Diffusion Auction Design

• Diffusion Auction Design is a framework for auctions over social networks where buyers both bid and invite neighbors to participate.
• Key mechanisms like IDM and CDM are developed to enforce truthful bidding and diffusion while balancing revenue, efficiency, and budget constraints.
• Advanced models extend to multi-unit and combinatorial settings, incorporating privacy, fairness, and Sybil-proof features to optimize auction outcomes.

Diffusion auction design studies auctions conducted over a social network in which a seller initially reaches only her direct neighbors, and informed buyers can strategically diffuse the sale information to their neighbors. In this model, a buyer’s report typically contains both a valuation and a subset of neighbors to invite, so mechanism design must jointly handle truthful bidding and truthful diffusion. The central objective is to make “invite all neighbors and report truthfully” a dominant strategy while maintaining individual rationality and some form of budget balance, and while improving seller revenue or allocation efficiency relative to auctions confined to the seller’s local neighborhood (Guo et al., 2021, Li et al., 2017).

1. Formal model and design objectives

A standard formalization represents the market as a graph $G=(V,E)$ with one seller $s$ and a set of buyers. In the single-item case, a buyer $i$ has type $t_i=(v_i,r_i)$, where $v_i$ is her private valuation and $r_i$ is the set of neighbors she can invite; her report is $t_i'=(v_i',r_i')$ with $r_i'\subseteq r_i$. Only buyers reachable from $s$ through reported invitation paths are qualified to participate. A diffusion auction mechanism specifies an allocation rule and a payment rule, and utilities are quasi-linear, $u_i=\pi_i v_i-x_i$ (Guo et al., 2021, Li et al., 2020).

The design criteria extend standard auction desiderata. Individual rationality requires truthful participation to yield nonnegative utility. Incentive compatibility, usually formulated as DSIC or IC, requires truthful reporting of both valuations and neighbor sets. Weak budget balance or non-deficit requires nonnegative seller revenue. In multi-unit and combinatorial variants, the same principles are combined with feasibility constraints on quantities or bundles, and with qualification constraints that ignore unqualified bids (Li et al., 2024, Fang et al., 2023).

The essential difficulty is that classical auctions do not reward information propagation. In the multi-unit literature, this is stated directly: in a classical VCG or Myerson auction among only $s$0, players never gain by inviting competitors, because doing so reduces their chance to win and gives no diffusion reward (Liu et al., 2022). Diffusion auction design is therefore a multidimensional mechanism-design problem in which social relations are strategically coupled to allocation and payment.

2. Foundational single-item mechanisms

The foundational negative result in this literature is that a direct VCG-style extension to social networks preserves incentive properties but may violate budget balance. In the undirected-network model of Li et al., the VCG extension allocates to the highest reachable bidder and charges each buyer the externality imposed on those who depend on her diffusion; it remains IR and IC, but in a line network it can run a deficit because intermediaries receive positive rewards while the winner pays little or nothing (Li et al., 2017).

To address this, Li et al. introduced the Information Diffusion Mechanism (IDM), built around diffusion-critical nodes and the diffusion-critical sequence of the highest bidder. IDM scans the critical sequence and may allocate the item to an earlier critical node rather than the highest bid itself; payments reward critical intermediaries and charge the eventual winner a threshold defined outside her downstream set. The mechanism is IR and IC, and it is weakly budget-balanced. Its revenue telescopes to the threshold at the first critical node and dominates the VCG extension in seller revenue (Li et al., 2017).

This structure was generalized by the Critical Diffusion Mechanism (CDM) framework. In the unweighted case, a CDM is parameterized by cut-sets $s$1 along the critical sequence of the highest bidder. Any choice of cut-sets satisfying information-blocking, node-independence, and diffusion-monotonicity yields an IR, DSIC mechanism. Theorem 2 in that line shows $s$2, where the right-hand side is the Vickrey revenue among the seller’s immediate neighbors; IDM appears as the extreme lowest-revenue member of this family (Li et al., 2019).

A more general characterization was later given for all DSIC+IR single-item diffusion auctions. The necessary and sufficient conditions are: value-monotonicity, bid-independence of win/lose payments, a payment gap equal to the critical bid, diffusion-monotonicity, and a zero-loss baseline. Under these conditions, any monotonic allocation rule is implementable, and for a fixed monotonic allocation the pointwise revenue-optimal payment rule is

$s$3

This result turns diffusion auction design from isolated constructions into a general implementation theory (Li et al., 2020).

3. Revenue, efficiency, and budget-balance trade-offs

The field repeatedly returns to the tension among efficiency, seller revenue, and budget balance. The survey literature summarizes the classical picture: VCG-diffusion is efficient, DSIC, and IR, but not BB; IDM is DSIC, IR, and weak-BB, but not maximally efficient; CDM provides a parametric family that trades efficiency for higher revenue by enlarging the critical edge cuts $s$4 (Guo et al., 2021, Li et al., 2019).

