Bounds and Heuristics for Multi-Product Personalized Pricing
BBounds and Heuristics for Multi-Product Personalized Pricing
Guillermo Gallego ∗ Gerardo Berbeglia † February 17, 2021
Abstract
We present tight bounds and heuristics for personalized, multi-product pricing prob-lems. Under mild conditions we show that the best price in the direction of a positivevector results in profits that are guaranteed to be at least as large as a fraction of theprofits from optimal personalized pricing. For unconstrained problems, the fractiondepends on the factor and on optimal price vectors for the different customer types.For constrained problems the factor depends on the factor and a ratio of the con-straints. Using a factor vector with equal components results in uniform pricing andhas exceedingly mild sufficient conditions for the bound to hold. A robust factor ispresented that achieves the best possible performance guarantee. As an application,our model yields a tight lower-bound on the performance of linear pricing relative tooptimal personalized non-linear pricing, and suggests effective non-linear price heuris-tics relative to personalized solutions. Additionally, our model provides guarantees forsimple strategies such as bundle-size pricing and component-pricing with respect tooptimal personalized mixed bundle pricing. Heuristics to cluster customer types arealso developed with the goal of improving performance by allowing each cluster to pricealong its own factor. Numerical results are presented for a variety of demand modelsthat illustrate the tradeoffs between using the economic factor and the robust factorfor each cluster, as well as the tradeoffs between using a clustering heuristic with aworst case performance of two and a machine learning clustering algorithm. In ourexperiments economically motivated factors coupled with machine learning clusteringheuristics performed best.
Studies of pricing and demand estimation go back to 17th century Davenant [1699] withsignificant work done in the 19th century by Cournot [1838] and others economists of thatera. Optimal pricing continues to attract researchers and practitioners to develop a deeper ∗ Industrial Engineering and Decision Analytics Hong Kong University of Science and Technology,Kowloon, Hong Kong, [email protected]. Supported by RGC project 16211619. † Melbourne Business School, The University of Melbourne, Australia, [email protected]. a r X i v : . [ ec on . T H ] F e b nderstanding and make price optimization more practical. Personalized pricing, a formof third degree price discrimination has emerge as more data is available, computing powerbecomes cheaper, and more transactions are done electronically enabling firms to chargedifferent prices to different customer types. One key question is how much better is per-sonalized pricing over non-personalized pricing. Another is to develop practical heuristicsthat perform well in practice and have a tight worst-case performance guarantee relativeto more sophisticated pricing policies.For the single product case, Bergemann et al. [2020] present a worst case bound of twofor the ratio of personalized to uniform pricing, when the profit function for each customertype is concave and the demand functions take positive values over a common compactset. Malueg and Snyder [2006] under more mild assumptions obtain a bound equal tonumber of types. Elmachtoub et al. [2020] obtains tight and robust bounds that dependson summary statistics of the aggregate demand distribution. Gallego and Topaloglu [2019]provide bounds and heuristics for specific families of demand distributions including linear,exponential and logistic demand functions. They also include robust procedures to clustercustomers types. Chen et al. [2019] present results for single-product distribution-freepricing.In this paper we provide bounds and heuristics for the multi-product pricing problemincluding bounds for optimal pricing under a single factor (with uniform pricing as a specialcase) and show how our results can be used to bound the performance of linear-pricingversus non-linear pricing for the single product case. Our work can be seen as a generaliza-tion of results in Berbeglia and Joret [2020] that provide performance guarantee of uniformpricing for a subset of multi-product demand pricing problems known as envy-free pricing .Besides extending the performance guarantees of uniform pricing to a much broader classof single factor models, and to linear versus non-linear pricing, our performance guaranteesare also stronger because they hold with respect to the optimal personalized pricing profitrather than the optimal non-personalized profit. Consider a firm with n products and m customers types. Let d ij ( p ) denote the demand forproduct i ∈ N := { , . . . , n } for customer type j ∈ M := { , . . . , m } at price vector p :=( p , . . . , p n ). We assume that all demand functions are non-negative. The