1 Introduction
In the context of risksensitive optimization, Conditional valueatrisk (CVaR
) is a popular measure. CVaR is a conditional expectation of a random variable (r.v.) that usually models the losses in an application (e.g. finance), where the conditioning is based on valueatrisk
. The latter denotes the maximum loss that could be incurred, with high probability. The advantage of employing
CVaR instead of VaR in a risksensitive optimization setting is that CVaR is a coherent risk measure [3], while VaR is not, as it violates the subadditivity assumption.Spectral risk measures are a generalization of CVaR, and are defined as follows:
(1) 
In the equation above, is a riskaversion function, which can be chosen to ensure that SRM is a coherent risk measure [1] and is the
quantile of the distribution of the r.v.
. In particular, a nonnegative, increasing that integrates to is sufficient for ensuring coherence. SRM can be seen as a weighted average of the quantiles (VaR) of the underlying distribution. Moreover, CVaR can be recovered by setting . The latter choice translates to an equal weight for all tailloss VaR values. In contrast, SRM can model a user’s risk aversion better, since the function can be chosen such that higher losses receive a higher weight, or at least, the same weight as lower losses [9].In this paper, we consider the problem of estimating SRM of a random variable (r.v.), given independent and identically distributed (i.i.d.) samples from the underlying distribution. In this context, our contributions are as follows: First, we provide a natural estimation scheme for SRM that uses the empirical distribution function (EDF) to estimate quantiles, together with a trapezoidal rulebased approximation to the integral in (1). Second, we provide concentration bounds for our proposed SRM estimate for the following two cases: first, when the underlying distribution is assumed to have bounded support; and second, when the distribution is either Gaussian or exponential. To the best of our knowledge, no concentration bounds are available for SRM estimation. Third, we perform simulation experiments to show the efficacy of our proposed SRM estimation scheme. In particular, we consider a synthetic setup, and show that our scheme provides accurate estimates of SRM. Next, we incorporate our SRM estimation scheme in the inner loop of the successive rejects (SR) algorithm [4], which is a popular algorithm in the best arm identification framework for multiarmed bandits. We test the resulting SR algorithm variant in a vehicular traffic routing application using SUMO traffic simulator [5]. The application is motivated by the fact that, in practice, human road users may not always prefer the route with lowest mean delay. Instead, a route that minimized worstcase delay, while doing reasonably well on the average, is preferable, and such a preference can be encoded into the risk aversion function in (1).
To the best of our knowledge, concentration bounds are not available for the SRM estimation. However, the bounds that we derive for SRM estimation could be specialized to the case of CVaR. In [7, 19] concentration bounds for the classic CVaR estimator are derived. Our bound, using a different estimator, exhibits a similar rate of exponential convergence around true CVaR. For the case of distributions with unbounded support, concentration bounds for empirical CVaR have been derived recently in [13, 14, 6]. In [13] (resp. [14, 6]), the authors derive an onesided concentration bound (resp. twosided bounds), when the underlying distributions are either subGaussian or subexponential [18]. In comparison to [13]
, we derive twosided concentration bounds for the special case of Gaussian and exponential distributions. Moreover, the bounds that we derive show an improved dependence on the number of samples, say
, and accuracy, say , when compared to the corresponding bounds in [14]. More precisely, the probability that the CVaR estimate is more than an away from the true CVaR is bounded above by , for large enough and some universal constants , in our bound. On the other hand, the corresponding tail bound in [14] is for the subGaussian case, and for the subexponential case. Finally, in comparison to a recent result in [6], our bound exhibits exponential concentration, while the corresponding bound in [6] shows a polynomial decay for . CVaRbased models have been explored in different contexts, for instance, in a bandit application [11], in a portfolio optimization problem [15], and in a general risk management setting [16]. In the simulation experiments, we consider each of these applications, and show the efficacy of our proposed estimation scheme in each application context.The rest of the paper is organized as follows: Section 2 introduces VaR, CVaR, SRM and their estimators from i.i.d. samples, Section 3 presents our estimate of SRM, together with concentration bounds for the case when the underlying distribution has bounded support. Section 4 presents a truncated SRM estimation scheme, and concentration bounds for the case when the underlying distribution is either Gaussian or exponential. Section 5 presents the simulation experiments, Section 6 provides the proofs of the concentration bounds in Sections 3–4, and finally, Section 7 concludes the paper.
