# Fair and Efficient Resource Allocation with Externalities

In resource allocation, it is common to assume that agents have a utility for their allocated items and zero utility for unallocated ones. We refer to such valuation domain as 1-D. This assumption of zero utility for unallocated items is not always valid. For example, in the pandemic, allocation of ventilators, oxygen beds, and critical medical help yields dis-utility to an agent when not received in time, i.e., a setting where people consume resources at the cost of others' utility. Various externalities affect an agent's utility, i.e., when an agent doesn't receive an item, it can result in their gain (positive externalities) or loss (negative externalities). The existing preference models lack capturing the setting with these externalities. We conduct a study on a 2-D domain, where each agent has a utility (v) for an item assigned to it and utility (v') for an item not allocated to it. We consider a generalized model, i.e., goods and chores. There is a vast literature to allocate fairly and efficiently. We observe that adapting the existing notions of fairness and efficiency to the 2-D domain is non-trivial. We propose a utility transformation (T_u) and valuation transformation (T_v) to convert from the 2-D domain to 1-D. We study the retention of fairness and efficiency property given this transformation, i.e., an allocation with property 𝒫 in a 1-D domain also satisfies property 𝒫 in 2-D, and vice versa. If a property is retainable, we can apply the transformation, and all the existing approaches are valid for the 2-D domain. Further, we study whether we can apply current results in a 2-D domain when they do not retain. We explore fairness notions such as Envy-freeness (EF), Equitability (EQ), Maxmin Share (MMS), and Proportionality and efficiency notions such as Pareto Optimality, Utilitarian Welfare, Nash Welfare, and Egalitarian Welfare.

• 2 publications
• 7 publications
• 33 publications
07/16/2019

### Almost Group Envy-free Allocation of Indivisible Goods and Chores

We consider a multi-agent resource allocation setting in which an agent'...
07/27/2018

### Fair allocation of combinations of indivisible goods and chores

We consider the problem of fairly dividing a set of items. Much of the f...
07/21/2020

### Dominant Resource Fairness with Meta-Types

Inspired by the recent COVID-19 pandemic, we study a generalization of t...
09/23/2019

### On Weighted Envy-Freeness in Indivisible Item Allocation

In this paper, we introduce and analyze new envy-based fairness concepts...
05/30/2020

### On the Number of Almost Envy-Free Allocations

Envy-freeness is a standard benchmark of fairness in resource allocation...
08/01/2018

### Almost Envy Freeness and Welfare Efficiency in Fair Division with Goods or Bads

We consider two models of fair division with indivisible items: one for ...
02/28/2021

### Fair and Efficient Allocations with Limited Demands

We study the fair division problem of allocating multiple resources amon...

## 1. Introduction

Division of resources among interested agents is well-explored by researchers. Economists have proposed various fairness and efficiency notions widely applicable in real-world settings, such as division of inheritance, land, task, vaccines, etc. (Brams and Taylor, 1996; Moulin, 2004; Segal-Halevi, 2019; STEIHAUS, 1948; Su, 2000). Computer Scientists, along with these notions, also explored computational aspects of some widely accepted fairness notions (Caragiannis et al., 2016; Aziz et al., 2019a, 2020; Barman et al., 2018b; de Keijzer et al., 2009; Freeman et al., 2019b; Procaccia and Wang, 2014). Such endeavors have led to web-based applications that offer readily available solutions, such as Spliddit 111www.spliddit.org, The Fair Proposals System 222www.fairproposals.com, Coursematch 333www.coursematch.io, Divide Your Rent Fairly , etc. In these applications, all the agents provide their valuations towards each item as input. The underlying algorithm aggregates the valuations and provides allocations that are fair and/or efficient as required. The algorithms used focus on resource allocation without externalities, i.e., an agent’s utility for not receiving a resource is zero. We believe this is inadequate when modeling resource allocation for necessary commodities.

The resources can be goods, chores, or a combination of both. Agents would want goods, i.e., they obtain a positive valuation for good for the resource while avoiding chores (negative valuation). Some resources could be goods for one and chore for another, which we call combination. Consider the situation during the Covid pandemic; there is a sudden and steep requirement for life-supporting resources like hospital beds, ventilators, and vaccines. Getting a free covid vaccination affects an agent positively. Even if someone else gets the free vaccination instead of the agent, the agent values it positively, possibly less. The more vaccinated people, the better the situation. If an agent does not receive resources like a ventilator bed can incur a loss. Thus, the valuation of an agent may not only depend on obtaining the resource but also on not getting the resource. Such an effect of external factors on agents’ utility is captured via externality.

