported by the Weissman School of Arts and Sciences at Baruch
College. The authors thank their colleagues Beat Ammon, Jim Gatheral, Andrew
Lesniewski, Yike Lu, Elena Kosygina, Avi Palley, Alex Tartakovsky, and anonymous
referees for their comments and suggestions.
References
[1] L. Bachelier. Théorie de la spéculation. Annales Scientifiques de L’Ecole Normale Supérieure,
17, pp. 21–88, 1900.
[2] J. Baumeister. Inverse problems in finance. in: Recent Developments in Computational Fi-
nance: Foundations, Algorithms and Applications, T. Gerstner, P. Kloeden (Eds.), World
Scientific Publishing, Volume 14, Interdisciplinary Mathematical Sciences, Chapter 3, pp.
81–158, 2013.
[3] H. Berestycki, J. Busca, and I. Florent. Computing the implied volatility in stochastic volatil-
ity models. Comm. Pure Appl. Math., 57(10), pp. 1352–1373, 2004.

[4] F. Black and M. Scholes. The pricing of options and corporate liabilities. Journal of Political
Economy, 81, pp. 637–654, 1973.
[5] M. Brenner and M. Subrahmanyam. A simple formula to compute the implied standard
deviation. Financial Analyst Journal, 44, pp. 80–83, 1988.
[6] M. Brenner and M. Subrahmanyam. A simple approach to option valuation and hedging in
the Black-Scholes model. Financial Analyst Journal, 50, pp. 25–28, 1994.
[7] D. M. Chance. A generalized simple formula to compute the implied volatility. The Financial
Review, 31, pp. 859–867, 1996.
[8] D. Chance, T. A. Hanson, W. Li, and J. Muthuswamy. A bias in the volatility smile. Review
of Derivatives Research, 20 (1), pp. 47–90, 2017.
[9] J. Choi, K. Kim, and M. Kwak. Numerical approximation of the implied volatility under
arithmetic Brownian motion. Applied Mathematical Finance, 16 (3), pp. 261–268, 2009.
[10] C. J. Corrado and T. W. Miller Jr. A note on a simple, accurate formula to compute implied
standard deviations. Journal of Banking and Finance, 20, pp. 595–603, 1996.
[11] M. R. Fengler. Option data and modeling BSM implied volatility. in: Handbook of Compu-
tational Finance, pp. 117–142. Springer Berlin Heidelberg, 2012.
[12] M. Forde and A. Jacquier. Small-time asymptotics for implied volatility under the Heston
model. International Journal of Theoretical and Applied Finance 12 (6), pp. 861–876, 2009.
[13] G. Frobenius. Ueber die Integration der linearen Differentialgleichungen durch Reihen. J.
reine angew. Math., 76, pp. 214–235, 1873.
[14] J. Gatheral. The Volatility Surface: A Practitioner’s Guide. John Wiley and Sons, Hoboken,
NJ, 2006.
[15] J. Gatheral, E. P. Hsu, P. Laurence, C. Ouyang, and T.-H. Wang. Asymptotics of implied
volatility in local volatility models. Mathematical Finance 22 (4), pp. 591–620, 2012.
[16] J. Gatheral, I. Matić, D. Stefanica, and R. Radoičić. Tighter bounds for implied volatility.
International Journal of Theoretical and Applied Finance 20 (5), pp. 1–14, 2017.
[17] S. Gerhold. Can there be an explicit formula for implied volatility? Appl. Math. E-Notes, 13,
pp. 17–24, 2013.
[18] K. Glau, P. Herold, D. Madan, C. Pötz. The Chebyshev method for the implied volatility.
Journal of Computational Finance, to appear, arXiv:1710.01797, 2017.
[19] C. Grunspan. Asymptotic Expansions of the Lognormal Implied Volatility: A Model Free
Approach. arXiv:1112.1652, 2011.
[20] C. Grunspan. A Note on the Equivalence between the Normal and the Lognormal Implied
Volatility: A Model Free Approach. arXiv:1112.1782, 2011.
[21] M. Hairer. A theory of regularity structures. Inventiones mathematicae, 198(2), pp. 269–504,
2014.
[22] P. Jäckel. By implication. Wilmott, November, pp. 60–66, 2006.
[23] P. Jäckel. Let’s be rational. Wilmott, January, pp. 40–53, 2015.
[24] P. Jäckel. Implied normal volatility. Wilmott, January, pp. 52–54, 2017.
[25] A. Jacquier, M. Keller–Ressel, and A. Mijatović. Large deviations and stochastic volatility
with jumps: asymptotic implied volatility for affine models. Stochastics 85 (2), pp. 321–345,
2013.
[26] A. Jacquier and M. Lorig. From characteristic functions to implied volatility expansions.
Advances in Applied Probability 47 (3), pp. 837–857, 2015.
[27] H. A. Latané and R. J. Rendleman. Standard deviation of stock price ratios implied by option
premia. Journal of Finance, 31, pp. 369–382, 1976.
[28] F. Le Floc’h. Fast and Accurate Analytic Basis Point Volatility (April 10, 2016). Available
at SSRN: https://ssrn.com/abstract=2420757
[29] F. Le Floc’h. Implied volatility from Black-Scholes price. (2017) Available at
http://chasethedevil.github.io/post/implied-volatility-from-black-scholes-price/
[30] R. Lee. The moment formula for implied volatility at extreme strikes. Mathematical Finance,
14 (3), pp. 469–480, 2004.
[31] S. Li. A new formula on computing the implied volatility. Applied Mathematics and Compu-
tation, 170, pp. 611–625, 2005.
[32] M. Li. Approximate inversion of the Black-Scholes formula using rational functions. European
Journal of Operational Research, 185 (8), pp. 743–759, 2008.
[33] M. Li and K. Lee. An adaptive successive over-relaxation method for computing the Black-
Scholes implied volatility. Quantitative Finance, 11 (8), pp. 1245–1269, 2011.

