---
title: Pessimistic Assortment Optimization (PASTA)
url: https://www.emergentmind.com/topics/pessimistic-assortment-optimization-pasta
type: topic
---

# Pessimistic Assortment Optimization (PASTA)

Pessimistic Assortment Optimization (PASTA) is an offline, data-driven approach to assortment optimization that applies the principle of pessimism to historical customer-choice data. In the original formulation, the seller observes logged assortments, customer choices, and revenues, does not know the underlying choice model, and seeks an assortment with maximal expected counterfactual revenue. PASTA replaces optimistic plug-in optimization with a max–min rule over a likelihood-based confidence region, thereby guarding against overestimation caused by sparse coverage in a combinatorial action space. Its central structural feature is that, under the multinomial logit (MNL) model, correct identification and finite-sample regret guarantees require only that the optimal assortment be covered by the offline data, rather than that all feasible assortments be observed [2302.03821]. Subsequent work has broadened the term from a specific algorithm to a more general pessimistic principle for offline assortment learning, and has extended its scope to weaker coverage notions, richer model classes, and joint assortment–pricing settings [2502.06777] [2510.01693] [2604.19008].

## 1. Offline formulation and the coverage problem

In the original problem setup, the item set is $[N]=\{1,2,\dots,N\}$, each item $i$ has a feature vector $x_i\in\mathbb{R}^d$, and the feasible family of assortments is $\mathbb{S}\subseteq 2^{[N]}\setminus\{\emptyset\}$. The offline dataset is
\[
D=\{(S_i,A_i,R_i)\}_{i=1}^n,
\]
where $S\in\mathbb{S}$ is the offered assortment, $A\in S\cup\{0\}$ is the observed choice, and $R$ is the realized revenue. Writing $\pi_S(s)=\mathbb{P}(S=s)$ for the probability that assortment $s$ appears in the log, the objective is to identify
\[
s^\star \in \arg\max_{s\in\mathbb{S}} \mathbb{E}[R(s)],
\]
where $R(s)$ is the potential revenue under intervention $S=s$ [2302.03821].

The original analysis adopts the standard causal assumptions of consistency and un-confoundedness. Under these assumptions, together with “positivity everywhere,” the expected revenue of every assortment is identifiable from observed data. The difficulty is that “positivity everywhere,” namely $\pi_S(s)>0$ for all $s\in\mathbb{S}$, is typically implausible when $|\mathbb{S}|$ grows exponentially in $N$ and firms rarely explore revenue-poor assortments. PASTA therefore replaces this with the strictly weaker “positivity at optimum” condition,
\[
\pi_S(s^\star)>0,
\]
meaning that the offline data covers the optimal assortment but need not cover the full combinatorial action space [2302.03821].

This distinction addresses a common misconception. PASTA is not designed around uniform identification of all assortments. Its premise is that the decisive quantity is estimation quality at the optimal assortment, because the pessimistic objective suppresses overoptimistic valuations of poorly covered alternatives [2302.03821].

## 2. MNL model and pessimistic objective

The original PASTA paper studies the MNL model with unknown parameter $\theta^\star\in\Theta\subset\mathbb{R}^d$, where $\Theta$ is compact. For an assortment $s$, the choice probabilities are
\[
\pi_A(i\mid s;\theta^\star)=\frac{\exp(x_i^\top \theta^\star)}{1+\sum_{j\in s}\exp(x_j^\top \theta^\star)},\qquad i\in s,
\]
and
\[
\pi_A(0\mid s;\theta^\star)=\frac{1}{1+\sum_{j\in s}\exp(x_j^\top \theta^\star)}.
\]
If $r_i\ge 0$ is the known per-item revenue and $r_0=0$, then the expected revenue of assortment $s$ under parameter $\theta$ is
\[
V(s;\theta)=\frac{\sum_{i\in s} r_i\exp(x_i^\top \theta)}{1+\sum_{j\in s}\exp(x_j^\top \theta)}.
\]
The true optimal assortment is $s^\star\in\arg\max_{s\in\mathbb{S}}V(s;\theta^\star)$ [2302.03821].