Externality in valuation signifies the effect of external factors on agents’ utility. The agent may incur a benefit – positive externalities or cost – negative externalities for an unassigned resource. In the most general form of externalities, the utility of not receiving an item depends on which other agent receives it. With agents, each agent’s valuations are -dimensional; hence we refer to this as a N-D domain. In the absence of externalities, it is a 1-D domain. In a N-D domain, the agent’s utility becomes complex, and in (Branzei et al., 2013), the authors focus only on positive externalities. In this work, which we motivate towards the pandemic, considering both positive and negative externalities is important. We consider externalities such that the agents incur a cost/benefit for not receiving a resource, yet it is independent of which other agent receives it. We refer to such a valuation domain as 2-D domain – valuation for a bundle of items if an agent receives it, and valuation otherwise. If is positive, it is a positive externality, and if is negative, it is a negative externality. In 1-D domain, is zero. In this work, we focus on fair and efficient resource allocation in a 2-D domain for indivisible goods or/and chores. There are widely studied fairness and efficiency notions in a 1-D domain, which we adapt to 2-D.

Fairness notions. Fair allocation of goods and chores is well studied for 1-D domain (Aziz et al., 2018b, 2019a, 2020; Caragiannis et al., 2016; Freeman et al., 2019a, 2020). The most desirable notions of fairness we consider are Envy-freeness (EF), which ensures that no agent has higher utility for a bundle received by another agent (Foley, 1967). Equitability (EQ) ensures that agents have equal utilities for their shares (Dubins and Spanier, 1961), Proportionality (PROP) ensures that every agent receives at least of its valuation of the entire bundle (STEIHAUS, 1948), and Maximin Share(MMS) ensures that each agent receive at least its MMS value (Budish, 2011).

We cannot ignore externalities in a 2-D domain if we want to ensure the above properties. For example, consider two agents , and two goods . Agent 1 has valuations, represented as , and Agent 2 has . If we consider 1-D valuation, allocating to 1 and to 2 is EF, which is not EF in a 2-D domain because agent 1 envies agent 2. Similarly, it is crucial to note that the existing definition of PROP does not capture the essence of proportional allocation in the 2-D domain; for example, it does not consider the dis-utility of not receiving good in case of a negative externality. For the above example, if we only consider the proportionality definition, each agent should receive goods worth at least . Guaranteeing this amount is impossible in a 2-D domain. We define PROP-E, an extension of proportionality for resource allocation with externalities. PROP-E reduces to the Average-share definition proposed in (Seddighin et al., 2019) for additive valuations, and both these definition reduces to proportionality under a 1-D domain(Section 2). We conduct a study for the possibility and challenges of ensuring these fairness notions and their approximations for a 2-D domain.

Efficiency notions. Efficiency is another essential aspect of resource allocation. Common efficiency notions like Pareto Optimality (PO), Maximum Utilitarian Welfare (MUW), Maximum Nash Welfare (MNW), and Maximum Egalitarian Welfare (MEW) also require further analysis in a 2-D domain.

### 1.1. Our Contributions.

In this paper, we study the 2-D domain for indivisible goods or chores with positive and negative externalities for different fairness (EF, PROP-E, EQ, and MMS) and efficiency notions (PO MUW, MNW, and MEW).

We aim to adapt the algorithms designed for obtaining the above properties in a 1-D domain. Towards this, our first approach proposes a reduction from a 2-D domain to 1-D through certain transformations such that ensuring fairness or efficiency in the transformed valuations also ensures them in the original 2-D domain. In 2-D domain, for general valuations we propose (Definition 3.1), and for additive valuations, we propose (Definition 3.2). Our second approach is to verify if an algorithm proposed for 1-D domain can be directly used without transformation. In particular,

• [noitemsep,leftmargin=*]

• EF and its relaxations like EF1, EFX can be retained through the transformation; hence we can adapt all the existing algorithms with appropriate transformation (Proposition 4.1).

