MlpMse

End-to-end analytic soundness for a first training target:

• 2-layer MLP (Linear → ReLU → Linear)
• MSE loss

We prove that the usual backprop formulas (including parameter gradients) coincide with the adjoint of the Fréchet derivative (fderiv) over ℝ.

Notes:

• This is a spec-level (ℝ) theorem.
• ReLU is not differentiable at 0, so we assume a "no kinks" hypothesis on the pre-activation.
• Reverse-mode is naturally stated as a VJP theorem for vector outputs; scalar loss gradients are obtained by choosing δ = 1 (equivalently VJP[loss, y] 1 = ∇ loss y).

PyTorch correspondence / citations

• MLP building blocks: torch.nn.Linear, torch.nn.functional.relu. https://pytorch.org/docs/stable/generated/torch.nn.Linear.html https://pytorch.org/docs/stable/generated/torch.nn.functional.relu.html
• MSE loss reference behavior: https://pytorch.org/docs/stable/generated/torch.nn.MSELoss.html
• “Backward equals adjoint of derivative” is exactly the Jacobian-transpose theorem for PyTorch autograd: https://pytorch.org/docs/stable/autograd.html

theorem Proofs.Autograd.reluDerivCLM_apply {n : ℕ} (x dx : Vec n) (i : Fin n) : ((reluDerivCLM x) dx).ofLp i = dx.ofLp i * Activation.Math.reluDerivSpec (x.ofLp i)

Coordinate formula for reluDerivCLM: it scales each coordinate by relu'(xᵢ).

theorem Proofs.Autograd.reluDerivCLM_inner {n : ℕ} (x dx δ : Vec n) : inner ℝ ((reluDerivCLM x) dx) δ = inner ℝ dx ((reluDerivCLM x) δ)

ReLU derivative map is self-adjoint w.r.t. the Euclidean inner product.

This is because it is a diagonal scaling map on ℝⁿ.

theorem Proofs.Autograd.reluDerivCLM_adjoint_apply {n : ℕ} (x δ : Vec n) : (ContinuousLinearMap.adjoint (reluDerivCLM x)) δ = (reluDerivCLM x) δ

The adjoint of reluDerivCLM equals itself (self-adjoint operator).

theorem Proofs.Autograd.getScalarE_linearInputDerivSpec_eq_adjoint {inDim outDim : ℕ} (W : TorchLean.Tensor ℝ [outDim, inDim]) (δ : TorchLean.Tensor ℝ [outDim]) : getScalarE (Spec.linearInputDerivSpec W δ) = (ContinuousLinearMap.adjoint (matCLM (tensorToMatrix W))) (getScalarE δ)

The spec-level “input derivative” for a linear layer agrees with the Euclidean adjoint.

In words: the tensor expression for ∂(W x)/∂x applied to an upstream δ is Wᵀ δ, and this is exactly the adjoint of the CLM x ↦ W x.

noncomputable def Proofs.Autograd.affineMat {inDim outDim : ℕ} (W : Mat outDim inDim) (b : Vec outDim) : Vec inDim → Vec outDim

Affine map using Mat parameters (same as affine, but with Mat instead of Matrix).

noncomputable def Proofs.Autograd.mlpVecMat {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) : Vec inDim → Vec outDim

2-layer MLP in Euclidean Vec form, parameterized by Mat weights and Vec biases.

This is the same computation as NN.Proofs.Autograd.FDeriv.Core’s mlpVec, but set up for parameter-gradient proofs where the parameter space is a Hilbert space.

noncomputable def Proofs.Autograd.mse {n : ℕ} (t : Vec n) : Vec n → ℝ

Mean-squared error loss (MSE) against a fixed target t:

mse t y = (1/n) * ‖y - t‖².

noncomputable def Proofs.Autograd.mseGrad {n : ℕ} (y t : Vec n) : Vec n

Gradient of MSE with respect to y:

∇_y mse(t)(y) = (2/n) * (y - t).

theorem Proofs.Autograd.hasFDerivAt_mse {n : ℕ} (t y : Vec n) : HasFDerivAt (mse t) ((2 / ↑n) • (innerSL ℝ) (y - t)) y

