Shephard's Lemma

How Shephard's lemma works

Picture an economist with the expenditure function in hand — a formula e(p, u) that tells you the minimum cost of reaching utility u at price vector p. The expenditure function encodes the entire compensated-demand structure of the consumer; it's the value function for the cost-minimization problem. Shephard's lemma is the rule for unpacking that information one good at a time:

h_i(p, u)  =  ∂e(p, u) / ∂p_i

Differentiate e in the i-th price and you get the i-th Hicksian demand. No Lagrangian, no FOCs, no constrained optimization. The expenditure function already solved that problem; Shephard's lemma reads the answer back out via differentiation. Shephard: ∂e/∂p_i = h_i directly — derivative replaces minimization.

Geometrically the lemma says something almost magical: at any point on the expenditure-vs-price curve, the slope equals the consumed quantity of that good. Plot e against p_1 holding p_2 and u fixed; the slope at every price is the compensated demand h_1 at that price. Slopes are demands.

Formal statement and proof

Let e(p, u) = min_x p · x subject to u(x) ≥ u. The Hicksian demand is the minimizer h(p, u) = argmin. Shephard's lemma:

∂e(p, u)/∂p_i  =  h_i(p, u)

Proof via the envelope theorem. By definition:

e(p, u)  =  p · h(p, u)  =  Σ_j  p_j · h_j(p, u)

Differentiate with respect to p_i, applying the chain rule:

∂e/∂p_i  =  h_i  +  Σ_j  p_j · ∂h_j/∂p_i

The first term is the direct effect of the price change at fixed quantities; the sum is the indirect effect through quantity adjustments. The miracle is that the sum equals zero at the optimum. Why? Because h(p, u) satisfies the FOCs of the constrained minimization: at the optimum, ∂u/∂x_j = λ · p_j for some Lagrange multiplier λ. Differentiating the constraint u(h(p, u)) = u with respect to p_i:

0  =  Σ_j (∂u/∂x_j) · ∂h_j/∂p_i  =  λ · Σ_j p_j · ∂h_j/∂p_i

For λ ≠ 0 (positive marginal utility of income), the sum vanishes. So ∂e/∂p_i = h_i. The envelope theorem in one line.

Intuition: why the substitution term vanishes

The non-obvious step is why the consumer's reoptimization doesn't add anything to the first-order effect on e. Here's the geometric story.

The expenditure function e(p, u) is the lower envelope of all hyperplanes of the form p · x for x on the indifference surface u(x) = u. Each x defines a hyperplane in price space whose value at price vector p is p · x; e(p, u) is the lowest such value as you vary x. Now, if you sit at a particular price p^0 with optimal bundle x^0 = h(p^0, u), the hyperplane of that bundle, p · x^0, agrees with e(p, u) at p = p^0 and lies above it everywhere else. So the two functions are tangent at p^0 — their derivatives match. The derivative of the hyperplane p · x^0 in p_i is just x^0_i = h_i. Hence ∂e/∂p_i = h_i.

This geometric view also reveals why the expenditure function is concave in prices (linearity of the hyperplane plus the min operator), and why the Slutsky matrix is negative semi-definite: e is the lower envelope of linear functions, so it inherits concavity.

Worked example: Cobb-Douglas expenditure

For U(x, y) = x^α · y^(1−α), the expenditure function (derived by minimizing p_x · x + p_y · y subject to the utility constraint) is:

e(p, u)  =  u · (p_x / α)^α · (p_y / (1 − α))^(1 − α)

Apply Shephard's lemma to find the Hicksian demand for x:

h_x  =  ∂e/∂p_x
     =  u · α · p_x^(α−1) / α^α · (p_y / (1 − α))^(1 − α)
     =  u · (p_x/α)^(α−1) · (p_y / (1 − α))^(1 − α)
     =  u · (α · p_y / ((1 − α) · p_x))^(1 − α)

For α = 0.5, this simplifies to h_x = u · (p_y / p_x)^0.5 — exactly what direct optimization would have given. And for the income-symmetric piece, h_y = ∂e/∂p_y = u · (p_x / p_y)^0.5. Same derivative trick, no need to redo the Lagrangian.

Verify by plugging into the original definition: h_x · p_x + h_y · p_y should equal e(p, u). For α = 0.5: u·(p_y/p_x)^0.5 · p_x + u·(p_x/p_y)^0.5 · p_y = u·(p_x p_y)^0.5 + u·(p_x p_y)^0.5 = 2u·(p_x p_y)^0.5 = e(p, u). Check.

Shephard, Roy, Slutsky — side by side

Identity	Recovers	From	Formula	Proof tool
Shephard's lemma	Hicksian demand h_i	Expenditure function e(p, u)	h_i = ∂e/∂p_i	Envelope theorem
Roy's identity	Marshallian demand x_i	Indirect utility v(p, m)	x_i = −(∂v/∂p_i)/(∂v/∂m)	Envelope theorem + Walras
Slutsky equation	Decomposition of ∂x_i/∂p_j	Both demands	∂x/∂p = ∂h/∂p − x·∂x/∂m	Chain rule on h = x(p, e(p,u))
Hotelling's lemma	Supply y_i	Profit function π(p)	y_i = ∂π/∂p_i	Envelope theorem
Shephard (producer)	Conditional factor demand x_i	Cost function C(w, y)	x_i = ∂C/∂w_i	Envelope theorem (original 1953)
McFadden's lemma	Net supply	Restricted-profit function	z_i = ∂π^R/∂p_i	Envelope theorem

All six identities follow from the same envelope-theorem template: the value function of a constrained-optimization problem inherits the derivatives of the Lagrangian at the optimum, with all substitution-margin adjustments killed by the FOCs. Once you've seen the pattern in one, you've seen it in all.