• We define PROP-E (Definition 4), an extension of PROP to suit the 2-D domain. We show that PROP-E reduces to Average-share as defined in (Seddighin et al., 2019) (Proposition 4.4). Finally, we show that PROP-E is also retained through transformation for additive valuations (Proposition 4.5).

• We observe that we cannot retain EQ and its relaxations. We show that the algorithm proposed in (Gourvès et al., 2014) for 1-D can be directly used in 2-D (Proposition 4.3) without requiring any transformation.

• Lastly, we study MMS and show that while we can retain MMS using transformation, we cannot retain -MMS under a specific setting. The algorithm proposed in (Ghodsi et al., 2018; Garg et al., 2018; Garg and Taki, 2020; Aziz et al., 2017; Huang and Lu, 2019) cannot be applied directly in 2-D, leaving scope to explore more in this domain. (Propositions [4.6,4.7,4.10,4.12,4.14])

• Efficiency Notions: We show that we can retain Pareto Optimality (PO) and Maximal Utilitarian Welfare (MUW) using transformation (Propositions [5.1,5.2]). We cannot define Maximum Nash Welfare (MNW), and we cannot retain Maximum Egalitarian Welfare (MEW) in a 2-D domain. (Proposition 5.3)

### 1.2. Related Work

There are few papers that consider externalities in resource allocation. Branzie, Procaccia, and Zhang (Branzei et al., 2013) and Seddighin, Saleh, and Ghodsi (Seddighin et al., 2019) study allocating divisible and indivisible goods with positive externalities. Authors in (Branzei et al., 2013) generalize notions of proportionality and envy-freeness indivisible goods allocation with positive externalities. Authors in (Seddighin et al., 2019) explore MMS allocation for goods allocation with positive externalities; it generalizes maximin share as EMMS and further provides approximation algorithms for calculating EMMS and EMMS allocations. (Li et al., 2019) analyze facility location games with externalities. To the best of our knowledge, there is no relevant work that considers negative externalities.

We summarize the existing algorithms for 1-D valuations available for each of the fairness notions.

Envy-freeness. EF may not exist for indivisible items, for example, two agents and one good. Hence we consider two popular relaxations of EF, Envy-freeness up to one item (EF1) (Budish, 2011; Lipton et al., 2004) and Envy-freeness up to any item(EFX) (Caragiannis et al., 2016). We have poly-time algorithms to find EF1 in general monotone valuations for goods (Lipton et al., 2004) and chores (Bhaskar et al., 2020). For additive valuations, EF1 can be found using Round Robin (Caragiannis et al., 2016) in goods or chores, and Double Round Robin (Aziz et al., 2018a) in combination. When the valuations are general, non-monotone, and non-identical, a poly-time algorithm is given by (Bérczi et al., 2020) for two agents. Authors in (Plaut and Roughgarden, 2017) present an algorithm to find EFX allocation under identical general valuations for goods.

Equitability. Relaxation to EQ are EQ1 and EQX (Budish, 2011; Caragiannis et al., 2016). Authors in (Gourvès et al., 2014) present a polynomial-time algorithm to find EQX allocation. They present a pseudo-polynomial time algorithm to find EQ1 and PO allocation in goods and chores in (Freeman et al., 2019b, 2020).

Proportionality. We have PROP1 and PROPX as the relaxation to PROP. For additive valuations, EF1 implies PROP1, and EFX implies PROPX.

MMS. MMS allocations do not always exist (Procaccia and Wang, 2014; Kurokawa et al., 2016). The papers (Procaccia and Wang, 2014; Amanatidis et al., 2017; Barman et al., 2018a; Garg et al., 2018) showed that 2/3-MMS for goods always exists. (Ghodsi et al., 2018; Garg and Taki, 2020) showed that 3/4-MMS for goods always exists. Authors in (Aziz et al., 2017) presented a polynomial-time algorithm for 2-MMS for chores. The algorithm presented in (Barman et al., 2018a) gives 4/3-MMS for chores. Authors in (Huang and Lu, 2019) showed that 11/9-MMS for chores always exists. Authors in (Kulkarni et al., 2021) explored MMS for a combination of goods and chores.

