Stochastic Linear Programming and Its Application to Economic Planning

Gerhard Tintner and Jati K. Sengupta’s Festschrift essay, originally published in 1964 and reprinted in 1969, develops stochastic linear programming as an instrument for economic planning under risk. Its central move is to replace a single optimal plan, calculated from fixed coefficients, with a probability distribution of optimal outcomes—and then to distinguish estimating that distribution from choosing policies according to it. The essay proceeds through theoretical bounds, static and dynamic applications to Indian planning, a discussion of decision rules, and proposals for further generalization. Its substantive conclusion is that the allocation yielding the highest expected national income need not be preferable to a planner concerned with the reliability of achieving a target.

The starting point is a linear program whose objective coefficients, technical coefficients, and resource availabilities may be random variables with known or estimable distributions. In the “passive” or “wait and see” approach, the distribution of optimal outcomes is derived across realizations of these parameters. Decisions can then refer to its central tendency, dispersion, or other characteristics. The “active” or “here and now” approach introduces variables specifying resource allocations and evaluates their consequences through a risk preference functional.

One of the specific advantages of the active approach in (2.0) is that it allows a flexible formulation of the decision-making problem.

This flexibility connects programming to economic policy: activity levels become targets, while resource-allocation proportions become instruments. The authors thus distinguish uncertainty in economic relationships from the policy choices made in response to it. An allocation is not justified merely because it maximizes an objective under one configuration of coefficients; it must be assessed through the distribution of outcomes it produces.

The theoretical section restricts attention to admissible parameter configurations yielding defined, bounded, feasible solutions. Its first theorem develops inequalities between passive and active objective values, grounded in the requirement that active solutions also satisfy the passive constraints. Corollaries address extrema across parameter selections, replacement of random coefficients by their expectations, and corresponding dual minimization problems. The authors qualify these comparisons: median orderings need not follow, and some minimum-value bounds depend on admitting the zero activity vector. A second theorem, under symmetry and mutual-independence assumptions and at least locally around parameter means, compares expected stochastic optima with solutions obtained using mean coefficients. These are conditional tools for approximation and sensitivity analysis, not an unrestricted endorsement of replacing uncertain data with averages.

The static application maximizes national income from investment-goods and consumption-goods production under capital and labor constraints. Using mean coefficients, the passive program and an appropriately chosen active allocation both yield 36.4103; equal resource shares yield only 31.1395. The stochastic calculations then compare two procedures: optimizing at observed coefficient pairs and fitting the resulting outcome distribution, or fitting coefficient distributions first and solving at selected points. Sixteen observations underpin the exercise. The reported distributional tests do not reject a common population for the two approximations, although the authors acknowledge the continuity assumptions involved. The point is methodological: a distribution of optimal outcomes can be approximated numerically without deriving it analytically.

The dynamic application reformulates Mahalanobis’s two-sector Indian growth model as a program maximizing terminal national income subject to investment, consumption, and resource restrictions. With a fixed one-third investment-sector allocation, mean coefficients give terminal income of 175.74, whereas the stochastic calculation gives an expectation of 180.10 and variance of 851.88. The distribution supplies information absent from the deterministic result:

Secondly, it can be used as a check (or a bound based on probabilities) against a purely deterministic result derived from a nonstochastic model.

The active comparison makes the policy stakes explicit. Among the three allocation ratios examined, assigning one-third of investment to the investment-goods sector gives the highest expected terminal income. Assigning two-thirds gives a lower expectation, but substantially smaller variance and a higher lower-five-percent income level. The ranking therefore changes when the criterion shifts from expected achievement to a cautious target.

Hence, the instability on account of the higher variance may offset the value of greater expectation of the end-period national income, as far as the national planner is concerned.

Risk preference is consequently part of the planning problem itself. The authors acknowledge that fully specifying preferences over future consumption, income, and uncertainty would generally introduce nonlinearity. Their linear illustration compares distributions under alternative allocations rather than deriving one universally optimal policy. They also leave time-varying allocation sequences incompletely explored.

The concluding conceptual discussion organizes decision rules into progressively broader classes: rules based on distributional characteristics, rules incorporating loss functions and observations, and mappings from uncertainty into a corresponding certainty space. A weighted combination of expected income and negative variance illustrates how risk preferences can produce a scalar objective. Further research would extend distribution-free bounds, incorporate new information sequentially, improve numerical estimation, and allow greater sectoral detail and time-varying coefficients.

This type of analysis may be relevant, particularly when some coefficients are estimated not statistically, but by extraneous information or by subjective belief.

This closing observation broadens the essay’s relevance beyond coefficients supported by ample statistical evidence. Planning requires sensitivity analysis around uncertain estimates as well as optimization. The contribution is therefore both computational and conceptual: it makes the dependability of an income target, and the planner’s valuation of that dependability, explicit components of resource allocation.