Linear-Algebra-Done-Right · Unit 17 · Video 2 · Interactive Practice

The Pseudoinverse T† and Its Three Inverse-Like Properties

IKey Formulas

FormulaNameWhat it says
T∣(null⁡T)⊥:(null⁡T)⊥→range⁡TT|_{(\operatorname{null} T)^\perp} : (\operatorname{null} T)^\perp \to \operatorname{range} TRestriction theoremInjective and onto range⁡T\operatorname{range} T, hence invertible
T†w=(T∣(null⁡T)⊥)−1Prange⁡T wT^\dagger w = (T|_{(\operatorname{null} T)^\perp})^{-1} P_{\operatorname{range} T}\, wPseudoinverse T†∈L(W,V)T^\dagger \in \mathcal{L}(W, V)Project onto range⁡T\operatorname{range} T, then invert the restriction
TT invertible ⇒T†=T−1\Rightarrow T^\dagger = T^{-1}Property (1)Nothing changes when an inverse exists
TT†=Prange⁡TTT^\dagger = P_{\operatorname{range} T},  T†T=P(null⁡T)⊥\ T^\dagger T = P_{(\operatorname{null} T)^\perp}Properties (2) and (3)Both products are orthogonal projections

Key Insight: T†T^\dagger sends (range⁡T)⊥(\operatorname{range} T)^\perp to 00, and sends each w∈range⁡Tw \in \operatorname{range} T to the one vector of (null⁡T)⊥(\operatorname{null} T)^\perp that TT maps to ww.

IIBuilding T†wT^\dagger w

For T(x,y)=(x+y, 12(x+y))T(x, y) = \big(x + y,\ \tfrac{1}{2}(x + y)\big), T†wT^\dagger w projects ww onto range⁡T\operatorname{range} T, then pulls it back into (null⁡T)⊥(\operatorname{null} T)^\perp.

💡 The solutions of Tv=Prange⁡T wTv = P_{\operatorname{range} T}\, w form the line T†w+null⁡TT^\dagger w + \operatorname{null} T; the next video proves that T†wT^\dagger w is the one of smallest norm.

IIIThe Round Trip T†TT^\dagger T

Sending vv to WW and back, T†TT^\dagger T keeps the (null⁡T)⊥(\operatorname{null} T)^\perp piece of vv and deletes the null⁡T\operatorname{null} T piece.

IVBoth Products Are Projections

TT†TT^\dagger and T†TT^\dagger T are orthogonal projections, and each is an identity exactly when TT is surjective or injective.

VQuiz Questions

Problem 1 · Compute a Pseudoinverse Value

Given: T∈L(R2)T \in \mathcal{L}(\mathbf{R}^2), T(x,y)=(x−y, 0)T(x, y) = (x - y,\ 0) — find T†(4,3)T^\dagger(4, 3).

✅ Correct! (2,−2)(2, -2) lies in (null⁡T)⊥(\operatorname{null} T)^\perp and T(2,−2)=(4,0)=Prange⁡T(4,3)T(2, -2) = (4, 0) = P_{\operatorname{range} T}(4, 3).
❌ Close, but… T(4,0)=(4,0)T(4, 0) = (4, 0), so (4,0)(4, 0) solves Tv=Prange⁡TwTv = P_{\operatorname{range} T} w, but it is not in (null⁡T)⊥=span⁡((1,−1))(\operatorname{null} T)^\perp = \operatorname{span}\big((1, -1)\big); T†wT^\dagger w must be the solution that is.
❌ Not quite. That is P(null⁡T)⊥(4,3)P_{(\operatorname{null} T)^\perp}(4, 3); the projection in the definition is onto range⁡T\operatorname{range} T in WW, and it is followed by the inverse of the restriction.
❌ Not quite. T†w=0T^\dagger w = 0 only when w∈(range⁡T)⊥w \in (\operatorname{range} T)^\perp. Here w∉range⁡Tw \notin \operatorname{range} T, but its projection (4,0)(4, 0) is nonzero.
❌ Not quite. Project ww onto range⁡T\operatorname{range} T first, then find the vector of (null⁡T)⊥(\operatorname{null} T)^\perp that TT sends there.
Show solution

range⁡T=span⁡((1,0))\operatorname{range} T = \operatorname{span}\big((1, 0)\big), so Prange⁡T(4,3)=(4,0)P_{\operatorname{range} T}(4, 3) = (4, 0).

null⁡T={(x,y):x=y}=span⁡((1,1))\operatorname{null} T = \{(x, y) : x = y\} = \operatorname{span}\big((1, 1)\big), so (null⁡T)⊥=span⁡((1,−1))(\operatorname{null} T)^\perp = \operatorname{span}\big((1, -1)\big).