[34] M. Lorig, S. Pagliarani, and A. Pascucci. A Taylor series approach to pricing and implied
volatility for local-stochastic volatility models. Risk, 17 (2), pp. 1–17, 2014.
[35] D. B. Madan. Adapted hedging. Annals of Finance, 12 (3-4), 305–334, 2016.
[36] D. B. Madan, R. H. Smith, and K. Wang. Laplacian risk management. Finance Research
Letters, 22, pp. 202–210, 2017.
[37] S. Manaster and G. Koehler. The calculation of implied variances from the Black-Scholes
model. The Journal of Finance, 38, pp. 227–230, 1982.
[38] I. Matić, R. Radoičić, and D. Stefanica. Pólya-based approximation for the ATM-forward
implied volatility. International Journal of Financial Engineering, 4 (2-3), 2017.
[39] I. Matić, R. Radoičić, and D. Stefanica. A PDE method for estimation of implied volatil-
ity, source code. open source, available at github.com/maticivan/PDE-method-for-implied-
volatility
[40] E. A. Nurminskii, A. A. Buryi. Parker-Sochacki method for solving systems of ordinary
differential equations using graphics processors. Numerical Analysis and Applications, 4:223,
2011.
[41] G. Orlando and G. Taglialatela. A review on implied volatility computation. Journal of
Computational and Applied Mathematics, 320, 202–220, 2017.
[42] G. E. Parker and J. S. Sochacki. Implementing the Picard iteration. Neural, Parallel Sci.
Comput., 4(1), pp. 97–112, 1996.
[43] M. Pistorius and J. Stolte. Fast computation of vanilla prices in time-changed models and
implied volatilities using rational approximations. International Journal of Theoretical and
Applied Finance 15 (4), 1–34, 2012.
[44] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery. Numerical Recipes in
C++: The Art of Scientific Computing. Second Edition. Cambridge University Press, New
York, NY, 2002.
[45] O. Salazar Celis. A parametrized barycentric approximation for inverse problems with appli-
cation to the BlackScholes formula. IMA Journal of Numerical Analysis, drx020, 2017.
[46] W. Schachermayer and J. Teichmann. How close are the option pricing formulas of Bachelier
and Black-Merton-Scholes? Mathematical Finance 18 (1), 155–170, 2008.
[47] K. Spanderen. Adaptive SOR method for implied volatility calculation.
https://hpcquantlib.wordpress.com/, 2017.
[48] D. Stefanica and R. Radoičić. A sharp approximation for ATM-forward option prices and
implied volatilities. International Journal of Financial Engineering, 3(1), 2016.
[49] D. Stefanica and R. Radoičić. An explicit implied volatility formula. International Journal
of Theoretical and Applied Finance, 20(7), 2017.
[50] G. Steinbrecher and W. T. Shaw. Quantile mechanics. European Journal of Applied Mathe-
matics, 19 (2), pp 87–112, 2008.
[51] M. R. Tehranchi. Uniform bounds for Black–Scholes implied volatility. SIAM Journal on
Financial Mathematics, 7(1), pp. 893–916, 2016.
[52] M. R. Tehranchi. A Black–Scholes inequality: applications and generalisations.
arXiv:1701.03897, 2017.
[53] A. Townsend and L. N. Trefethen. An extension of chebfun to two dimensions. SIAM Journal
on Scientific Computing, 35(6), pp. 495-518, 2013.
[54] Y. Xia and Z. Cui. An Exact and Explicit Implied Volatility Inversion Formula. International
Journal of Financial Engineering, 5(3), 29 pages, 2018.
Baruch College, City University of New York