Fair and Efficient. In (Caragiannis et al., 2016), the authors showed that MNW allocation is EF1 and PO for indivisible goods and gave a pseudo-polynomial time algorithm (Barman et al., 2018b). For a combination of resources, the authors in (Aziz et al., 2018a) presented a polynomial-time algorithm to find EF1 and PO for two agents. An Algorithm to find PROP1 and PO was proposed by (Aziz et al., 2019b) for a combination of resources. Authors in (Aziz et al., 2020) proposed a pseudo-polynomial time for finding Utilitarian maximizing among EF1 or PROP1 in goods.

Organization. In Section 2, we define all the notions and definitions related to fairness and efficiency used in this paper. In Section 4, we present results related to preserving fairness under transformation along with algorithmic analysis. In Section 5, we study efficiency preservation upon transformation and its corresponding algorithmic analysis.

## 2. Preliminaries

We consider a resource allocation problem for determining an allocation of indivisible items among interested agents, .

• [noitemsep, leftmargin=*]

• An allocation is an -way partition of . Here, is the bundle assigned to agent and and . We consider a complete allocation of items, i.e., . We represent as .

• Typically in a 1-D domain, agents have valuation function for items. In 2-D domain, , , and for subset , . For goods, , and for chores, . For positive externality, , and for negative externality, .

• The valuation naturally induces utility structure to the agents, i.e., the utility of an agent for a bundle in a 2-D domain as, . Also, . In the case of 1-D domain and , and , and utility .

Note that, in 2-D, additive valuations does not imply additive utility, i.e., , and . However, . Also when agents have identical valuations,

• We assume monotonicity for goods, i.e., , . Similarly, We assume anti-monotonicity for chores, i.e, , . As we increase goods in an agent’s bundle, the utility increases, and as we increase chores, the utility decreases. Note that we use the notation or to represent the valuation of agent for item .

We now formally define the fairness and efficiency criteria that we analyze for the 2-D domain.

###### Definition 0 (Envy-free (EF) and relaxations (Aziz et al., 2018a; Budish, 2011; Bérczi et al., 2020; Caragiannis et al., 2016; Foley, 1967)).

An allocation that satisfies,

 (1) ∀i,j∈N,ui(Ai)≥ui(Aj) is EF (2) (3) ∃k∈{Ai∪Aj} s.t. ui(Ai∖{k})≥ui(Aj∖{k}) is EF1

Informally, is EF when no agent envies another. It is EF1, i.e. Envy-free up to one item, if each agent’s envy can be eliminated by either removing a good from the envied agent’s allocation or removing a chore from its allocation. For EFX, i.e., Envy-free up to any item, the envy is removed when any good or chore is removed. Hence EF EFX EF1.

###### Definition 0 (Equitable (EQ) and relaxations (Dubins and Spanier, 1961; Freeman et al., 2019a, 2020)).

An allocation is said to be equitable, when , . An allocation is said to be EQ1, i.e., Equitable up to one item, . An allocation is said to be EQX, i.e., Equitable up to any item, and , and , and .

Informally, in EQ, all agents are treated equally, i.e., everyone values their bundles equally. It is EQ1, i.e., Equitable up to one good, if the utility of an agent is less than other agents. There exists some good in other agents’ bundle or some chore in its bundle; on removing it, the utility of agent is at least the utility of other agents. Similarly, for EQX, the above satisfies all goods in other’s bundles and for all chores in their bundles.

###### Definition 0 (Proportionality (PROP) (Steihaus, 1948)).

An allocation is said to be proportional, if , .

For 2-D domain, achieving PROP is impossible (e.g., given in Section 1). Hence, we adapt the definition and propose PROP-E for valuations with externalities as follows,

###### Definition 0 (Proportionality with externality (PROP-E)).

An allocation is said to PROP-E if

 (4) ∀i∈N,ui(Ai) ≥1/n.∑j∈Nui(Aj)

We now define the relaxations for PROP-E as below, similar to the relaxations of PROP.

###### Definition 0 (PROP-E relaxations).

An allocation that satisfies,

 (5) (6)
###### Definition 0 (Maxmin Share MMS (Budish, 2011)).

An allocation is said to be MMS if , where

 μi=max(A1,A2,…,An)∈∏n(M)mini∈Nui(Ai)

An allocation is said to be -MMS if it guarantees for , and when , where

We consider the following efficiency criteria,

###### Definition 0 (Pareto-Optimal (PO)).

An allocation is PO if s.t., , and , .