2 Preliminaries
For a r.v. , VaR and CVaR at the level , , are defined as follows:
(2) 
where for a real number . can be interpreted as the minimum loss that will not be exceeded with probability . Note that, if
has a continuous and strictly increasing cumulative distribution function (CDF)
, then is a solution to the following:Further, can be interpreted as the expected loss, conditional on the event that the loss exceeds , i.e.,
Let , denote i.i.d. samples from the distribution of . Then, the estimate of , denoted by , is formed as follows [17]:
(3) 
where is the EDF of . Letting denote the order statistics, i.e., , we have .
3 Distributions with bounded support
3.1 Estimation scheme
We estimate , given i.i.d. samples from the distribution of , by approximating the integral in SRM definition (1). Notice that the integrand in (1) has to be estimated using the samples. Recall that is the estimate of , given by (3). We use the weighted VaR estimate to form a discrete sum to approximate the integral, an idea motivated by the trapezoidal rule [8]. The estimate of is formed as follows:
(4) 
In the equation above, is a partition of such that and , where is the length of each subinterval.
In the next section, we present concentration bounds for the estimator presented above, assuming that the underlying distribution has bounded support.
3.2 Concentration bounds
For notational convenience, we shall use and to denote and , for any .
For all the results presented below, we let denote the SRM estimate formed from i.i.d. samples of and with m subintervals, using (4). Let and denote the distribution and density of , respectively.
For the sake of analysis, we make one of the following assumptions:
(A1) Let be a riskaversion function such that and .
(A1) The conditions of (A1) hold. In addition, .
[SRM concentration: bounded case] Let the r.v. be continuous and a.s. Fix .
(i) Assume (A1) holds and , . If , and , then
(5) 
where and , is a constant that depends on the value of the density of the r.v. in a neighborhood of , with as in (4).
(ii) Assume (A1) holds and ,
.
If , and , then
(6) 
where is as in the case above.
Proof.
See Section 6.1. ∎
For small values of , the bound in (6) is better than that in (5). However, the bound in (5) is derived under weaker assumptions on the r.v. and the riskaversion function , as compared to the bound in (6).
In part (i) of the theorem above, we assumed that the density of is bounded below by . This implies that the derivative of VaR is bounded above. The latter condition is required for the trapezoidal rule to provide a good approx to the integral in (1). Moreover, the assumption that the first derivative of VaR w.r.t the confidence level is bounded implies that the underlying r.v. is bounded. This claim can be made precise as follows: For any , it can be shown that (see Lemma B in the Appendix for a proof)
(7) 
Notice that the first derivative of VaR involves a term, and if the random variable is unbounded, then for every , there is an such that . This implies cannot be bounded above uniformly w.r.t , and hence the derivative of VaR cannot be bounded either.
The stronger condition used in part (ii) of Theorem 3.2, in conjunction with (7), implies that the second derivative of VaR is bounded. Now, as before, a bounded second derivative implies that the underlying r.v. is bounded. To see this, the expression for the second derivative of VaR involves a term, and if the r.v. is unbounded, then a uniform bound on would mean that, as , decays too slowly to integrate to something finite, leading to a contradiction. More precisely, the differential inequality can be “solved” to get for large and suitable constants , , and . However, the expression on the RHS integrates to infinity, and hence, no density with unbounded support can have bounded.
4 Gaussian and exponential distributions
4.1 Estimation scheme
Let denote i.i.d. samples from the distribution of . We form a truncated set of samples as follows:
where
is a truncation threshold that depends on the underlying distribution. For the case of Gaussian distribution with mean zero and variance
, we set , and for the case of exponential distribution with mean , we set .We form an SRM estimate along the lines of (4), except that the samples used are truncated samples, i.e.,
(8) 
In the above, , with .
4.2 Concentration bounds
Next, we present concentration bounds for our SRM estimator assuming that the samples are either from a Gaussian distribution with mean zero and variance , or from the exponential distribution with mean . Note that the estimation scheme is not provided this information about the underlying distribution. Instead is formed from i.i.d. samples and with subintervals, using (8).
[SRM concentration: Gaussian case] Assume (A1). Suppose that the r.v. is Gaussian with mean zero and variance , with . Fix . If , then
where is as in Theorem 3.2 (i).
Proof.
See Section 6.2. ∎
[SRM concentration: Exponential case] Assume (A1). Suppose that the r.v. is exponentially distribution with parameter , and . Fix . If , then
where is as in Theorem 3.2 (i).
Proof.
See Section 6.3. ∎
Note that concentration bounds for CVaR estimation can be derived using a completely parallel argument to that of the proof of the theorems above, together with following choice for risk aversion function . The CVaRspecific results are provided in Appendix D.