The pessimistic construction begins from the empirical negative log-likelihood
\[
\widehat L_n(\theta)= -\frac{1}{n}\sum_{i=1}^n \log \pi_A(A_i\mid S_i;\theta),
\]
with MLE $\hat\theta_{\mathrm{ML},n}\in\arg\min_{\theta\in\Theta}\widehat L_n(\theta)$. PASTA then forms the likelihood-ratio confidence region
\[
\Omega_n(\alpha_n)=\left\{\theta\in\Theta:\widehat L_n(\theta)-\widehat L_n(\hat\theta_{\mathrm{ML},n})\le \alpha_n\right\}.
\]

Its decision rule is the max–min program
\[
\hat s_{\mathrm{PASTA},n}\in\arg\max_{s\in\mathbb{S}} \min_{\theta\in\Omega_n} V(s;\theta).
\]
The inner minimization penalizes assortments whose revenue estimates are uncertain due to insufficient coverage; the outer maximization selects assortments with high robust value. This directly contrasts with estimate-then-optimize baselines that first fit $\hat\theta_{\mathrm{ML},n}$ and then maximize $V(s;\hat\theta_{\mathrm{ML},n})$, a procedure that can overestimate poorly covered assortments and therefore mis-rank the action space [2302.03821].

## 3. Identification logic and regret guarantees

The regret of an assortment estimator $\hat s$ is
\[
\mathrm{Regret}(D)=V(s^\star;\theta^\star)-V(\hat s;\theta^\star).
\]
For PASTA, the core calibration lemma states that if $\theta^\star\in\Omega_n$, then
\[
R(\hat s_{\mathrm{PASTA},n})\le \max_{\theta\in\Omega_n}\{V(s^\star;\theta^\star)-V(s^\star;\theta)\}.
\]
Thus the regret is controlled entirely by the worst-case drop in value at the optimal assortment, not by uniform control over all $s\in\mathbb{S}$ [2302.03821].

Under Assumption 4, $\pi_S(s^\star)>0$, and a likelihood-based concentration condition ensuring both confidence-set coverage and concentration of empirical likelihood around population likelihood, the original paper proves
\[
R(\hat s_{\mathrm{PASTA},n}) \lesssim r_{s^\star} C_{s^\star}\sqrt{\alpha_n},
\qquad
C_{s^\star}=\frac{1}{\pi_S(s^\star)},
\]
with high probability. The dependence on $C_{s^\star}$ makes explicit that the guarantee worsens when the optimal assortment is only rarely observed, but it does not require positivity away from the optimum [2302.03821].

For MNL, the theory specializes further. Let $\theta_{\max}=\sup_{\theta\in\Theta}\|\theta\|_2<\infty$, and define $C_A>0$ so that $1/\pi_A(i\mid s;\theta)\le C_A$ for all $\theta\in\Theta$, $s\in\mathbb{S}$, and $i\in s$. With
\[
\alpha_n \asymp \frac{C_A d}{n}\log\frac{\theta_{\max}}{\delta},
\]
the high-probability regret bound becomes
\[
R(\hat s_{\mathrm{PASTA},n})
\lesssim
r_{s^\star} C_{s^\star}
\sqrt{\frac{C_A d}{n}\log\frac{\theta_{\max}}{\delta}}.
\]
The paper summarizes this as regret scaling like $O(\sqrt{d/n})$ up to coverage-related constants. An important limitation is that under MNL one has $C_A\ge N$, so the bound contains terms that grow at least as $\sqrt{N}$; the paper explicitly notes that sharp minimax lower bounds are an open question in that formulation [2302.03821].

Later work generalized this template. The unified framework of "PASTA: A Unified Framework for Offline Assortment Learning" defines a likelihood-based confidence set over a general model class $P$,
\[
\Omega_n(\alpha_n)=\{p\in P:\widehat L_n(p)-\widehat L_n(\widehat p_n)\le \alpha_n\},
\]
and solves
\[
\widehat s_{p,n}\in\arg\max_{s\in S}\min_{p\in\Omega_n}\mathcal V(s;p).
\]
It proves a general regret bound of the form
\[
R(\widehat s_{p,n})\lesssim r_{s^\star}\sqrt{\frac{\alpha_n(\delta)}{\pi_S(s^\star)}},
\]
and gives model-specific rates for MNL, latent class logit, and nested logit, together with a minimax regret lower bound [2510.01693].