E-mail address: ivan.matic@baruch.cuny.edu

E-mail address: rados.radoicic@baruch.cuny.edu

E-mail address: dan.stefanica@baruch.cuny.edu

Stefanica 2019 - Volatility

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Stefanica 2019 - Volatility

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Stefanica 2019 - Volatility

Uploaded by

Copyright:

Available Formats

October 15, 2019

A PDE METHOD FOR ESTIMATION OF IMPLIED VOLATILITY

IVAN MATIĆ, RADOŠ RADOIČIĆ, AND DAN STEFANICA

Abstract. In this paper it is proved that the Black-Scholes implied volatility

1. Introduction and formulation of the problem

Electronic copy available at: https://ssrn.com/abstract=3264356

Algorithm 1 Precomputation Phase

The functions KAxisPartition, CAxisPartition, and CreatePolynomial are based

3This structure is identical with the one recently proposed in [18].

Electronic copy available at: https://ssrn.com/abstract=3264356

that approximates the integrated volatility (k, c) 7→ ΣBS with precision ε.

Figure 1. Partition of the (k, c) plane.

Algorithm 2 Runtime Phase

Our implementation of the above algorithm in C++ and OpenCL is available as

4Because of the divergence of our approximation at k = 0 as explained later in the text, in

Electronic copy available at: https://ssrn.com/abstract=3264356

3. Overview of the results from the literature

Electronic copy available at: https://ssrn.com/abstract=3264356

5It should be noted that one still needs to divide by

Electronic copy available at: https://ssrn.com/abstract=3264356

Electronic copy available at: https://ssrn.com/abstract=3264356

3.2. PDE-based method. Our new numerical procedure is relatively simple in

Electronic copy available at: https://ssrn.com/abstract=3264356

Electronic copy available at: https://ssrn.com/abstract=3264356

general enough to be applicable in models other than the Black–Scholes model. In

4. Implied volatility in Black-Scholes model

Electronic copy available at: https://ssrn.com/abstract=3264356

Our domain of approximation can be restricted to k ≥ 0, due to the following

5. Piecewise approximation of the implied volatility

Electronic copy available at: https://ssrn.com/abstract=3264356

We can re-write (17) as

6. Power series representation

Electronic copy available at: https://ssrn.com/abstract=3264356

Placing k = k0 in equation (19) and differentiating twice implies

Electronic copy available at: https://ssrn.com/abstract=3264356

the sequence of coefficients LC is the convolution of four sequences. Three of them

Let us denote by LC,lower

Electronic copy available at: https://ssrn.com/abstract=3264356

As in Subsection 6.1 we introduce the sequences αnKD = (n + 1)αn+1 K

LK,lower αK ∗ αK ∗ αK ∗ αKDD v−1 − Σ30 · (v + 1) · v · αv+1

6.3. Calculation of αm,n . As before, we assume that we are given k0 , c0 , and

Electronic copy available at: https://ssrn.com/abstract=3264356

then their product satisfies

Equation (3) contains a variable k. Unlike k0 , the quantity k is not a constant,

Electronic copy available at: https://ssrn.com/abstract=3264356

Electronic copy available at: https://ssrn.com/abstract=3264356

Algorithm 3 Partition of the k axis

Electronic copy available at: https://ssrn.com/abstract=3264356

Algorithm 4 Partition of the c axis

I0 = · · · = I27 = 0, I28 = · · · = I334 = 1, I335 = · · · = I999 = 2.

Electronic copy available at: https://ssrn.com/abstract=3264356

Algorithm 5 Creation of the polynomial

7.5. Approximation to volatility surface. For our chosen level of precision

Method Time (in seconds) Domain Error

Electronic copy available at: https://ssrn.com/abstract=3264356

Figure 2. Graphs of c 7→ Σ(k, c) for k ∈ {0.1, 0.3, 0.5, 0.7, 0.9}

Electronic copy available at: https://ssrn.com/abstract=3264356

9. Implied volatility in Bachelier model