Find tt with T(t(1,−1))=(4,0)T\big(t(1, -1)\big) = (4, 0):

T(t,−t)=(2t,0)=(4,0) ⇒ t=2.T(t, -t) = (2t, 0) = (4, 0) \ \Rightarrow\ t = 2.

Hence T†(4,3)=(2,−2)T^\dagger(4, 3) = (2, -2). The other solutions of Tv=(4,0)Tv = (4, 0), such as (4,0)(4, 0), differ from it by a vector of null⁡T\operatorname{null} T.

Problem 2 · Injective but Not Surjective

Given: VV is finite-dimensional and T∈L(V,W)T \in \mathcal{L}(V, W) is injective but not surjective — which statement is true?

✅ Correct! Injectivity gives (null⁡T)⊥=V(\operatorname{null} T)^\perp = V, so T†T^\dagger is a left inverse; since range⁡T≠W\operatorname{range} T \neq W, it is not a right inverse.
❌ Close, but… TT†=Prange⁡TTT^\dagger = P_{\operatorname{range} T}, which is IWI_W only when range⁡T=W\operatorname{range} T = W, that is, when TT is surjective.
❌ Not quite. TT is not surjective, so it is not invertible and T−1T^{-1} does not exist.
❌ Not quite. T†TT^\dagger T projects onto (null⁡T)⊥(\operatorname{null} T)^\perp, not onto null⁡T\operatorname{null} T; here null⁡T={0}\operatorname{null} T = \{0\}, so Pnull⁡TP_{\operatorname{null} T} would be the zero map.
❌ Not quite. Use TT†=Prange⁡TTT^\dagger = P_{\operatorname{range} T} and T†T=P(null⁡T)⊥T^\dagger T = P_{(\operatorname{null} T)^\perp}.
Show solution

TT injective means null⁡T={0}\operatorname{null} T = \{0\}, so (null⁡T)⊥=V(\operatorname{null} T)^\perp = V and property (3) gives

T†T=P(null⁡T)⊥=PV=IV.T^\dagger T = P_{(\operatorname{null} T)^\perp} = P_V = I_V.

TT not surjective means range⁡T≠W\operatorname{range} T \neq W, so property (2) gives TT†=Prange⁡TTT^\dagger = P_{\operatorname{range} T}, which sends every nonzero vector of (range⁡T)⊥(\operatorname{range} T)^\perp to 00 and is therefore not IWI_W.

So T†T^\dagger is a left inverse of TT but not a right inverse.

Problem 3 · A Map from F³ to F²

Given: T∈L(R3,R2)T \in \mathcal{L}(\mathbf{R}^3, \mathbf{R}^2), T(x,y,z)=(x+z, y+z)T(x, y, z) = (x + z,\ y + z) — find T†(3,0)T^\dagger(3, 0) and T†T(1,0,0)T^\dagger T(1, 0, 0).

What is T†(3,0)T^\dagger(3, 0)?

What is T†T(1,0,0)T^\dagger T(1, 0, 0)?

✅ Correct! T†(3,0)=(2,−1,1)T^\dagger(3, 0) = (2, -1, 1), and T†TT^\dagger T removes the null⁡T\operatorname{null} T component (13,13,−13)(\tfrac{1}{3}, \tfrac{1}{3}, -\tfrac{1}{3}) from (1,0,0)(1, 0, 0).
❌ Check T†(3,0)T^\dagger(3, 0). Every option solves Tv=(3,0)Tv = (3, 0); only one of them is orthogonal to null⁡T=span⁡((1,1,−1))\operatorname{null} T = \operatorname{span}\big((1, 1, -1)\big).
❌ Not quite. T†T=IT^\dagger T = I only for injective TT, and this TT has null⁡T=span⁡((1,1,−1))\operatorname{null} T = \operatorname{span}\big((1, 1, -1)\big).
❌ Close, but… That is the projection of (1,0,0)(1, 0, 0) onto null⁡T\operatorname{null} T, the piece T†TT^\dagger T deletes; property (3) keeps the other piece.
❌ Check T†T(1,0,0)T^\dagger T(1, 0, 0). Use T†T=P(null⁡T)⊥T^\dagger T = P_{(\operatorname{null} T)^\perp} and subtract the null⁡T\operatorname{null} T component.
Show solution

TT is surjective (it sends (1,0,0)(1, 0, 0) and (0,1,0)(0, 1, 0) to the standard basis), so range⁡T=R2\operatorname{range} T = \mathbf{R}^2 and Prange⁡T(3,0)=(3,0)P_{\operatorname{range} T}(3, 0) = (3, 0).

null⁡T\operatorname{null} T: x+z=0x + z = 0 and y+z=0y + z = 0, so null⁡T=span⁡((1,1,−1))\operatorname{null} T = \operatorname{span}\big((1, 1, -1)\big) and (null⁡T)⊥={(x,y,z):x+y−z=0}(\operatorname{null} T)^\perp = \{(x, y, z) : x + y - z = 0\}.