5 Simulation experiments
In this section, we demonstrate the efficacy of our proposed method for SRM estimation (4), which we shall refer to as SRMTrapz. In our experiments, we set the risk aversion function as follows: . In the following subsection, we consider a synthetic experimental setting to compare the accuracy of SRM estimators. Subsequently, we use SRMTrapz as a subroutine in a vehicular traffic routing application (see section 5.2).
5.1 Synthetic setup
Figure 1 presents the estimation error as a function of the sample size for SRMTrapz. The algorithm is run with two different subdivisions. The samples are generated using a Gaussian distribution with mean and variance . We observe that SRMTrapz with subdivisions performs on par with SRMTrapz
with 150 subdivisions for every sample size. Further, as expected, increasing sample size leads to lower estimation error, while also increasing the confidence (demonstrated by the shrinkage in standard error).
5.2 Vehicular traffic routing
In the vehicular routing application, the traditional objective is to find a route with the lowest expected delay. However, such an objective ignores risk factors. An alternative is to consider the weightedsum delay of each route, and we use SRM to quantify this objective. Thus, given a set of routes, the aim is to find the route (by adaptive sampling) with the lowest SRM of the delay. Simulation of Urban MObility (SUMO) [5]
is an open source, highly portable, microscopic road traffic simulation package designed to handle large road networks. Traffic Control Interface (TraCI)
[20] is a library, providing extensive commands to control the behavior of the simulation online, including vehicle state, road configuration, and traffic lights. We implement our routing algorithm using SUMO and TRACI.For the experiments, we use the street map of the area around IIT Madras, Chennai, India (see Figure 2) obtained from OpenStreetMap (OSM) [12], and then used Netconvert tool to load the map in SUMO. The network has 426 junctions and a total edge length of 123 km. We ran SUMO on this network for timesteps, in which cars, buses, bikes, cycles, and pedestrians were added at different timesteps and in different lanes uniformly. We choose routes between two fixed points, marked as S and D in Figure 2. On these selected routes, we add cars and track them. In Table 2, is the estimated average delay of the th route, and is the SRM estimate for the th route, , using (4), and with samples. We set the number of subdivisions .
route  route  route  route  route  

283.81  287.15  306.80  266.85  325.86  
431.28  361.81  455.83  378.68  390.95 
From Table 2 it is apparent that ROUTE has the lowest expected delay, and ROUTE has the lowest SRM. We consider a bestarm identification (BAI) bandit framework [4], where an algorithm is given a fixed budget. Here, the budget refers to the total number of samples across routes. After the sampling budget, the algorithm is expected to recommend a route, and is judged by the probability that the recommended route is correct (i.e. the best route).
We ran successive rejects (SR), which is a popular BAI algorithm, except that SR is modified to find the route with lowest SRM. Note that the regular SR algorithm finds the route with the lowest expected delay. Algorithm 1 presents the pseudocode for the SRMSRTrapz algorithm, with SRMTrapz used to form SRM estimates for each route. The setting of SUMO is as noted above. We set the budget , number of routes , and subdivisions for SRMTrapz. We observed that Algorithm 1 picks ROUTE with probability .
6 Convergence proofs
6.1 Proof of Theorem 3.2
For establishing the bound in Theorem 3.2, we require a result concerning the error of a trapezoidalrulebased approximation, and a concentration bound for the estimate in (3). We state these results below, and subsequently provide a proof of Theorem 3.2. Let , and be a partition of such that and , is length of each subinterval.
(i) If for , then
(9) 
(ii) If for , then
(10) 
Proof.
See Appendix A. ∎
[VaR concentration] Let the r.v. X be continuous. Fix , then we have
where is a constant that depends on the value of the density of the r.v. in a neighborhood of .
Proof:.
See Proposition 2 in [13]. ∎
Proof of Theorem 3.2..
First, we prove the claim in part (i). Notice that
where the final inequality follows by using Lemma 6.1(i) to infer that for , we have . Now, we have
We now apply Lemma 6.1 to bound each of the terms on the RHS above, to obtain
where is a constant that depends on the value of the density in the neighborhood of , for i = . Thus,
(11)  
Note that in (11). The claim in part (i) follows.
The proof of the result in part (ii) follows in a similar manner. In particular, using part (ii) in Lemma 6.1, with , we obtain
∎
6.2 Proof of Theorem 4.2
Proof..
Recall that the truncation threshold . Letting , we have
(12) 
where , and .
We bound as follows:
(13) 
since is Gaussian with mean zero, and variance . Using , we obtain
leading to
(using (13))  
(since ) 
Hence,
(14) 
s Applying the bound in the Theorem 3.2 to the truncated r.v. , we bound as follows:
(15) 
Hence,
where the final inequality follows from the fact that , which holds since the underlying Gaussian distribution is truncated at .