## 4. Computational procedure under MNL

The original solver alternates between assortment optimization given a model parameter and worst-case parameter search given an assortment. Step 0 computes $\hat\theta_{\mathrm{ML},n}$ by standard optimization such as gradient descent, and constructs
\[
\Omega_n=\{\theta\in\Theta:\widehat L_n(\theta)-\widehat L_n(\hat\theta_{\mathrm{ML},n})\le \alpha_n\}.
\]
Step 1 fixes $\theta$, sets $v_i=\exp(x_i^\top\theta)$, and solves
\[
\max_{\gamma\in\Gamma}\frac{\sum_{i\in[N]} r_i v_i\gamma_i}{1+\sum_{i\in[N]} v_i\gamma_i},
\]
where $\gamma\in\{0,1\}^N$ is the assortment indicator and $\Gamma$ is encoded by $M$ linear inequalities with a totally unimodular matrix. This fractional program is equivalent to a linear program in variables $\{w_j\}_{j\in[N]\cup\{0\}}$, and the assortment is recovered via
\[
\gamma_j^\star=\frac{w_j^\star}{v_j w_0^\star},\qquad j\in[N].
\]
Total unimodularity yields polynomial-time solvability and integral solutions [2302.03821].

Step 2 fixes the current assortment $s$ and minimizes $V(s;\theta)$ over $\Omega_n$ by gradient descent with line search. The iteration is initialized at some $\theta^{(0)}\in\Omega_n$, computes the gradient $\xi_\ell=\nabla_\theta V(s;\theta^{(\ell-1)})$, proposes
\[
\theta^{(\ell)}=\theta^{(\ell-1)}-\beta_\ell \xi_\ell,
\]
and shrinks $\beta_\ell$ by a factor $c\in(0,1)$ until feasibility $\theta^{(\ell)}\in\Omega_n$ is restored. The algorithm alternates the two steps for $T$ iterations, starting from $\theta_0=\hat\theta_{\mathrm{ML},n}$ and an initial assortment obtained from Step 1, and returns the final assortment $s_T$ [2302.03821].

The computational significance of this design is that it avoids explicit enumeration of exponentially many assortments. The MLE step scales in $n$ and $d$, the assortment subproblem is polynomial in $N$ and $M$ under totally unimodular constraints, and the inner minimization is first-order in a $d$-dimensional parameter space [2302.03821].

## 5. Numerical behavior, applicability, and limitations

The original numerical studies use synthetic datasets with known $\theta^\star$ and $s^\star$, allowing direct evaluation of regret. The experiments vary the number of products $N$, a cardinality constraint $K$, feature dimension $d$, sample size $n$, and the coverage parameter $p=\pi_S(s^\star)$. Item revenues satisfy $r_i\in[0.5,0.8]$, features are random unit vectors, and $\exp(x_i^\top\theta^\star)\le \exp(-0.6)$ is imposed to avoid degenerate cases. The baseline is estimate-then-optimize: compute $\hat\theta_{\mathrm{ML},n}$ by gradient descent and then solve the assortment LP without pessimism. Evaluation uses average regret and assortment accuracy, defined as the fraction of items in $s^\star$ correctly selected, across 50 random datasets per configuration [2302.03821].

The reported pattern is consistent across the tested regimes. When sample size varies, including configurations such as $N=40$, $K=8$, $d=16$, $p=0.9$ and also $N=60$, $K=15$, PASTA achieves regret less than 25% of the baseline’s regret across $n$ values. When coverage varies over $p\in\{0.1,0.3,0.5,0.7,0.9\}$, its gains remain robust, including when suboptimal assortments are not covered. When feature dimension varies over $d\in\{8,20,32,64,128\}$, both methods deteriorate, but PASTA maintains an advantage in both regret and accuracy. The reported hyperparameters are $\alpha_n=2\cdot \widehat L_n(\hat\theta_{\mathrm{ML},n})$, $T=30$, and GDLS parameters $L=2$, $\widetilde\beta=0.01$, $c=1/2$ [2302.03821].

The method is particularly suited to offline retail settings with large $N$ and combinatorial $\mathbb{S}$, where logged data likely omit many assortments but include optimal or near-optimal assortments with nonzero probability. The paper recommends estimating $\hat\theta_{\mathrm{ML},n}$, building $\Omega_n$ from the likelihood-ratio constraint, running the alternating LP-plus-GDLS procedure, and conducting sensitivity analysis over $\alpha_n$ and $\delta$ while monitoring $\min_{\theta\in\Omega_n}V(s;\theta)$ across candidate assortments [2302.03821].

The assumptions and limitations are explicit. The guarantees rely on the MNL model with compact $\Theta$, uniform strong convexity of $L(\theta)$ and $\widehat L_n(\theta)$, known revenues $r_i$ independent of assortment, stationarity through i.i.d. offline samples, causal consistency and un-confoundedness, and positivity at the optimum only. The regret bound degrades with $C_{s^\star}=1/\pi_S(s^\star)$, very small coverage of $s^\star$ is unfavorable, the analysis is tailored to MNL, and handling assortment-dependent revenues is possible but complicates optimization [2302.03821].