We consider Utilitarian welfare, for which an allocation is the sum of agents utilities. Similarly, Nash welfare corresponds to the product of agents’ utilities, and Egalitarian welfare corresponds to the minimum individual agents’ utility. Formally, the optimal welfare is,

###### Definition 0 (Maximum Welfare).

The optimal welfare is,

 (7) Maximum Utilitarian Welfare, MUW(u)=maxAn∑i=1ui(Ai) (8) Maximum Nash Welfare, MNW(u)=maxAn∏i=1ui(Ai) (9) Maximum Egalitarian Welfare, MEW(u)=maxAminiui(Ai)

## 3. Reduction from 2-D to 1-D

We define two transformations that convert 2-D utilities or valuations to corresponding 1-D valuations. We then represent the transformed valuations in 1-D as . We denote the utility transformation as and valuation transformation as .

###### Definition 0 (Transformation (Tu)).

Given a resource allocation problem where represents 2-D valuations we obtain the 1-D valuations denoted by as follows, ,

 (10) wi(S)=uWi(S)=uVi(S)−v′i(M)

Each agent , has utility function for goods and for chores.

Given that is monotone, is normalized (i.e. ), monotone and non-negative for goods and accordingly for chores, i.e., anti-monotone and negative.

transforms utility from to utility in . Now, in terms of valuations, the transformation that can reduce valuations from to for additive settings is known and is given below. However, such a transformation is unknown in general valuations.

###### Definition 0 (Transformation (Tv)).

We consider the following valuation transformation from 2-D domain to a 1-D domain . Each agent , has valuation function for goods and for chores. We define utility

 (11) wi(S)=vi(S)−v′i(S)
###### Proposition 0.

For additive valuations, utility transformation is equivalent to valuation transform.

Proof Sketch.

 uWi(S) =uVi(S)−v′i(M) =vi(S)+v′i(M∖S)−v′i(M) =vi(S)−v′i(S)

Similarly, we can proof vice versa. ∎

## 4. Fairness in 2-D Domain

In this section, we analyze the challenges in adapting the existing fairness criteria and algorithms for the 2-D domain. We will explore how transformation can help us retain certain fairness properties, and thus we can transform into 1-D and apply the existing approaches. However, in the cases where transformation doesn’t work, we analyze their corresponding algorithms.

### 4.1. Analysis for EF and its relaxation

###### Proposition 0.

Given a resource allocation problem , an allocation is EF/EFX/EF1 in , iff is EF/EFX/EF1 in

Proof Sketch. The utility transformation retains these properties.

 ∀i ∀j,uWi(Ai) ≥uWi(Aj) uVi(Ai)−v′i(M) ≥uVi(Aj)−v′i(M) uVi(Ai) ≥uVi(Aj)

Similarly, the proof follows for EFX and EF1. Accordingly in additive setting retains these properties by Proposition 3.3. ∎

Given the above proposition, we make the following remarks for algorithms to find EF1 and EFX in additive and general 2-D domain.

###### Remark 1.

EF1 for Additive Valuations: Round Robin for goods or chores and Double Round Robin for a combination (Aziz et al., 2018a; Caragiannis et al., 2016) find EF1 in polynomial time in 1-D domain. But is not additive; hence we cannot directly use the algorithms. We can apply the transformation , i.e. we can transform 2-D domain to 1-D, , and find EF1 allocations due to Proposition 4.1.

###### Remark 2.

EF1 for General Valuations: Envy-cycle elimination algorithm (Lipton et al., 2004; Bhaskar et al., 2020) finds EF1 allocations in 1-D general valuations. We can apply the utility transformation , and apply envy-cycle elimination algorithm to to obtained EF1 allocation in due to Proposition 4.1.

Proof Sketch. Envy-cycle elimination algorithm is applicable for indivisible goods with general 1-D valuations. The assumption for the algorithm is that , is non-negative, normalized, and monotone. In 1-D domain , , i.e., the utility function is non-negative, normalized, and monotone. Now, is non-negative, normalized, and monotone. We can apply the envy cycle elimination algorithm to to find EF1 allocation in . Similarly, for chores presented in paper (Bhaskar et al., 2020). ∎

###### Remark 3.

EFX for Identical valuations: Leximin++ finds EFX allocations for a 1-D domain with identical general valuations (Plaut and Roughgarden, 2017). Leximin++ selects the allocation which maximizes the minimum individual utility. If there are multiple such allocations, it will further maximize the size of the bundle of the minimum utility, and then go on to maximizing the second minimum, and so on.