profit function for type j customers is given by R j ( p ) := (cid:80) i ∈ N p i d ij ( p ). Denote by R ∗ j := max p ≥ R j ( p ) themaximum profit for type j customers and let θ j > , j ∈ M be distribution of the customertypes, so (cid:80) j ∈ M θ j = 1. Then ¯ R := (cid:80) j ∈ M θ j R ∗ j is the optimal profit from personalizedpricing, also known as third-degree price discrimination. Let R ( p ) := (cid:80) j ∈ M θ j R j ( p ) be theprofit over all types at price p and let R ∗ = max p ≥ R ( p ). We will refer to R ∗ as the optimalprofit from non-personalized pricing. Solving for R ∗ may be difficult even when solving After a simple transformation p ← p − c and d ( p ) ← d ( c + p ) if there is a non-zero unit cost vector c . R ∗ j , j ∈ M is easy. This is because the aggregate demand function may be significantlymore complex than the underlying demands for the customer types. As a result, heuristicsare often used. Here we consider a class of single factor heuristics (or pricing policies)where pricing is done along a positive vector f , with R f := max q> R ( qf ). This is a singledimensional optimization problem that can be solved numerically. Notice that f = e ,the vector of ones, results in the uniform pricing policy with R e the optimal profit underuniform pricing. Clearly R f ≤ R ∗ with equality holding if f ∈ arg max p ≥ R ( p ). Severalquestions arise from this setting, including finding tight bounds of the form (1) R ∗ ≤ β R f that provide performance guarantees for simple heuristics (including uniform pricing) fornon-personalized pricing, or (2) ¯ R ≤ β R ∗ to assess the benefit of personalized over non-personalized pricing. While both of these questions have been partially answered for specificpricing models, no tight bounds of type (1) and (2) are known for general pricing modelswith more than one product. To answer these questions simultaneously we will provide atight upper-bound of the form ¯ R ≤ β R f which implies the former bounds on account of R ∗ ≤ ¯ R ≤ β R f ≤ β R ∗ . Brief preview our results : Let ¯ p ij , i ∈ N be an optimal price vector for type j ∈ M .If there are positive scalers h and a k such that hf i ≤ ¯ p ij ≤ h exp( k ) f i for all i ∈ N, j ∈ M ,then under mild conditions ¯ R ≤ (1 + k ) R f . For the special case f = e , if ¯ p ij ∈ [ h, . h ]for some h >
0, then ¯
R ≤ R e . A sufficient condition for f = e is that the products areweek substitute and have the connected substitute property for each j ∈ M , see Berryet al. [2013]. The bounds work verbatim if prices are constrained to the stated intervals. Assumption 0 (A0) : We assume that optimal prices are positive and finite.Assumption 0 (A0) holds for virtually all practical pricing problems as few firms price theirgoods at zero when optimizing profits and there is no demand at infinite prices. In ouranalysis we will first obtain a bound for the unconstrained case assuming that we are ableto solve the pricing problem for each market segment. We then consider the case whereprices are constrained to compact sets of the form p i ∈ [ f i q min , f i q max ] , i ∈ N, j ∈ M for exogenously given 0 < q min < q max < ∞ .Let ¯ p ij , i ∈ N is a vector of optimal prices for R j ( p ) , j ∈ M . Define δ ij ( q ) := 1 if q ≤ ¯ p ij /f i and δ ij ( q ) := 0 otherwise. Let ¯ p j be the vector with components ¯ p ij , i ∈ N andconsider G ( q ) := (cid:88) j ∈ M θ j (cid:88) i ∈ N f i d ij (¯ p j ) δ ij ( q ) q ≥ . Then G ( q ) is the f -weighted demand at the personalized optimal solution ¯ p j , j ∈ M filteringout i, j combinations for which ¯ p ij < qf i , so some combinations are dropped as q increases.On the other hand consider H ( q ) := (cid:88) j ∈ M θ j (cid:88) i ∈ N f i d ij ( qf )3e the f -weighted demand at price vector qf . We are now ready to state our main as-sumption: Assumption 1 (A1) : G ( q ) ≤ H ( q ) for all q ≥ q , the LHS of A1 accumulates the f -weighted demands at optimal per-sonalized pricing for i, j combinations with ¯ p ij ≥ qf i whereas the RHS is the f -weighteddemand over all products and market segments at price vector qf . In many cases the in-equality holds even if we filter terms on the right hand side by δ ij ( q ). Later we will providesufficient conditions for A1, but we state here the weaker A1 to highlight the generality ofour results.To gain intuition of how this inequality will be used, define G j ( q ) := (cid:80) i ∈ N f i d ij (¯ p j ) δ ij ( q )for q ≥ R ∗ j = (cid:82) ∞ G j ( q ) dq . Multiplying by θ j and adding over j ∈ M weconclude that ¯ R = (cid:82) ∞ G ( q ) dq . Notice that δ ij ( q ) = 1 for all q ∈ [0 , q min ] and δ ij ( q ) = 0 forall q > q max , where q min := min i ∈ N,j ∈ M ¯ p ij /f i and q max := max i ∈ N,j ∈ M ¯ p ij /f i . This impliesthat ¯ R = q min G ( q min ) + (cid:90) q max q min G ( q ) dq. Finally, notice that R ( qf ) = qf (cid:48) d ( qf ) = qH ( q ). We are now ready to state our main result. Theorem 1.