By using a parallel argument, a concentration result for bounding the lower semideviations can be derived, and we omit the details. ∎
6.3 Proof of Theorem 4.2
7 Conclusions
We proposed a novel SRM estimation scheme that is based on numerical integration, and derived concentration bounds for our SRM estimator for the case of distributions with bounded support, Gaussian and exponential. As future work, it would be interesting to generalize the bounds for Gaussian/exponential distributions to the class of subGaussian/subexponential distributions. An orthogonal direction for future work is to derive a lower bound for SRM estimation, and close the gap (if any) in the upper bound that we have derived.
References
 [1] C. Acerbi. Spectral measures of risk: A coherent representation of subjective risk aversion. Journal of Banking & Finance, 26(7):1505–1518, 2002.
 [2] C. Acerbi and D. Tasche. On the coherence of expected shortfall. Journal of Banking & Finance, 26(7):1487–1503, 2002.
 [3] P. Artzner, F. Delbaen, J. M. Eber, and D. Heath. Coherent measures of risk. Mathematical finance, 9(3):203–228, 1999.
 [4] J. Y. Audibert, S. Bubeck, and R. Munos. Best arm identification in multiarmed bandits. In Conference on Learning Theory, pages 41–53, 2010.
 [5] M. Behrisch, L. Bieker, J. Erdmann, and D. Krajzewicz. SUMO–Simulation of Urban MObility. In The Third International Conference on Advances in System Simulation (SIMUL 2011), Barcelona, Spain, volume 42, 2011.
 [6] S. P. Bhat and L.A. Prashanth. Concentration of risk measures: A Wasserstein distance approach. pages 11739–11748, 2019.
 [7] D. B. Brown. Large deviations bounds for estimating conditional valueatrisk. Operations Research Letters, 35(6):722–730, 2007.
 [8] D. CruzUribe and C. J. Neugebauer. An elementary proof of error estimates for the trapezoidal rule. Mathematics magazine, 76(4):303–306, 2003.
 [9] K. Dowd and D. Blake. After VaR: the theory, estimation, and insurance applications of quantilebased risk measures. Journal of Risk and Insurance, 73(2):193–229, 2006.
 [10] J. M. Dufour. Distribution and quantile functions. Technical report, McGill University, Montreal, Canada, 1995.

[11]
N. Galichet, M. Sebag, and O. Teytaud.
Exploration vs exploitation vs safety: Riskaware multiarmed
bandits.
In
Asian Conference on Machine Learning
, pages 245–260, 2013.  [12] M. Haklay and P. Weber. Openstreetmap: Usergenerated street maps. IEEE Pervasive Computing, 7(4):12–18, 2008.
 [13] R. K. Kolla, L. A. Prashanth, S. P. Bhat, and K. P. Jagannathan. Concentration bounds for empirical conditional valueatrisk: The unbounded case. Operations Research Letters, 47(1):16–20, 2019.
 [14] R. K. Kolla, L. A. Prashanth, and K. P. Jagannathan. Riskaware multiarmed bandits using conditional valueatrisk. CoRR, abs/1901.00997, 2019.
 [15] P. Krokhmal, J. Palmquist, and S. Uryasev. Portfolio optimization with conditional valueatrisk objective and constraints. Journal of risk, 4:43–68, 2002.
 [16] J. M. Mulvey and H. G. Erkan. Applying CVaR for decentralized risk management of financial companies. Journal of Banking & Finance, 30(2):627–644, 2006.
 [17] Robert J Serfling. Approximation theorems of mathematical statistics, volume 162. John Wiley & Sons, 2009.
 [18] M. J. Wainwright. Highdimensional statistics: A nonasymptotic viewpoint, volume 48. Cambridge University Press, 2019.
 [19] Y. Wang and F. Gao. Deviation inequalities for an estimator of the conditional valueatrisk. Operations Research Letters, 38(3):236–239, 2010.
 [20] A. Wegener, M. Piórkowski, M. Raya, H. Hellbrück, S. Fischer, and J. P. Hubaux. TraCI: an interface for coupling road traffic and network simulators. In Proceedings of the 11th communications and networking simulation symposium, pages 155–163. ACM, 2008.
Appendix A Proof of Lemma 6.1
Proof..
The proof follows in a similar fashion as a result in [8], and we provide the details below for the sake of completeness.
Let and . We look at a single interval and operate integrate by parts twice:
(16)  