• [noitemsep,leftmargin=*]

• Note that Leximin/Leximin++ allocation in 2-D domain and 1-D domain might not be same.

###### Example 0.

Consider two agents and two goods . Agents have additive valuations. , , , . , and .

• Similarly, we can adapt the cut and choose algorithm for two agents where an agent makes a cut using Leximin++ solution, which gives EFX and PO.

For EF and its relaxation, we show how few of the current algorithms result in ; however, given that the transformations retain this notion, we can adapt any algorithm via .

### 4.2. Analysis for EQ and its relaxations

###### Proposition 0.

Given a resource allocation problem ,

1. With identical general valuations, an allocation is EQ/EQX/EQ1 in iff is EQ/EQX/EQ1 in .

2. For non-identical valuations an allocation is EQ/EQX/EQ1 in does not necessarily mean is EQ/EQX/EQ1 in , and vice versa, even for additive valuations.

Proof Sketch a. When the valuations are identical and additive EF/EFX/EF1 EQ/EQX/EQ1 (Freeman et al., 2019a). Hence, from Proposition 4.1, we can prove the first statement.

Proof Sketch b. Consider the example, where and . The 2-D additive valuations for agent for , , , and . For agent 2, the additive valuations for , , , and .

• is EQ in , but is not even EQ1 in .

• is EQ in but is not even EQ1

Given the above proposition, applying to adapt algorithms proposed for 1-D is not possible unless identical valuations. Despite this, we make the following observation to ensure EQX.

###### Remark 4.

EQX for Additive Valuations: Authors in (Gourvès et al., 2014) proposed a greedy algorithm to find EQX for goods when agents have 1-D additive valuations. In each round, the agent with the least utility picks its favourite good. We can apply this algorithm in .

Proof Sketch. , and an allocation , assume that , and consider the last good added to is , then before adding , suppose the bundles were and for , and respectively. Since was added to ’s bundle, , We know that ,

 ∀k′∈A′i, uVik≤uVik′ vik−v′ik≤vik′−v′ik′ vi(A′i)+v′i(A′i)+vik−v′ik≤vi(A′i)+v′i(A′i)+vik′−v′ik′ vi(A′i)+vik−vik′+v′i(A′i)+v′ik−v′ik′≤vi(A′i)+v′i(A′i) uVik∪{A′i∖k′})≤uVi(A′i)
 ∀k′∈A′i,uVi({k}∪{A′i∖k′})≤uVi(A′i)≤uVj(A′j)≤uVj(Aj)

Even though the transformation doesn’t retain EQX, the algorithm in (Gourvès et al., 2014) can be directly applied in .

### 4.3. Analysis for PROP-E

Ensuring PROP given by Eq. 2.3 is impossible in a 2-D domain; we use the definition given in Eq. 4. In (Seddighin et al., 2019) the authors have also modified PROP to include externalities when valuations are additive. An allocation is proportional when it ensures each agent the Average-share. In the domain , the average value of item for agent , denoted by

 (12) avg[vik]=1/n⋅[vik)+(n−1)[v′ik]]

The average-share of agent ,

 (13) ¯¯¯¯¯¯¯¯¯¯¯¯¯¯vi(M)=∑k∈Mavg[vik]
###### Proposition 0.

For additive valuations, PROP-E reduces to Average-share, and both reduce to proportionality definition in a 1-D domain, i.e., no externalities.

Proof Sketch.

 ∀i∈N,ui(Ai) ≥1/n.∑j∈[n]ui(Aj)=1/n.∑j∈[n]vi(Aj)−v′i(M∖Aj) =1/n.∑k∈[m]vik−1/n.∑j∈[n]v′i(M∖Aj) =1/n.∑k∈[m]vik−1/n.∑j∈[n](n−1)v′ik =¯¯¯¯¯¯¯¯¯¯¯¯¯¯vi(M)

Similarly, we can prove vice versa. ∎

###### Proposition 0.

Given a resource allocation problem , an allocation is PROP-E/PROPX-E/PROP1-E in , iff is PROP/ PROPX/PROP1 in for additive valuations.