Suppose that A0 and A1 hold. Then ¯ R ≤ β R f where β := 1 + ln( q max /q min ) .Moreover, the bound is tight.Proof. ¯ R = q min G ( q min ) + (cid:90) q max q min qG ( q ) dqq ≤ R ( q min f ) + (cid:90) q max q min R ( qf ) dqq ≤ R f (cid:20) (cid:90) q max q min dqq (cid:21) = R f [1 + ln( q max /q min )] = β R f where the first equality follows form G ( q ) = qG ( q ) /q for q > q min >
0; the first inequalityfollows from qG ( q ) ≤ qH ( q ) = R ( qf ) due to A1. The second inequality follows from R ( qf ) ≤ R f .We now address the tightness of the bound. Suppose there are a continuum of typeswith willingness to pay in the interval [1 , ρ ] for some ρ >
1. Assume that tail of thedistribution of types is given by d ( p ) = k min(1 , /p ) over p ∈ [0 , ρ ] with k = 1 /β and β = 1 + ln( ρ ), so that it integrates to one. Then, personalized pricing results in expectedprofit 1, so ¯ R = 1. For uniform pricing, any price in the interval [1 , ρ ] is optimal, resultingin expected profit R e = 1 /β . Clearly ¯ R / R e = β = 1 + ln( ρ ) so the bound is tight.4 orollary 1. Suppose the firm imposes the constraint p ij ∈ [ q min f i , q max f i ] ∀ i ∈ N, j ∈ M for exogenously given f , q min > and q max = exp( k ) q min for some k > . If A0-A1 holdthen ¯ R ≤ (1 + k ) R f where ¯ R is the maximum personalized profit subject to the statedconstraints and R f is the optimal price along qf constrained to q ∈ [ q min , q max ] .Proof. The result follows directly from Theorem 1 since ln( q max /q min ) = k resulting in β = 1 + k . Consequently ¯ R ≤ (1 + k ) R f .We remark that Theorem 1 also holds for continuous customer types as long as q min is bounded away from zero and q max is finite. Slightly sharper bounds can be obtainedif the set of allowable prices is a finite set bounded away from zero. As an example, if Q = { q , . . . , q K } with q k strictly decreasing in k , we obtain β = (cid:80) Kk =1 ( q k − q k +1 ) /q k ≤ K where for convenience we set q K +1 := 0. It is also possible to prove that the sharpenedbound is tight regardless of the number of customer types. The proof follows the samelogic replacing sums with integrals and changing the order of summation and is omittedfor brevity. Proposition 1.
The following two properties are sufficient conditions for A .P1 d ij ( p ) is increasing in p k for all k (cid:54) = i , for all j ∈ M .P2 (cid:80) i ∈ N f i d ij ( p ) is decreasing in p i for all i ∈ N and all j ∈ M . We remark that A1 need only hold at ¯ p j , j ∈ so conditions P1 and P2 are much strongerthan needed. If f = e , then P1 and P2 together state that the products are weak substitutesand have the connected substitute property, see Berry et al. [2013]. These properties aresatisfied by most pricing models studied in the literature including linear demand models,MNL models, Exponomial choice models [Alptekino˘glu and Semple, 2016], envy-free pricingmodels [Rusmevichientong et al., 2006] and any mixture of them. Moreover, P1 and P2 arealso satisfied in pricing models for which even finding a price vector that guarantees some(positive) constant fraction of the optimal non-personalized profit R ∗ is N P -hard . If d j ( p )is differentiable in p for all j ∈ M , then the aggregate demand d ( p ) = (cid:80) j ∈ M θ j d j ( p ) is alsodifferentiable. Then P2 implies that ∇ d ( p ) e ≤ ∇ d ( p ) is the Jacobian matrix. ByP1, ∂d i ( p ) /∂p k ≤ i (cid:54) = k , so together P1 and P2 imply that ∇ d ( p ) is a P-matrix.As a result, its inverse exists and is non-negative and the demand function is injective. For We use the terms increasing and decreasing in the weak sense unless otherwise stated. Consider a demand model where each customer type has a preference list. Assume further that con-sumers remove product i from the list if its price is higher than r i , and then select their top choice from theremaining products, if any. This model satisfies P1 and P2 for f = e and it is equivalent to an assortmentoptimization problem in which r i is the profit for product i . Thus, the strongest negative result to dateabout the inapproximability of assortment optimization ( N P -hardness to approximate to within a factorof Ω(1 /n − (cid:15) ) for every (cid:15) > f (cid:54) = e we can transform the demand function via d ( p ) ← diag(f) d (diag( f ) − p ) andone can verify that the test of weak substitution and connected substitute properties areare equivalent to P1 and P2 above. Details of the proof can be found in the Appendix.We next show how to construct reasonable choices of f when p ∗ is hard to find and thefirm knows ¯ p j , j ∈ M . An economically motivated choice is f = (cid:80) j ∈ M α j ¯ p j with weights α j = θ j R j / ¯ R , j ∈ M . For such f , R f is called the