21+. BLP Contraction Mapping
Application
Given outer $\theta$, invert observed shares $S$ for mean utilities $\delta$. The map $\delta \mapsto s(\delta;\theta)$ is nonlinear with no closed form; Berry (1994) supplies a fixed-point iteration whose convergence rests on the [[21. Contraction Mapping Theorem|Banach Fixed-Point Theorem]].
Also see [BLP(1995)'s appendix 1](/home/haruki/Documents/Mistcraft/Learning/University/Economics/Research Repository/1. Articles Library/1. Automobile Prices in Market Equilibrium.pdf)
1. Setup
Market $t$, $J$ inside goods + outside good $j=0$:
Solve $s(\delta,\theta) = S$ for $\delta$.
2. The Iteration
Pure logit ($\mu_{ij}\equiv 0$) admits closed-form inversion via the log-odds identity:
Random coefficients break the closed form (the integral (2) has no analytic inverse). Since $\delta_j$ enters shares only through $\exp(\delta_j)$, the natural correction is **multiplicative on $\exp(\delta_j)$** — scale it by the observed/predicted ratio $S_j / s_j(\delta^{(n)},\theta)$:
Taking $\ln$ turns the multiplicative update into an additive one:
The $\ln S_j - \ln s_j$ term is the **log-gap** between observed and predicted shares — vanishes iff $s_j = S_j$. In pure logit (5) is **Newton-exact**: one iteration lands at $\delta^\star$. With $\mu_{ij}\ne 0$ it stays a strict contraction (Section 3 bounds the Jacobian by $\bar q < 1$), so iteration converges from any $\delta^{(0)}$ — just no longer in one step.
Fixed point: $\ln s_j(\delta^\star,\theta) = \ln S_j \iff s(\delta^\star,\theta) = S$.
3. Banach Conditions
Space $X = \mathbb{R}^J$ with $\|\cdot\|_\infty$; complete. Multivariate MVT (均值定理) gives
so contraction reduces to **uniformly bounding the sup-norm of the Jacobian $J$** by some $\bar q < 1$.
3.1 Jacobian of $T$
BLP's key decomposition: factor $e^{\delta_j}$ out of the share numerator,
so $T_j(\delta) = \ln S_j - \ln D_j(\delta,\theta)$ — the fixed-point map depends on $\delta$ only through $D_j$, and $D_j$ is strictly **decreasing** in every $\delta_k$ (denominator rises with each $\delta_k$).
Individual logit derivative (own effect minus IIA cross-effect, from (1)):
Chain rule ($\partial \ln s_j = \frac{1}{s_j}\partial s_j$) + expectation from (2) ($\partial s_j/\partial \delta_k = \mathbb{E}[\partial s_{ij}/\partial \delta_k]$):
Since $\partial\delta_j/\partial\delta_k = \mathbf{1}\{j=k\}$ and $\ln S_j$ is constant in $\delta$:
Substituting (7) into (8), and using $\mathbb{E}[s_{ij}] = s_j$:
Every entry non-negative — the load-bearing structural fact.
3.2 Row-Sum Bound
Sup-norm (a.k.a. row-sum norm) is defined as $\|A\|_\infty = \max_j \sum_k |A_{jk}|$. Two features of (9) make it the natural choice: (i) all entries non-negative, so $|J_{jk}| = J_{jk}$; (ii) the individual share identity $\sum_{k=0}^{J} s_{ik} = 1$ rewrites $\sum_{k=1}^{J} s_{ik} = 1 - s_{i0}$, collapsing the row sum:
Strict positivity $s_{ij},\,s_{i0} > 0$ a.s. gives $\mathbb{E}[s_{ij}s_{i0}] > 0$, so the subtracted term is strictly positive:
3.3 Compact Invariant Set
BLP-Berry explicit construction of $K \subset \mathbb{R}^J$ (BLP 1995 Assumption 2 + Berry 1994):
- **Lower bound**: since $D_j$ is decreasing, $\lim_{\delta \to -\infty} D_j = \int e^{\mu_{ij}}\,dF(\nu)$ is its max. So $s_j = S_j$ forces
(smallest $\delta_j$ consistent with $S_j$ when all other $\delta_k \to -\infty$).
- **Upper bound**: Berry (1994) picks $\bar\delta_j$ = the $-\delta_j$ that would explain outside share $s_0$ when $\delta_0 = 0$ and all other $\delta_k \to -\infty$; take $\bar\delta > \max_j \bar\delta_j$. Any $\delta > \bar\delta$ would force $s_k(\delta) > S_k$ for some $k$, so iterates can't escape upward.
Iterates from any $\delta^{(0)}$ enter $K \equiv [\underline\delta,\,\bar\delta]^J$ and stay. Continuous $q(\delta) < 1$ on compact $K$ attains $\bar q \equiv \max_{\delta \in K} q(\delta) < 1$. Multivariate MVT:
4. Banach Consequences
- **Existence & uniqueness**: unique $\delta^\star(\theta)$ with $s(\delta^\star,\theta) = S$.
- **Global convergence**: any $\delta^{(0)}$ works; BLP default is the pure-logit inversion $\ln S_j - \ln S_0$.
- **Linear rate**:
5. Why the Crawl
$\bar q \approx 1 - \frac{\mathbb{E}[s_{ij}s_{i0}]}{s_j}$. Small outside share $\Rightarrow \bar q \to 1 \Rightarrow$ slow linear rate. Newton gets quadratic at $\delta^\star$ but needs $\partial s/\partial \delta$ each step. pyBLP default `iteration='squarem'` accelerates the same contraction without forming the Jacobian; Banach still supplies the guarantee.
Conclusion
Q.E.D.