Revenue optimization under Bayesian assumptions leads to a sharper impossibility. In “Optimal Diffusion Auctions,” valuations are independent and regular, and the seller’s expected revenue on a network $s$5 is upper-bounded by the Myerson optimum on the full reachable set. For each fixed network size $s$6, the $s$7-Partial Winner of Myerson’s mechanism is IC, IR, and attains exactly $s$8. However, no single diffusion mechanism is revenue-optimal on all network structures. The same work therefore proposes approximation mechanisms such as CWM, with guarantee

$s$9

which becomes a $i$0-approximation when $i$1 is uniform on $i$2 (Zhang et al., 2023).

A reserve-price line makes this trade-off explicit. “Approximate Revenue Maximization for Diffusion Auctions” proposes APX-R, a reserve-based diffusion mechanism with an explicit reserve

$i$3

Because $i$4 is type-independent, APX-R preserves DSIC in bidding and diffusion, as well as IR and WBB. The mechanism strictly outperforms classical Myerson revenue on the seller’s initial $i$5 bidders and guarantees a $i$6 approximation to the theoretical upper bound, defined as the maximum possible revenue from any network of size $i$7 (Huang et al., 19 Jul 2025).

These results show that diffusion changes the meaning of “optimal auction.” Revenue depends not only on valuation distributions but also on graph structure, reachability, and the incentives of intermediaries. A plausible implication is that reserve design in networks is less a scalar threshold problem than a structural design problem.

4. Fairness, privacy, robustness, and redistribution

Beyond welfare and revenue, later work broadens the objective set. “Incentivize Diffusion with Fair Rewards” argues that rewarding only cut-points is inadequate in well-connected networks, because cut-points rarely exist there. Its Fair Diffusion Mechanism (FDM) rewards more related participants with fairer rewards while guaranteeing that the seller’s revenue is not reduced; the paper states $i$8 (Zhang et al., 2019).

A more general redistribution perspective treats any truthful diffusion auction as a black box. The network-based redistribution mechanism framework keeps the original allocation rule and replaces each payment by $i$9, where branch-wise rebates are computed on a diffusion-critical tree. If the input mechanism is truthful, IR, and non-deficit, the transformed mechanism preserves IR and IC; under revenue-invariance and growth conditions, it can asymptotically return almost 100% of the original revenue (Gu et al., 2023).

Fairness was subsequently formalized through marginal contribution. “Fair Diffusion Auctions” defines Shapley fairness using each buyer’s contribution to efficient social welfare over all permutations. Existing mechanisms such as IDM, CDM, FDM, SCM, CWM, and LDM are said not to approximate this benchmark. The paper introduces the Permutation Diffusion Auction (PDA), a randomized mechanism for $t_i=(v_i,r_i)$0 homogeneous items that is the first diffusion auction satisfying $t_i=(v_i,r_i)$1-Shapley fairness, incentive compatibility, and individual rationality. Its combinatorial extension CPDA is $t_i=(v_i,r_i)$2-Shapley fair (Gu et al., 2024).

Privacy introduces a different robustness criterion. “Differentially Private Diffusion Auction: The Single-unit Case” observes that auction outcomes can leak hidden valuations. It proposes two differentially private diffusion mechanisms, recursive DPDM and layered DPDM, and proves that both guarantee differential privacy, incentive compatibility, and individual rationality for both valuations and neighborhood reports. The main theorems give privacy levels $t_i=(v_i,r_i)$3 for recursive DPDM and $t_i=(v_i,r_i)$4 for layered DPDM. The same work reports that recursive DPDM nearly matches the full exponential-mechanism benchmark in welfare for moderate $t_i=(v_i,r_i)$5, while layered DPDM can recover about $t_i=(v_i,r_i)$6–$t_i=(v_i,r_i)$7 of that benchmark depending on the choice of $t_i=(v_i,r_i)$8 (Jia et al., 2023).

Strategic robustness against fake identities is handled by Sybil-proof mechanisms. “Sybil-Proof Diffusion Auction in Social Networks” introduces the Sybil tax mechanism (STM) and the Sybil cluster mechanism (SCM), the first single-item diffusion mechanisms to achieve both Sybil-proofness and incentive compatibility. STM uses dominator-tree pricing with safe downstream sets; SCM adds a randomized cluster backbone to preserve positive expected diffusion incentives. The paper states that these protections are obtained with a mild sacrifice of social welfare and revenue (Chen et al., 2022).

5. Multi-unit and combinatorial generalizations

Single-item diffusion is the core case, but the literature has progressively moved to richer allocation environments.