Economic single factor profit. While theeconomic factor works well in practice, it does not minimize ρ among all possible vectors f . To find a robust f , let p Hi := max j ∈ M ¯ p ij , p Li := min j ∈ M ¯ p ij . Define the robust factor by f ∗ i := (cid:113) p Li p Hi , i ∈ N , and the Robust single factor profit by R f ∗ . Let ρ i := p Hi /p Li ∀ i ∈ N ,and ρ ∗ = max i ∈ N ρ i . Theorem 2. ρ ≥ ρ ∗ . Moreover, ρ ∗ is attained by f ∗ .Proof. Let f be any positive vector. Then for any k ∈ Nρ = max i ∈ N,j ∈ M ¯ p ij /f i min i ∈ N,j ∈ M ¯ p ij /f i ≥ max j ∈ M p kj /f k min i ∈ N,j ∈ M ¯ p ij /f i ≥ max j ∈ M p ∗ kj /f k min j ∈ M p ∗ kj /f k = ρ k . Since this holds for all k ∈ N , it follows that ρ ≥ ρ ∗ . We next show that ρ ∗ is attainedby f ∗ . By construction, p ij /f ∗ i ∈ [( p Li /p Hi ) . , ( p Hi /p Li ) . ] with the two bounds attained.Consequently, q max = max i ∈ N ( p Hi /p Li ) . , and and q min = min i ∈ N ( p Li /p Hi ) . . Clearly theproduct that attains the maximum in q max also attains the minimum in q min , so ρ =max i ∈ N p Hi /p Li = max i ∈ N ρ i = ρ ∗ . Suppose that m is large and the price vectors ¯ p j , j ∈ M are dissimilar resulting in alarge ρ and therefore a poor performance guarantee. The firm can potentially improve theworst case performance if it can partition M into K collectively exhaustive and mutuallyexclusive clusters, so M = ∪ Kk =1 M k . We have already dealt with the case K = 1, while K = m corresponds to personalized pricing. The problem is interesting for 1 < K < m .For a given partition, and a given positive vector f k for cluster k , the worst case per-formance is 1 + ln( ρ ( M k )) where ρ ( M k ) is the corresponding worst case ratio for cluster M k and factor f k and ρ ∗ ( M k ) ≤ ρ ( M k ) is the minimal ratio corresponding to the ro-bust factor for cluster M k . More formally, ρ ∗ ( M k ) = max i ∈ N ρ ik , where ρ ik = p Hik /p Lik , p Hi,k := max j ∈ M k ¯ p ij , and p Li,k := min j ∈ M k ¯ p ij . We define the clustering problem as findinga partition to min M ,...,M K max k ρ ∗ ( M k ).This problem is known in the literature as that of minimizing the maximum inter-clusterdistance in the context of graphs [Gonzalez, 1985] and also as the bottleneck problem[Hoschbaum and Shmoys, 1986]. To see that equivalence, consider a graph G with vertices( i, j ) for all i ∈ N and j ∈ M . There are edges between any two nodes that share the same6roduct, say ( i, j ) and ( i, l ). The distance between the two adjacent nodes in the networkis given by max(¯ p ij , p ∗ il ) / min(¯ p ij , p ∗ il ). The reader can confirm that the maximum distancefor a graph G , is equal to ρ ∗ ( G ), and that the distance satisfies the triangle inequality.Fortunately, there is a 2-factor approximation polynomial time algorithm for this prob-lem, which is the best possible unless P = N P , see Gonzalez [1985] and Hoschbaum andShmoys [1986]. We also used the k -means clustering heuristic and compare their perfor-mance. We remark that once the clusters are formed, each cluster will price along a positivevector which may or may not be the robust choice for that cluster. This is because fre-quently the economic factor performs better than the robust factor even though the robustfactor gives the best performance guarantee. In this section we discuss several applications to our results, including linear demands, thelatent class MNL, and Non-Linear Pricing.
We first briefly review the representative consumer problem that results in the linear de-mand model. The task of the representative consumer is to solve the problem max q ≥ [( u − p ) (cid:48) q − q (cid:48) Sq ] where u is the vector of gross utilities, u − p is the vector of net utilities, and S is a positive definitive matrix. The solution that ignores the non-negativity constraintsyields q ∗ = B ( u − p ) = a − Bp where a := Bu and B := ( S + S (cid:48) ) − . Notice that byconstruction B is symmetric and positive definitive. Let P := { p ≥ a − Bp ≥ } . Weassume that a has positive components. Then d ( p ) = a − Bp for all p ∈ P . Maximizing R ( p ) = p (cid:48) ( a − Bp ) yields p ∗ = 0 . B − a = 0 . u , d ( p ∗ ) = 0 . a and R ( p ∗ ) = 0 . u (cid:48) Bu > p ∗ ∈ P . For p / ∈ P , the solution to the representative consumer’s problem is equivalentto solving the linear complementarity problem y ≥ d ( p − y ) ≥
0, and y (cid:48) d ( p − y ) = 0, seeGallego and Topaloglu [2019].Suppose that for all j ∈ M , d j ( p ) = B j ( u j − p ) = a j − B j p for p ∈ P j := { p ≥ a j − B j p ≥ } . where u j and a j are positive vectors, and B j is a symmetric positivedefinitive matrix. Then ¯ p j = 0 . B − j a j = 0 . u j is in P j , and ¯ R = (cid:80) j ∈ M θ j R j can becomputed without problems. Corollary 2.