Practical uses

• Empirical demand estimation. Postulate a flexible functional form for e(p, u) (Translog, Almost Ideal, CES) and differentiate to get the Hicksian demand equations. Estimate via Generalized Method of Moments on observed prices and budget shares. The Almost Ideal Demand System (Deaton-Muellbauer 1980) is the industry standard built directly from Shephard's lemma.
• Welfare measures. Compensating variation CV = e(p^1, u^0) − e(p^0, u^0) and equivalent variation EV use the expenditure function directly. Shephard's lemma gives the integral form: CV = ∫ h(p, u^0) dp along the price path. Closed-form welfare numbers from observed price changes.
• Cost-of-living indices. The Konüs (true) cost-of-living index is e(p^1, u)/e(p^0, u) — a ratio of expenditure functions. Differentiating in prices gives the Hicksian quantity weights that approximate a chained-CPI index, fixing the substitution bias of a Laspeyres index.
• Optimal taxation. Ramsey's inverse-elasticity rule and Mirrlees-Saez optimal-income tax rely on compensated elasticities, recovered from estimated expenditure functions via Shephard's lemma. The compensated channel isolates the distortion margin from the income margin.
• Trade gravity equations. Armington-style CES expenditure functions for varieties from different countries get their bilateral trade demand equations directly from Shephard's lemma — modern quantitative trade theory (Eaton-Kortum, Anderson-van Wincoop) is built on this scaffolding.
• Producer-side cost minimization. The original Shephard (1953) result. Estimate a cost function from input prices and output, then differentiate to get the conditional factor demands. Used in productivity studies, input-substitution analysis, and CGE modeling.

Where Shephard's lemma breaks (and what to do)

Failure mode	Where it occurs	Fix
e(p, u) non-differentiable in p	Perfect complements (Leontief preferences); kinked indifference curves	Replace derivative by subdifferential; h becomes set-valued
Multiple optimal bundles	Linear utility (perfect substitutes) at the price ratio	Hicksian demand is a correspondence; pick any element
Corner solutions	One good's price very high relative to MU	Lemma still holds — h_i = 0 means slope of e in p_i is 0
Non-smooth e(p, u)	Discrete choice; price-induced regime switches	Use directional derivatives or stochastic-choice generalizations
Discontinuous demand	Indivisibilities (cars, houses)	Local Shephard + interval inference; full discrete-choice machinery

Related theorems and identities

• Envelope theorem. The general result that the derivative of a constrained value function with respect to a parameter equals the partial derivative of the Lagrangian at the optimum. Shephard's lemma is its consumer-side specialization.
• Roy's identity. The indirect-utility counterpart. Together with Shephard, the two halves of consumer duality.
• Slutsky equation. Derived from Shephard plus the duality identity h(p, u) = x(p, e(p, u)). The bridge between Hicksian and Marshallian demand.
• Hotelling's lemma. Profit-function version: y_i = ∂π/∂p_i. Same envelope-theorem proof.
• McFadden's lemma. Restricted-profit function with some inputs fixed. Generalizes Hotelling to short-run analysis.
• Adding up. Σ_i p_i · h_i = e(p, u) — the budget identity for Hicksian demand. Differentiate and use Shephard to get cross-Slutsky constraints used to check empirical demand systems.

Common pitfalls

• Applying Shephard's lemma to non-differentiable expenditure functions. At kinks (Leontief preferences, perfect complements) the derivative doesn't exist as a function. The lemma generalizes via the subdifferential but requires careful handling.
• Confusing the expenditure function with the cost-minimization Lagrangian. e(p, u) is the value function — the minimum value. The Lagrangian is the auxiliary object used in the optimization. Shephard differentiates the value function, not the Lagrangian.
• Forgetting that h depends on u not m. Shephard returns Hicksian demand indexed by utility level u, not Marshallian demand indexed by income m. The two are equal at the optimum (h(p, v(p, m)) = x(p, m)) but differ off-optimum.
• Misapplying the producer version. Shephard's original 1953 result is about firms: x_i = ∂C/∂w_i where C is the cost function and w_i is input price. Hotelling's lemma uses the profit function and output prices. The two are easily confused.
• Treating Shephard as evidence of cardinality. The expenditure function is ordinal — it depends on the ranking of bundles, not cardinal utility levels. Shephard's lemma holds for any monotonic transformation of u; only the specific u-level indexing changes.
• Reading h_i = ∂e/∂p_i as a definitional identity. It's a theorem requiring the envelope-theorem hypotheses (continuous differentiability of e, interior optimum, regularity conditions on preferences). At boundary points (corner solutions, kinks) the theorem still holds in modified form but the simple gradient identity needs care.