Mechanism family	Setting	Main guarantees
LDM / LDM-Tree	$t_i=(v_i,r_i)$9 identical items, diminishing marginal utilities	IC, IR, non-wastefulness; $v_i$0, $v_i$1
MUDAN / MUDAN-$v_i$2	Identical items, single-demand or multi-demand	IC, IR, ND, NW, $v_i$3-weak efficiency; optimal SW in the no-reward class
DCAF	Combinatorial bundles over a social network	If the single-item subroutine and structural subroutines satisfy stated conditions, output is IC, IR, and WBB
Tree fixed-price mechanism	Multiple identical items in a rooted tree	IR, diffusion incentive compatibility, $v_i$4 worst-case approximation to the optimal fixed-price auction

For homogeneous items with diminishing marginal utilities, Liu et al. propose the Layer-based Diffusion Mechanism. LDM allocates layer by layer in an invitation tree or BFS tree, removing carefully chosen competitor sets to block manipulations that couple bidding and inviting. Theorems establish IR and IC, every unit is allocated, and the mechanism weakly improves both welfare and revenue relative to running VCG on the seller’s immediate neighbors. The implementation is polynomial, with total complexity $v_i$5. The same paper notes that the need for an upper bound $v_i$6 on diffusing children is a limitation, and removing $v_i$7 while maintaining IC is open (Liu et al., 2022).

A different design paradigm is graph exploration. “Multi-unit Auction over a Social Network” localizes competition in $v_i$8 stages so that a buyer’s report cannot globally reshape the market. Its mechanisms MUDAN and MUDAN-$v_i$9 satisfy IC and $r_i$0-weak efficiency, and the tightness theorem states that no IC, ND, no-reward diffusion auction can do better than $r_i$1 weak efficiency. The experiments report that MUDAN-$r_i$2 loses at most $r_i$3–$r_i$4 in social welfare relative to optimum, whereas LDM-Tree loses over $r_i$5 in the tested instances (Fang et al., 2023).

The combinatorial setting requires a more radical reduction. “Combinatorial Diffusion Auction Design” states that one cannot directly extend solutions from simpler settings to combinatorial settings. Its DCAF framework decomposes the market into candidate distributors, bundle division, and diffusion resale processes, each invoking a chosen single-item diffusion auction $r_i$6 as a subroutine. The main theorem shows that if $r_i$7 is IC, IR, and Revenue-Consistent, the candidate-distributor process is consistent, and the bundle division process is resale-diffusion-monotonic, then DCAF is IC, IR, and WBB. Worst-case bundle division is exponential in $r_i$8, but greedy or random single-item heuristics give polynomial-time implementations (Li et al., 2024).

A separate fixed-price branch studies tree-structured networks under uniform valuations. There, the seller allocates multiple items across level-1 branches, posts branch-specific prices, and rewards nodes on invitation paths by geometrically discounted payments. The mechanism is IR and incentive-compatible with regard to buyers’ action, and its worst-case approximation ratio to the optimal fixed-price auction is $r_i$9; Monte-Carlo simulations report that the ratio to the optimum rises from $t_i'=(v_i',r_i')$0 at $t_i'=(v_i',r_i')$1 to $t_i'=(v_i',r_i')$2 at $t_i'=(v_i',r_i')$3 (Yu, 2024).

6. Current frontier and open problems

The survey literature identifies several open directions: multi-unit auctions with combinatorial demands, Sybil attacks and collusion, the exact efficiency–budget-balance frontier in DSIC families such as CDM, double auctions on social networks, networked public-goods provision, and richer network metrics such as clustering and centrality in mechanism design (Guo et al., 2021).

Recent probabilistic work addresses one of the longest-standing gaps: obtaining IC, nonnegative revenue, and a constant approximation to efficiency simultaneously. “Probabilistic Mechanism Design in Diffusion Auctions” proposes the Probabilistic Diffusion Mechanism (PDM) for path graphs, where allocation probabilities are defined directly from ordered valuations and payments split winner charges and referral rewards. On a path, PDM is feasible, IR, WBB, and Sybil-proof, and it satisfies $t_i'=(v_i',r_i')$4-efficiency. The mechanism extends to arbitrary graphs through an incentive-diffusion map $t_i'=(v_i',r_i')$5, producing $t_i'=(v_i',r_i')$6-PDM; when $t_i'=(v_i',r_i')$7 is breadth-first order or generalized BFS, the mechanism is also Sybil-proof and provides approximate revenue. The same work further gives collusion-proof and multi-unit variants MUPDM and SP-MUPDM (Zhang et al., 17 May 2026).

Another persistent conclusion is impossibility of universal optimality. The Bayesian revenue literature shows that no globally optimal mechanism exists across all network structures (Zhang et al., 2023). Multi-unit work shows that $t_i'=(v_i',r_i')$8-weak efficiency is the best possible in the no-reward, IC, ND class (Fang et al., 2023). Privacy work shows that stronger privacy can be bought by sacrificing some welfare, while recursion often performs well in shallow networks (Jia et al., 2023). Fairness work shows that Shapley-based benchmarks are fundamentally misaligned with the payment patterns of earlier critical-path mechanisms (Gu et al., 2024).

Taken together, these strands indicate that diffusion auction design is no longer a single problem but a family of network mechanism-design problems indexed by structural objective: truthful diffusion, budget balance, welfare approximation, revenue maximization, privacy, fairness, and robustness against strategic identities or coalitions. This suggests that future progress is likely to continue through modular designs—black-box lifting, redistribution layers, probabilistic orderings, and privacy-preserving randomization—rather than through a single canonical mechanism.