Proof Sketch. Consider an allocation is PROP in , i.e., . For additive valuations, is equal to . Now, , and we get . Hence is PROP-E in as given in Eq.4. The proof for PROPX-E and PROP1-E will follow in similar fashion. ∎

[noitemsep, leftmargin=*]

:

It is well know that in a 1-D domain, EF PROP for sub-additive valuation. Whereas in a 2-D domain, EF is always PROP-E, even for general arbitrary valuations, , . Also EF1 PROP1-E, and EFX PROPX-E in additive setting.

###### Remark 5.

PROP-E and relaxations: Like PROP, PROP-E may not always exist, so we have relaxed versions and algorithms. As the transformation retains these properties, we can apply the algorithms in (Aziz et al., 2019b, 2020) via .

### 4.4. Analysis for MMS

###### Proposition 0.

Given a resource allocation problem , an allocation is MMS in , iff is MMS in

Proof Sketch a. Proof follows similar to Proposition 4.1, resulting into

 uVi(Ai)−v′i(M)≥−v′i(M)+maxA∈∏nM minj∈NuVj(Aj)

MMS allocation doesn’t always exists, and even if it does computing it is NP-hard, but a PTAS exists (Woeginger, 1997). There are various algorithms for approximate MMS or -MMS (Procaccia and Wang, 2014; Amanatidis et al., 2017; Barman et al., 2018a; Garg et al., 2018; Ghodsi et al., 2018; Garg and Taki, 2020; Aziz et al., 2017; Huang and Lu, 2019). Unfortunately, -MMS is not retained through .

We know that is positive for goods and negative for chores in a 1-D domain. However, we can no longer guarantee that for goods with negative externalities and chores with positive externalities. For a resource allocation problem, some agents can have , while others can have . We consider the generalized definition of MMS as defined in Def. 2.6.

###### Proposition 0.

Given a resource allocation problem ,

1. If an allocation is -MMS in then is -MMS in , for positive externalities in goods, and negative externalities in chores but not vice versa.

2. If an allocation is -MMS in then is -MMS in , for negative externalities in goods, and for positive externalities in chores but not vice versa.

Proof Sketch [a]. If allocation is -MMS in , then

 ∀i∈N,uWi(Ai) ≥α⋅μWi uVi(Ai)−v′i(M) ≥−α⋅v′i(M)+α⋅μVi

So in the case of positive externalities for goods, , and , and in the case of negative externalities for chores, , and , and so

However, if is MMS in , does not mean it is in .

###### Example 0.

Consider and , both having additive identical valuations, given by, , , , , , and . and . In , is MMS, while it is not in .

Proof Sketch [b]. If allocation is -MMS in , then there are two conditions. Consider in the case of goods in , - i) ii) . We will denote for . Note that in .

If allocation is -MMS in , then

 ∀i∈N,uVi(Ai) ≥αVi⋅μVi uWi(Ai)+v′i(M) ≥αVi⋅v′i(M)+αVi⋅μWi
• , i.e.

 uWi(Ai)+v′i(M) ≥1/αW⋅v′i(M)+1/αW⋅μWi ≥αW⋅v′i(M)+αW⋅μWi

Now,

• , i.e. The proof follows similarly.

However, if allocation is MMS in , does not mean it is MMS in .

###### Example 0.

Consider and , both having additive identical valuations, given by, , , , , , and . and Allocation is in , while it is not in .

The proof follows similarly for chores. ∎

Considering point (a) from Proposition 4.7, we conclude that we can apply transformation and adapt algorithms when we have goods with positive externalities or chores with negative externalities. Yet, due to point (b) from the same proposition, we cannot apply transformation when we have goods with negative externalities and chores with positive. Hence for the latter setting, we explore if we can directly use the algorithms in the 2-D domain to obtain -MMS. We make the following remarks towards some negative results along these lines.

###### Remark 6.

The MMS algorithm for goods in (Ghodsi et al., 2018; Garg et al., 2018; Garg and Taki, 2020) consist of a sub-routine Bag Filling. It is a greedy way of allocating low valued items. For normalized valuation (i.e. , ), if and , , then bag filling guarantees -MMS allocation, where .