Part 1. Solve x+z=3x + z = 3, y+z=0y + z = 0, x+y−z=0x + y - z = 0. From the first two, x=3−zx = 3 - z and y=−zy = -z; then (3−z)−z−z=0(3 - z) - z - z = 0 gives z=1z = 1:

T†(3,0)=(2,−1,1),T(2,−1,1)=(3,0).T^\dagger(3, 0) = (2, -1, 1), \qquad T(2, -1, 1) = (3, 0).

Part 2. By property (3), T†T=P(null⁡T)⊥T^\dagger T = P_{(\operatorname{null} T)^\perp}. With n=(1,1,−1)n = (1, 1, -1) and ∥n∥2=3\|n\|^2 = 3:

P(null⁡T)⊥(1,0,0)=(1,0,0)−⟨(1,0,0),n⟩3 n=(1,0,0)−13(1,1,−1)=(23,−13,13).P_{(\operatorname{null} T)^\perp}(1, 0, 0) = (1, 0, 0) - \frac{\langle (1, 0, 0), n\rangle}{3}\, n = (1, 0, 0) - \tfrac{1}{3}(1, 1, -1) = \big(\tfrac{2}{3}, -\tfrac{1}{3}, \tfrac{1}{3}\big).

Check: T(1,0,0)=(1,0)T(1, 0, 0) = (1, 0) and T(23,−13,13)=(1,0)T\big(\tfrac{2}{3}, -\tfrac{1}{3}, \tfrac{1}{3}\big) = (1, 0), and (23,−13,13)⊥n\big(\tfrac{2}{3}, -\tfrac{1}{3}, \tfrac{1}{3}\big) \perp n.

Problem 4 · The Pseudoinverse of a Projection

Given: UU is a subspace of a finite-dimensional inner product space VV with {0}≠U≠V\{0\} \neq U \neq V, and T=PU∈L(V)T = P_U \in \mathcal{L}(V) — find T†T^\dagger.

✅ Correct! The restriction of PUP_U to (null⁡PU)⊥=U(\operatorname{null} P_U)^\perp = U is the identity on UU, so T†=IUPU=PUT^\dagger = I_U P_U = P_U.
❌ Not quite. T†T^\dagger sends (range⁡T)⊥=U⊥(\operatorname{range} T)^\perp = U^\perp to 00, so it cannot be II when U≠VU \neq V.
❌ Not quite. PU⊥P_{U^\perp} kills U=range⁡TU = \operatorname{range} T, but T†T^\dagger must send each u∈range⁡Tu \in \operatorname{range} T to a vector that TT maps back to uu.
❌ Not quite. T†T^\dagger is defined for every T∈L(V,W)T \in \mathcal{L}(V, W) with VV finite-dimensional; invertibility is not needed.
❌ Not quite. Compute null⁡PU\operatorname{null} P_U and range⁡PU\operatorname{range} P_U, then apply the definition.
Show solution

For T=PUT = P_U: range⁡T=U\operatorname{range} T = U and null⁡T=U⊥\operatorname{null} T = U^\perp, so (null⁡T)⊥=(U⊥)⊥=U(\operatorname{null} T)^\perp = (U^\perp)^\perp = U.

The restriction T∣UT|_U sends each u∈Uu \in U to PUu=uP_U u = u, so it is the identity on UU and its inverse is the identity on UU too. By the definition,

T†w=(T∣U)−1PUw=PUwfor every w∈V.T^\dagger w = (T|_U)^{-1} P_U w = P_U w \quad \text{for every } w \in V.

Hence T†=PUT^\dagger = P_U. Consistently with properties (2) and (3), TT†=PU2=PU=Prange⁡TTT^\dagger = P_U^2 = P_U = P_{\operatorname{range} T} and T†T=PU=P(null⁡T)⊥T^\dagger T = P_U = P_{(\operatorname{null} T)^\perp}.

Solved: 0 / 4