## 6. Later developments and evolving uses of the term

Later papers use “PASTA” in related but not identical senses, and this terminological expansion is itself part of the topic.

| Work | Relation to PASTA | Coverage or guarantee emphasized |
|---|---|---|
| [2302.03821] | Original max–min pessimistic assortment optimization | Positivity at optimum, $\pi_S(s^\star)>0$ |
| [2502.06777] | PRB as a PASTA-style pessimistic pipeline | Optimal item coverage is sufficient and necessary |
| [2510.01693] | Unified PASTA framework | Offline data distribution contains an optimal assortment |
| [2604.19008] | Offline LCB assortment–pricing method mapped to “PASTA” terminology | Local information around the optimal assortment–price pair |

"Learning an Optimal Assortment Policy under Observational Data" does not define PASTA as a formal algorithm; instead, it presents Pessimistic Rank-Breaking (PRB) and interprets it as an instance of the same pessimistic principle. PRB rank-breaks observational choices into pairwise comparisons between item $j$ and the no-purchase option, constructs lower-confidence estimates of $P(j\mid\{0,j\})$, converts them to pessimistic attraction parameters $v^{\mathrm{LCB}}$, and solves
\[
\widehat S\in\arg\max_{|S|\le K} R(S;v^{\mathrm{LCB}}).
\]
Its main conceptual advance is the replacement of full optimal-assortment coverage by optimal item coverage: each item in $S^\star$ must appear sufficiently often in historical offers, even if the full set $S^\star$ never appears as a logged assortment. The paper proves that this condition is both sufficient and necessary for efficient offline learning under MNL, and gives near-matching upper and lower bounds, making PRB nearly minimax optimal up to logarithmic factors [2502.06777].

"PASTA: A Unified Framework for Offline Assortment Learning" broadens the original construction from MNL to a model class including MNL, latent class logit, and nested logit. It retains the likelihood-ratio uncertainty set and pessimistic max–min selection rule, proves finite-sample regret bounds under coverage at the optimum, and derives a minimax regret lower bound. In its parametric summary, the critical value satisfies
\[
\alpha_n(\delta)\eqsim \frac{D_P}{n}\log\frac{C_P n}{D_P}+\frac{1}{n}\log\frac{1}{\delta},
\]
so the regret scales as
\[
R(\widehat s_{p,n})\lesssim r_{s^\star}\sqrt{\frac{\alpha_n(\delta)}{\pi_S(s^\star)}}.
\]
The paper presents this as the first finite-sample regret bounds for offline assortment optimization across several widely used choice models and as a minimax-optimal result in sample and model complexity up to logarithmic factors [2510.01693].

"Optimal Online and Offline Algorithms for Contextual MNL with Applications to Assortment and Pricing" does not literally use the name “PASTA,” but its offline method is explicitly mapped to that terminology in the supplied summary. There the pessimistic object is an assortment–price pair $(S,p)$ rather than an assortment alone. The algorithm estimates a regularized MLE, constructs a price-uniform confidence region for price-dependent utilities, defines lower-confidence utilities
\[
u_i^{\mathrm{LCB}}(p)=\widetilde x_i(p)^\top \hat\vartheta
-
16\sqrt{3}\,
\|\widetilde x_i(p)\|_{H_{\widetilde D}(\hat\vartheta)^{-1}}
\sqrt{\log(N/\delta)},
\]
and maximizes the induced lower-confidence revenue
\[
R^{\mathrm{LCB}}(S,p)=\frac{\sum_{i\in S} p_i \exp(u_i^{\mathrm{LCB}}(p_i))}{1+\sum_{i\in S}\exp(u_i^{\mathrm{LCB}}(p_i))}.
\]
Its offline suboptimality guarantee is governed by local information around the optimal assortment–price pair, via inverse-Hessian geometry and weights $q_j^\star q_0^\star$, rather than by exact coverage of the optimal action [2604.19008].

Taken together, these developments suggest two stable features of PASTA as a research concept. First, pessimism is the invariant algorithmic principle: optimize a lower-confidence or worst-case revenue functional rather than a point estimate. Second, the data requirement has progressively weakened from exact coverage of the optimal assortment, to itemwise coverage of the optimal set, to more local information conditions around the optimal action. That trajectory does not erase the original formulation; rather, it locates the 2023 PASTA algorithm as the starting point of a broader pessimistic program in offline assortment learning [2302.03821] [2502.06777] [2510.01693] [2604.19008].

Source: https://www.emergentmind.com/topics/pessimistic-assortment-optimization-pasta