Suppose that B j has non-positive off-diagonal elements for all j ∈ M . ThenTherorem 1 holds for positive vectors f such that B j f ≥ for all j ∈ M . Moreover, if B j is an M-matrix for all j ∈ M then the Theorem 1 holds for all positive f .Proof. From the assumptions of the corollary, P1 and P2 holds for all j ∈ M . Since P1 andP2 are sufficient for A1, the result of Theorem 1 hold. Moreover, if B j is an M-matrix then B − j has positive components. This implies that the vector (cid:15)B − j e has positive components7hat can be made arbitrarily small, so for any vector positive vector f we can find an (cid:15) > f ≥ (cid:15)B − j e . Multiplying both sides by B j shows that B j f ≥ (cid:15)e > R ≤ β R f ≤ β R ∗ where R f and R ∗ are interpreted as R f = max q ≥ qf (cid:48) ( a − qBf ) and R ∗ = max p ≥ q (cid:48) ( a − Bp ) where a = (cid:80) j ∈ M θ j a j and B = (cid:80) j ∈ M θ j B j . The caveat is that optimal solutions to this problems, say q ∗ f and p ∗ ,may be outside P j for some j ∈ M resulting in negative demands for some products forsome customer types. This cast as question of whether the profits from pricing at q ∗ f orat p ∗ will actually satisfy the guarantees of Theorem 1. We will show that ¯ R ≤ β R f and¯ R ≤ β R ∗ continue to hold even after the adjustments required to ensure that all demandsare non-negative. To see this in a generic form, we will argue that the actual profit when d ( p ) has negative components is at least as large as R ( p ) = p (cid:48) d ( p ). Proposition 2.
The profit under the representative consumer model is equal to R ( p ) + p (cid:48) By ≥ R ( p ) when p / ∈ P .Proof. The expected profit associated with a vector p / ∈ P is given by p (cid:48) d ( p − y ) = p (cid:48) d ( p ) + p (cid:48) By = R ( p ) + p (cid:48) By where y ≥ , d ( p − y ) ≥ y (cid:48) d ( p − y ) = 0. The complementaryslackness condition y (cid:48) d ( p − y ) = 0 can be written as y (cid:48) [ a − Bp + By ] = 0. By the symmetryof B , p (cid:48) By = y (cid:48) Bp = y (cid:48) [ a + By ] = y (cid:48) a + y (cid:48) By ≥
0. The inequality follows from a ≥ , y ≥ y (cid:48) By ≥ B is positive definitive.We can now apply the result to each customer type that has negative demands at p .In particular, if there is a customer type with negative demands at q ∗ f , then the firm willsee profits R j ( q ∗ f ) + q ∗ f (cid:48) B j y j ≥ R j ( q ∗ f ) from type j customers, so R f = max q R ( qf )is a lower bound on the aggregate profit over all customer types at q ∗ f . Consequently,the profit from the single factor model is at least R f ≥ ¯ R /β . In a similar way, R ∗ =max p R ( p ) is a lower bound of the aggregate profits at p ∗ ∈ arg max R ( p ), so the profitunder p ∗ is at least R ∗ ≥ ¯ R /β . One must be aware, however, that for multiple customertypes, p ∗ = 0 . B − a is not necessarily optimal if there are customer types with negativedemands at this price vector. To find a true optimal solution to the problem the firmneeds to solve max p ≥ ,y j ,j ∈ M (cid:80) j ∈ M θ j [ R j ( p ) + p (cid:48) B j y j ] subject to d j ( p − y j ) ≥ , y j ≥ y ji d ij ( p − y j ) = 0 for all i ∈ N, j ∈ M . The solution p ∗ = 0 . B − a together with thecorresponding y j s that solve the linear complementarity problem for market segments withnegative demands is only a heuristic for this problem, so our results continue to hold ifthe more complex problem is solved, with a similar more sophisticated program for pricingalong a factor f .Figure 1 reports a series computational results in order to evaluate the performance ofdifferent pricing strategies. For each value of n and m , we generated 20 random instancesand we reported the average profit as a percentage of the maximum profit that can be ob-tained using personalized pricing. The weight of each segment j was set to θ j = x j / (cid:80) ml =1 x l x l is a uniform random number between 0 and 1. For each segment j we ran-domly generated the matrix B j ensuring it is symmetric and positive definitive and satisfiesP1 and P2. The vector a j was generated uniformly random from (0 , n . The percentagesshown are the average percentage of the profit obtained with respect to the best personal-ized pricing strategy. As we can see, the economic factor outperformed the robust factor,and both performed significantly better than uniform pricing based on f = e . The lowerright-hand table for the optimal price uses the heuristic p ∗ = 0 . B − a adjusting demandsby solving the complementary slackness for customer types with negative demands. Itperformance is similar to that of the economic factor.Figure 1: Average performance of different pricing strategies under the linear demandmodel.Figure 2 reports another set of experiments to quantify the advantages of clusteringconsumer segments into two clusters ( k = 2) under the economic and the robust pricingstrategies. Two clustering algorithms were implemented. The first one is the standardk-means algorithm in which each segment j was assigned the price vector ¯ p j as its rep-resentative point in an n -dimensional space. The second is the farthest point first (FPF)proposed by Gonzalez [1985] where the distance matrix is set as described in Section 3.For these experiments the number of consumer segments was set to m = 6. As can beseen, the best combination was the economic factor coupled with k -means and the worsewas the robust factor with FPF. 9igure 2: Average performance of Economic and Robust pricing strategies for instanceswith 6 segments with and without clustering with k = 2. Suppose that d j ( p ) is the expected demand from an MNL model, so d ij ( p ) = exp( a ij − b ij p i )1 + (cid:80) k ∈ N exp( a kj − b kj p k ) ∀ k ∈ N. The matrix of partial derivatives is given by ∇ d j ( p ) = diag( b j ) (cid:2) d j ( p ) d j ( p ) (cid:48) − diag( d j ( p )) (cid:3) .Since the off-diagonal elements are non-negative we see that d ij ( p ) is increasing in p k , k (cid:54) = i and P1 holds. P2 hold for all positive vectors f such that ∇ d j ( p ) f ≤
0, or equivalently forall positive vectors f such that d j ( p ) (cid:48) f ≤ min i ∈ N f i for all j ∈ M . We can scale f withoutloss of generality so that min i ∈ N f i = 1, which reduces the condition to (cid:80) i ∈ N f i d ij ( p ) ≤ f = e on account of (cid:80) i ∈ N d ij ( p ) = 1 − d j ( p ) ≤
1. For any positive f the condition reduces to (cid:80) i ∈ N ( f i −
1) exp( a ij − b ij p i ) ≤ j ∈ M . We remarkthat A1 requires the condition to hold only at p = ¯ p j , for j ∈ M . We know that for eachconsumer type the optimal price is of the form ¯ p ij = 1 /b ij + m j where m j is the adjustedmarkup for product type j consumers, so the condition is easy to check. In particular, if b ij = b j for each i ∈ N , and j ∈ M , then both the economic and the robust factors can betaken to be equal to e . Corollary 3.