###### Proposition 0.

Bag Filling does not guarantee -MMS allocation for goods with negative externalities when applied in .

###### Example 0.

Consider 3 agents - , having additive valuations for 12 goods - provided in Table 1. The valuation profile in Table 1 is based on the MMS doesn’t exist example in the paper (Kurokawa et al., 2018). For calculating MMS Share, Agent 1 divides items as , Agent 2 divides items as , and Agent 3 divides items as .

In Bag Filling, similar to the famous moving knife algorithm, for goods valued , we keep adding goods in a bag until some agent claims it, i.e., the agent values it at least , and so on. We set , i.e., -MMS allocation. In Table 1, for , , , and hence we can apply bag filling to obtain MMS allocation. Note that in , we start filling the bag. till , no one claims it. As the bag reaches , all three claim it, and we break ties lexicographically and assign it to agent , , and . We again start filling, till , no one claims it. As the bag reaches , both agents and claims it. We assign the bag to agent , , and . and we assign remaining items to agent , , and . This allocation is not -MMS in ,i.e., in 2-D domain proving our proposition. ∎

Further in , the allocation found using bag filling is -MMS in .

###### Remark 7.

In the 1-D domain, Bag Filling works for any additive valuations; however, algorithms in (Garg et al., 2018; Garg and Taki, 2020) first converted these valuations into ordered instances and then apply bag filling for low-valued items. However, in 2-D domain, Bag filling doesn’t work even for IDO (Identical Ordinal Preference) instance

###### Remark 8.

For chores, authors in (Aziz et al., 2017) presented a Round Robin algorithm for -MMS allocation, and authors in (Huang and Lu, 2019) presented an algorithm that gives -MMS. However, both these algorithms don’t work as expected in

###### Proposition 0.

For chores with positive externalities in , Round Robin allocation doesn’t guarantee 2-MMS allocation.

###### Proof.

Consider the following example,

###### Example 0.

Consider 2 agents - and 3 chores - , both having additive identical valuations, given by, , , and . , and . In Round Robin, the agents arrive in a fixed order and pick the item with the maximum utility of the remaining items sequentially. Agent choose first, and Round Robin allocation is . is 2-MMS in , but not in

###### Proposition 0.

For chores with positive externalities in , the algorithm in (Huang and Lu, 2019) doesn’t guarantee 11/9-MMS allocation.

###### Proof.

Consider the following example,

###### Example 0.

Consider 3 agents - , having additive valuations for 12 chores - . In Table 1, we negated all the valuations, i.e., , and . Now , and . We set a threshold value for each agent, and then iterate times to create bundles of chores to allocate among agents. In each iteration, we add chores from largest (more negative) to lowest until, , we allocate the bundle to and repeat the process with remaining agents and chores. We set the threshold value to of MMS for each agent, , and .

In , we add chores to a bundle; it doesn’t violate any agent’s constraint. Now, as we add chore ; it violates all three agents’ threshold value. Similarly, adding a chore from to this bundle will again violate. On adding , the bundle satisfies , and further adding any chore will only violate. We break ties lexicographically and assign the bundle to agent 1, , . Similarly, we continue with the algorithm, and we get , , leaving chores unallocated. Assigning them to any agent will violate MMS. On apply this in , we get which is -MMS in .

We cannot retain MMS via transformation for goods with negative externality and chores with positive externality, and we cannot apply the algorithm presented in (Ghodsi et al., 2018; Garg et al., 2018; Garg and Taki, 2020; Aziz et al., 2017; Huang and Lu, 2019) in . It leaves further scope to explore the algorithms for the same.

## 5. Efficiency in 2-D Domain

We will now analyze if the transformation defined in Equation 10 also helps us find efficient allocations in the 2-D domain.

We find utility transformation holds for PO and MUW and does not hold for MEW and MNW. MNW cannot be defined in as some agents might have non-negative utilities while some might have negative utilities.

###### Proposition 0.

Given a resource allocation problem , an allocation is PO in , iff is PO in

###### Proposition 0.

Given a resource allocation problem , an allocation is MUW in , iff is MUW in

Proof for Proposition [5.1,5.2] follows in similar fashion.

###### Proposition 0.

Given a resource allocation problem an allocation is MEW in does not necessarily mean is MEW in , even with additive valuations.