Fréchet derivative of MSE, packaged as a continuous linear map Vec n →L ℝ.

theorem Proofs.Autograd.adjoint_mseDeriv_one {n : ℕ} (t y : Vec n) : (ContinuousLinearMap.adjoint ((2 / ↑n) • (innerSL ℝ) (y - t))) 1 = mseGrad y t

The adjoint of the MSE derivative, applied to the seed 1, is the gradient mseGrad y t.

theorem Proofs.Autograd.mseGrad_eq_adjoint_fderiv {n : ℕ} (t y : Vec n) : (vjp (mse t) y) 1 = mseGrad y t

The VJP of MSE at y with upstream seed 1 equals the usual gradient mseGrad y t.

This is the scalar-loss specialization: for scalar loss ℓ, the gradient is (fderiv ℓ)† 1.

theorem Proofs.Autograd.adjoint_fderiv_comp_apply_one {E : Type u_1} {F : Type u_2} [NormedAddCommGroup E] [InnerProductSpace ℝ E] [CompleteSpace E] [NormedAddCommGroup F] [InnerProductSpace ℝ F] [CompleteSpace F] {f : E → F} {g : F → ℝ} {f' : E →L[ℝ] F} {g' : F →L[ℝ] ℝ} {x : E} (hf : HasFDerivAt f f' x) (hg : HasFDerivAt g g' (f x)) : (ContinuousLinearMap.adjoint (fderiv ℝ (fun (x : E) => g (f x)) x)) 1 = (ContinuousLinearMap.adjoint f') ((ContinuousLinearMap.adjoint g') 1)

Convenience lemma: adjoint of the derivative of a scalar composition, applied to seed 1.

For scalar loss g ∘ f, this is the reverse-mode “chain rule” in adjoint form.

We name the intermediate activations so the gradient statements read like textbook backprop: mlpPreActivation is z₁ = W₁ x + b₁, mlpHidden is a₁ = relu z₁, and mlpOutput is the network output y = W₂ a₁ + b₂.

noncomputable def Proofs.Autograd.mlpPreActivation {inDim hidDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (x : Vec inDim) : Vec hidDim

Pre-activation W1 x + b1 (the ReLU differentiability hypothesis is stated on it).

noncomputable def Proofs.Autograd.mlpHidden {inDim hidDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (x : Vec inDim) : Vec hidDim

Hidden activation relu (W1 x + b1).

noncomputable def Proofs.Autograd.mlpOutput {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) : Vec outDim

Network output mlpVecMat W1 b1 W2 b2 x.

theorem Proofs.Autograd.hasFDerivAt_mlp_wrt_W2 {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) : HasFDerivAt (fun (W2 : Mat outDim hidDim) => mlpVecMat W1 b1 W2 b2 x) (matApplyLin (mlpHidden W1 b1 x)) W2

Fréchet derivative of the network output with respect to the second-layer weights W2.

Informally: ∂y/∂W2 is the linear map dW2 ↦ dW2 a1.

theorem Proofs.Autograd.grad_W2_mse {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) (t : Vec outDim) : (ContinuousLinearMap.adjoint (fderiv ℝ (fun (W2 : Mat outDim hidDim) => mse t (mlpVecMat W1 b1 W2 b2 x)) W2)) 1 = outer (mseGrad (mlpOutput W1 b1 W2 b2 x) t) (mlpHidden W1 b1 x)

Closed-form gradient of scalar loss mse t (mlpVecMat …) with respect to W2.

Result: ∂L/∂W2 = (∂L/∂y) ⊗ a1, i.e. outer product of the output gradient and hidden activation.

theorem Proofs.Autograd.grad_b2_mse {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) (t : Vec outDim) : (ContinuousLinearMap.adjoint (fderiv ℝ (fun (b2 : Vec outDim) => mse t (mlpVecMat W1 b1 W2 b2 x)) b2)) 1 = mseGrad (mlpVecMat W1 b1 W2 b2 x) t

Closed-form gradient of the scalar loss with respect to the second-layer bias b2.

