Independent scholarship · Open accessThe monthly digest

Indexed text attachment

games-and-information-in-higher-categories.tex

Source for Games and Information in Higher Categories

View paper · Download original file

Lines 3601–3700 of 6559. Source text is displayed literally.

3601noting. The incentive is \emph{coinductive}: it is provided not by any payment contingent on performance but3602by the continuation value of the relationship, which is exactly the promised-value state of3603\cref{thm:APS}. And it has a striking aggregate implication. If all firms pay the premium, a dismissed worker must not be3604able to step into an equally good job at once; equilibrium requires $V_U<V_N$, hence a pool of involuntarily3605unemployed workers who would strictly prefer to work at the going wage. Unemployment is the punishment3606state of the automaton.3607 3608\section{Tournaments}3609 3610With several agents the principal can compare them. Let $y_i=e_i+\eta+\varepsilon_i$, where $\eta$ is a3611shock common to all agents and the $\varepsilon_i$ are idiosyncratic. A \term{tournament} pays a prize3612$W_1$ to the agent with the higher output and $W_2<W_1$ to the other \cite{LR81}. Agent $i$ wins when3613$e_i+\varepsilon_i>e_j+\varepsilon_j$; the common shock cancels. If $G$ is the distribution function of3614$\varepsilon_j-\varepsilon_i$ and $g$ its density, agent $i$'s problem is to maximize3615$W_2+(W_1-W_2)G(e_i-e_j)-c(e_i)$, and in a symmetric equilibrium3616\[ (W_1-W_2)\,g(0)=c'(e). \]3617Effort is governed by the \emph{spread} of prizes and by how responsive the probability of winning is to3618effort; the level of prizes serves only to satisfy participation. For risk-neutral agents the first best3619$c'(e)=1$ is attained with $W_1-W_2=1/g(0)$: the more luck matters, the larger the prize for winning must3620be.3621 3622Why compare agents rather than pay each on her own output? The informativeness principle answers3623(\cref{thm:informativeness}). When $\eta$ is present, $y_j$ is informative about the noise in $y_i$, so3624$y_i$ alone is not a sufficient statistic for $e_i$; the difference $y_i-y_j$ is free of $\eta$. Relative3625performance evaluation is valuable exactly to the extent that agents face common shocks \cite{Hol82}. A3626rank-order tournament is a coarse way to use this information---it throws away the margin of victory---but3627it has the merit that the principal's total payment is fixed, so that she has no incentive to misreport3628who won. Its costs are equally structural: agents gain from each other's failure, and so may sabotage or3629collude.3630 3631\section{Teams}3632 3633Now let the agents produce a single joint output $x=x(e_1,\dots,e_n)$, increasing and concave, with3634private costs $c_i(e_i)$. Only $x$ is observed. The agents form a partnership: output is divided among3635them according to sharing rules $s_i(x)$ with $\sum_is_i(x)=x$ for all $x$.3636 3637\begin{theorem}[Holmstr\"om]\label{thm:teams}\Scl[\cite{Hol82}]3638No budget-balanced sharing rule makes the efficient effort profile a Nash equilibrium.3639\end{theorem}3640 3641\begin{proof}[Proof for differentiable rules]3642Efficiency requires $\partial x/\partial e_i=c_i'(e_i)$ for each $i$. Agent $i$'s best response satisfies3643$s_i'(x)\,\partial x/\partial e_i=c_i'(e_i)$. For the efficient profile to be an equilibrium every agent must3644therefore face $s_i'=1$ at the efficient output; but budget balance gives $\sum_is_i'=1$.3645\end{proof}3646 3647Each agent must be the residual claimant of the whole marginal product, and there is only one marginal3648product to go round. The cure is to break the budget: introduce a principal who is not a member of the3649team, and pay $s_i(x)=b_i$ if $x$ reaches the efficient level and nothing otherwise, the principal keeping3650any shortfall. Then every agent is pivotal, and efficient effort is an equilibrium. On this account the3651economic function of an owner is neither to monitor nor to supply capital, but to be a sink.3652 3653The theorem has a clean reading in the calculus of \cref{ch:open}. A partnership is a \emph{closed} game:3654all utility generated inside is distributed inside. Holmstr\"om's result says that the efficient profile3655is not an equilibrium of any closed game of this shape, and that it becomes one when a boundary wire is3656added along which value can leave. The Groves mechanism of \cref{ch:mechanism} is the same phenomenon for3657hidden information: making each agent internalize the total surplus is possible, and incompatible with3658budget balance.3659 3660\section{Many tasks}3661 3662Let the agent supply a vector of efforts $e\in\R^k$ at cost $\tfrac12e^{\top}Ce$, generating benefits3663$b^{\top}e$ for the principal and signals $x=e+\varepsilon$, $\varepsilon\sim N(0,\Sigma)$. With a linear3664contract $w=\alpha+\beta^{\top}x$ and constant absolute risk aversion $r$, the agent chooses3665$e=C^{-1}\beta$, and maximizing total surplus gives3666 3667\begin{proposition}[Multitask incentives]\label{prop:multitask}\Scl[\cite{HM91}]3668\[ \beta^*=(I+rC\Sigma)^{-1}b. \]3669If there are two tasks and the second cannot be measured at all ($\sigma_2^2\to\infty$, so $\beta_2=0$), then3670\[ \beta_1^*=\frac{b_1-b_2\,C_{12}/C_{22}}{1+r\sigma_1^2\,\bigl(C_{11}-C_{12}^2/C_{22}\bigr)}. \]3671\end{proposition}3672 3673The first formula generalizes \cref{ex:LEN}. The second is the interesting one. When the tasks are3674substitutes in the agent's cost ($C_{12}>0$), rewarding the measured task draws effort away from the3675unmeasured one, and the optimal incentive on the measured task is \emph{reduced} by the value of what is3676crowded out. If the unmeasured task matters enough, $\beta_1^*$ is zero or negative: the best contract is a3677flat salary, not because effort cannot be measured but because \emph{some} of it cannot.3678 3679This is the formal theory of a familiar complaint. When a measure of performance becomes a target, effort3680flows to what is measured, and the measure ceases to track what it was introduced to track. The complaint3681goes under the names of Goodhart and Campbell; the proposition says precisely when it has force (tasks are3682substitutes; the unmeasured task is valuable) and what to do (weaken the incentive; or redesign jobs so3683that measurable and unmeasurable tasks are assigned to different agents \cite{HM91}). It also explains why3684incentives inside organizations are so much weaker than a single-task model would recommend, and why the3685strongest incentives are found where the output is one-dimensional.3686 3687\section{Renegotiation}3688 3689A contract is a commitment by both sides. Suppose the parties can tear it up and write another whenever3690both wish. After the agent has chosen her effort but before the outcome is known, the incentive scheme has3691done its work: effort is sunk, and all that remains is risk, which the risk-neutral principal should bear.3692Both parties gain by replacing the contract with a fixed payment. But if the agent anticipates this, she3693has no reason to work.3694 3695\begin{theorem}\Scl[\cite{FT90}]3696In the hidden-action model with renegotiation after effort is chosen and before the outcome is realized,3697no contract induces high effort with probability one. In equilibrium the agent randomizes over effort, and3698the renegotiated contract is a menu that screens the agent according to the effort she privately knows she3699has taken.3700\end{theorem}