For the LC-MNL ¯ R ≤ β R e ≤ β R ∗ without any further conditions since P1and P2 hold for f = e for all MNL models with arbitrary price sensitivities. In addition,if the b kj = b j is independent of k ∈ N for each j ∈ M , then the economic and the robustfactors are equivalent to e . Finally, ¯ R ≤ β R f ≤ β R ∗ holds for all positive vectors f suchthat d j (¯ p j ) f ≤ min i ∈ N f i for all j ∈ M . We tested the performance of the different heuristics under the LC-MNL model andreported the results in Figure 3. The percentages shown are the average percentage of theprofit obtained with respect to the best personalized pricing strategy. For each value n and value m reported, we generated 20 random instances with n products and m segments.The mean utility of product i to consumer segment j is u ij = a ij − b ij p i where the intrinsic10roduct utility a ij were randomly chosen following a procedure proposed by Rusmevichien-tong et al. [2014] and the (segment and product dependent) price sensitivities b ij wererandomly chosen from a symmetric triangular distribution between 0 and 2. We can ob-serve that while uniform pricing does relatively well (obtaining between 60.6% to 91% ofthe optimal personalized profit) it is surpassed by the Economic and Robust strategieswith get at least 76.3% and 75% respectively. The values for the non-personalized pricingare not necessarily the optimal ones since they were obtained using a multi-variable non-linear solver in Python. This strategy requires much broader computational resources thanthe other three methods which simply rely on a single variable optimization. For example,when n = 100 the solver took on average over 22 times more time than any of the other 3strategies.Figure 3: Average performance of different pricing strategies when the demand model isthe LC-MNL.Similarly to the clustering results for the linear demand model, Figure 4 reports com-putational results that quantify and compare the benefits of clustering consumer segmentsusing k-means and FPF under the Economic single factor and the Robust single factorpricing strategies. The values represent the average percentage of the profit obtained withrespect to the best personalized pricing strategy. As in the linear demand model, m = 6for all these experiments. As can be seen, the best combination was the economic factorcoupled with k -means and the worse was the robust factor with FPF. Specifically, the intrinsic utility of product i for consumer segment j is defined as a ij := ln((1 − σ i ) v ij /n )with probability p = 0 . a ij ::= ln((1 + σ i ) v ij /n ) otherwise. The values v ij and σ i are realizations froma uniform distribution [0 ,
10] and [0 ,
1] respectively. This is an NP-hard problem. k = 2. We now consider the non-linear pricing scheme where the firm sells a single product indifferent bundle sizes- for a broad overview of non-linear pricing see Wilson [1993] and Oren[2012]. Let p i be the price of a size i ∈ N bundle and d i ( p ) is the demand for a size i bundleat the price vector p . Let R ∗ = max p R ( p ) yielding an optimal non-linear price schedule.Let f be a vector with components f i = i, i ∈ N . Then R f := max q qf (cid:48) d ( qf ) correspondsto the linear price schedule p i = iq, i ∈ N . Theorem 2 holds if d i ( p ) is increasing in p k , k (cid:54) = i and if f (cid:48) d ( p ) is decreasing in p . We remark that f (cid:48) d ( p ) = (cid:80) i ∈ N id i ( p ) is the totalnumber of units demanded at price p , so the condition is that the total number of unitsdemanded goes down if the price of any bundle is increased.As an example, suppose that d ( p ) = a − Bp and B is an M -matrix, then Bf ≥ f and in particular for f i = i . Let v := B − a . Then p ∗ = v/
2. If v i is increasing concavethen q min = v n /n and q max = v resulting in β = 1 + ln( nv /v n ).To our knowledge this is the first result that gives a performance guarantee for linearversus non-linear pricing, but we can go further as Theorem 1 works for the personalizedversion as well. More precisely, if d j ( p ) is the demand vector for bundles of size i ∈ N forevery j ∈ M , d ij ( p ) is increasing in p k for all k (cid:54) = i , and f (cid:48) d j ( p ) is decreasing in p for all j ∈ M then Theorem 1 applies and bounds how much better personalized non-linear pricingcan be relative to linear pricing. Theorem 1 can also be used to bound the performanceof non-personalized non-linear pricing schemes relative to personalized non-linear pricingschemes. We summarize the results for personalized non-linear pricing here. Corollary 4. If d ij ( p ) represent the demand for bundles of size i ∈ N in market segment j ∈ M where p i is the price of a size i ∈ N bundle, d ij ( p ) is decreasing in p k , k (cid:54) = i and (cid:80) i ∈ N id ij ( p ) is decreasing in p for all j ∈ M , then ¯ R ≤ β R f ≤ β R ∗ where R ∗ is the profit from the optimal non-personalized non-linear pricing policy and R f is the optimal under non-personalized linear pricing policy.
12e remark that f here was selected as f i = i, i ∈ N for the purpose of comparinga common linear price schedule for all customer types to optimal personalized non-linearpricing. One can instead use the robust f or the economic f to obtain a potentially bettercommon (non-linear) price schedule.Figure 5 reports a computational results about the performance of different pricingstrategies for a non-linear pricing problem as described above. Each consumer segmentfollows a linear model as explained in Section 4.1. The matrix B j associated to segment j was generated in the same way as for the experiments of Section 4.1 whereas insteadof generating a random vector a j , we produced a random vector of utilities u satisfyingthat u i +1 > u i and u i +1 / ( i + 1) < u i /i for all i ∈ [ n − m and each maximum bundle size ( n ). As can be seen from the tables, linearpricing performance relatively well for n = 10 but deteriorates as the maximum bundle sizeincreases achieving only about 77% of optimal personalized non-linear pricing. The resultsdo not seem to be sensitive to the number of customer types. The robust factor slightlyoutperforms the economic factor, and the non-personalized non-linear pricing strategy isvery close to the optimal personalized strategy with little need for clustering types.Figure 5: Performance of different pricing methods under non-linear pricing. Here n rep-resents the maximum bundle size. Consider a set of n products that can be sold as bundles. Suppose that given prices for eachof the 2 n − non-trivial bundles, the firm can obtain the demand for each of the bundles.Selecting the bundle prices to maximize profits is know as the mixed bundle problem.Suppose that p ( x ) , x ∈ { , } n is an optimal solution to the mixed bundle problem. We13ay wonder about the performance of several pricing strategies relative to mixed bundling.The simplest strategy is to use uniform pricing. This gives rise to f ( x ) = e (cid:48) x . A secondstrategy, known as component pricing is to set f ( x ) = p (cid:48) x where p i is the price of component i . Finally, we can have a non-linear function f ( x ) = g ( e (cid:48) x ) where g is an increasing non-linear function. Notice this form yields the same price factor for all bundles of size e (cid:48) x . Thisis known as bundle-size pricing and contains uniform pricing as a special case if g ( e (cid:48) x ) = e (cid:48) x .In all cases, the firm will find and optimal q for the pricing strategy q · f ( x ) for size x bundles, Chenghuan Sean Chu and Sorensen [2001] shows through extensive numericalstudies that bundle-size pricing can do a good job of approximating the benefits of themore complicated mixed bundling strategy but to our knowledge there are no theoreticalwork on tight bounds. Let q max = max x (cid:54) =0 p ( x ) /f ( x ) and q min = min x (cid:54) =0 p ( x ) /f ( x ). Undermild conditions on the demand function for bundles (A0 and A1, or A0 together with P1and P2) we obtain performance guarantees R ∗ ≤ β R f where here R ∗ is the optimal profitunder mixed bundling and R f is the optimal pricing along the vector f ( x ) , x ∈ { , } n ,where β = 1 + ln( ρ ) where ρ = q max /q min . As an example, if n = 2 then the non-trivialbundles are e , e and e = e + e where e i is the i th unit vector in (cid:60) . If p ( e ) = 1 , p ( e ) = 2and p ( e ) = 2 . f (1) = 1 and f (2) = 2 for bundles of size 1 and 2, then q min = 1and q max = 2 so ρ = 2 and β = 1+ln(2) (cid:39) . q min and q max has to be over all bundles and all types. It is also possibleto fit a robust or an economic mixed bundle strategy and as long as the conditions A0 andA1 hold the bound from Theorem 1 holds. If the model is linear or latent class MNL wewill get similar numerical results with the exception that the bundles are interpreted asproducts and the vector f is either the bundle size or either the economic or robust factorin the case of multiple market segments. This paper presents tight performance guarantees for multi-product single factor pricingrelative to personalized pricing with applications to a variety of demand models. Theresults apply to d i ( p i , p − i ), where d i is the demand vector for firm i at price vector p i giventhat competitors offer price p − i for their own goods provided A1 or P1 and P2 hold forfixed p − i . This opens the door to study competition under a variety of pricing scenarios. Proof.
For convenience, we first consider the single customer type case dropping the index j . Let p ∗ ∈ arg max p (cid:48) d ( p ) and δ i ( q ) = 1 for q ≤ p ∗ i /f i and 0 otherwise. Since qf ≤ max( p ∗ , qf ), P2 implies that f (cid:48) d (max( p ∗ , qf )) ≤ f (cid:48) d ( qf ) . p ∗ , qf ) to qf decreases the price of products for which δ i ( q ) = 1.By P1 this has a negative effect on the demand of products for which δ i ( q ) = 0. Thus, (cid:88) i ∈ N f i d i ( qf )(1 − δ i ( q )) ≤ (cid:88) i ∈ N f i d i (max( p ∗ , qf ))(1 − δ i ( q )) . Consequently, (cid:88) i ∈ N f i d i ( qf ) δ i ( q ) ≥ (cid:88) i ∈ N f i d i (max( p ∗ , qf )) δ i ( q ) . Moreover, moving from max( p ∗ , qf ) to p ∗ decreases the prices of products for which δ i ( q ) =0, so by P1 d i (max( p ∗ , qf )) δ i ( q ) ≥ d i ( p ∗ ) δ i ( q ) ∀ i ∈ N. Multiplying by f i , adding and collecting the inequalities we obtain f (cid:48) d ( qf ) ≥ (cid:88) i ∈ N f i d i (max( p ∗ , qf )) δ i ( q ) ≥ (cid:88) i ∈ N f i d i ( p ∗ ) δ i ( q ) . This completes the proof for a single customers type, and implies under the stated assump-tions that (cid:88) i ∈ N f i d ij ( qf ) ≥ (cid:88) i ∈ N f i d ij (¯ p j ) δ ij ( q ) ∀ j ∈ M. Multiplying by the weights θ j and adding over M yields H ( q ) ≥ G ( q ) which is A1. Acknowledgements
We would like to thank Adam Elmachtoub, Pin Gao, Wentau Lu, Preston McAfee, ShmuelOren, and Zhuodong Tang for their help and suggestions, as well as Daniel Aloise forpointing us to the minimax diameter clustering problem and the FPF heuristic used inthis paper.
References
Aydın Alptekino˘glu and John H Semple. The exponomial choice model: A new alternativefor assortment and price optimization.
Operations Research , 64(1):79–93, 2016.Ali Aouad, Vivek Farias, Retsef Levi, and Danny Segev. The approximability of assortmentoptimization under ranking preferences.
Operations Research , 66(6):1661–1669, 2018.Gerardo Berbeglia and Gwena¨el Joret. Assortment optimisation under a general discretechoice model: A tight analysis of revenue-ordered assortments.
Algorithmica , 82(4):681–720, 2020. 15irk Bergemann, Francisco Castro, and Gabriel Y Weintraub. Uniform pricing versusthird-degree price discrimination.
Cowles Foundation Discussion Paper , 2020.Steven Berry, Ami Gandhi, and Phillip Haile. Connected substitutes and invertibility ofdemand.
Econometrica , 81(5):2087–2111, 2013.Hongqiao Chen, Ming Hu, and Georgia Perakis. Distribution-free pricing.
Available atSSRN 3090002 , 2019.Phillip Leslie Chenghuan Sean Chu and Alan Sorensen. Bunndle-size pricing as an approx-imation to mixed bundling.
The American Economic Review , 101(1):263–303, 2001.A. A. Cournot.
Researches into the Mathematical Principles of the Theory of Wealth . TheMacmillan Company, New York, 1838.C. Davenant.
An essay upon the probable methods of making a people gainers in the balanceof trade . James Knapton, London, 1699.Adam N Elmachtoub, Vishal Gupta, and Michael Hamilton. The value of personalizedpricing.
Management Science (forthcoming) , 2020.Guillermo Gallego and Huseyin Topaloglu.
Revenue Management and Pricing Analytics .Springer Verlag, New York, NY, 2019.Teofilo F Gonzalez. Clustering to minimize the maximum intercluster distance.
TheoreticalComputer Science , 38:293–306, 1985.D. S Hoschbaum and D. B Shmoys. A unified approach to approximation algorithms forbottleneck problems.
J. ACM , 33, 1986.David A Malueg and Christopher M Snyder. Bounding the relative profitability of pricediscrimination.
International Journal of Industrial Organization , 24(5):995–1011, 2006.Shmuel S Oren. Nonlinear pricing. In
The Oxford Handbook of Pricing Management . 2012.Paat Rusmevichientong, Benjamin Van Roy, and Peter W Glynn. A nonparametric ap-proach to multiproduct pricing.
Operations Research , 54(1):82–98, 2006.Paat Rusmevichientong, David Shmoys, Chaoxu Tong, and Huseyin Topaloglu. Assort-ment optimization under the multinomial logit model with random choice parameters.
Production and Operations Management , 23(11):2023–2039, 2014.Robert B Wilson.