Result: ∂L/∂b2 = ∂L/∂y.

theorem Proofs.Autograd.hasFDerivAt_mlp_wrt_b1 {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) (hx : ∀ (i : Fin hidDim), (mlpPreActivation W1 b1 x).ofLp i ≠ 0) : HasFDerivAt (fun (b1 : Vec hidDim) => mlpVecMat W1 b1 W2 b2 x) (matCLM (toMatrix W2) ∘SL reluDerivCLM (mlpPreActivation W1 b1 x)) b1

Fréchet derivative of the network output with respect to the first-layer bias b1, under the ReLU “no kinks” hypothesis.

Informally: ∂y/∂b1 = W2 ∘ ReLU'(z1).

theorem Proofs.Autograd.grad_b1_mse {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) (t : Vec outDim) (hx : ∀ (i : Fin hidDim), (mlpPreActivation W1 b1 x).ofLp i ≠ 0) : (ContinuousLinearMap.adjoint (fderiv ℝ (fun (b1 : Vec hidDim) => mse t (mlpVecMat W1 b1 W2 b2 x)) b1)) 1 = (reluDerivCLM (mlpPreActivation W1 b1 x)) ((ContinuousLinearMap.adjoint (matCLM (toMatrix W2))) (mseGrad (mlpOutput W1 b1 W2 b2 x) t))

Closed-form gradient of the scalar loss with respect to the first-layer bias b1 (under the ReLU “no kinks” hypothesis).

theorem Proofs.Autograd.hasFDerivAt_mlp_wrt_x {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) (hx : ∀ (i : Fin hidDim), (mlpPreActivation W1 b1 x).ofLp i ≠ 0) : HasFDerivAt (mlpVecMat W1 b1 W2 b2) (matCLM (toMatrix W2) ∘SL reluDerivCLM (mlpPreActivation W1 b1 x) ∘SL matCLM (toMatrix W1)) x

Fréchet derivative of the network output with respect to the input x, under the ReLU “no kinks” hypothesis.

Informally: ∂y/∂x = W2 ∘ ReLU'(z1) ∘ W1.

theorem Proofs.Autograd.grad_x_mse {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) (t : Vec outDim) (hx : ∀ (i : Fin hidDim), (mlpPreActivation W1 b1 x).ofLp i ≠ 0) : (ContinuousLinearMap.adjoint (fderiv ℝ (fun (x : Vec inDim) => mse t (mlpVecMat W1 b1 W2 b2 x)) x)) 1 = (ContinuousLinearMap.adjoint (matCLM (toMatrix W1))) ((reluDerivCLM (mlpPreActivation W1 b1 x)) ((ContinuousLinearMap.adjoint (matCLM (toMatrix W2))) (mseGrad (mlpOutput W1 b1 W2 b2 x) t)))

Closed-form gradient of the scalar loss with respect to the input x, under the ReLU “no kinks” hypothesis.

theorem Proofs.Autograd.hasFDerivAt_mlp_wrt_W1 {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) (hx : ∀ (i : Fin hidDim), (mlpPreActivation W1 b1 x).ofLp i ≠ 0) : HasFDerivAt (fun (W1 : Mat hidDim inDim) => mlpVecMat W1 b1 W2 b2 x) (matCLM (toMatrix W2) ∘SL reluDerivCLM (mlpPreActivation W1 b1 x) ∘SL matApplyLin x) W1

Fréchet derivative of the network output with respect to the first-layer weights W1, under the ReLU “no kinks” hypothesis.

The derivative is linear in W1 through the slice dW1 ↦ dW1 x, then propagated through ReLU' and W2.

theorem Proofs.Autograd.grad_W1_mse {inDim hidDim outDim : ℕ} (W1 : Mat hidDim inDim) (b1 : Vec hidDim) (W2 : Mat outDim hidDim) (b2 : Vec outDim) (x : Vec inDim) (t : Vec outDim) (hx : ∀ (i : Fin hidDim), (mlpPreActivation W1 b1 x).ofLp i ≠ 0) : (ContinuousLinearMap.adjoint (fderiv ℝ (fun (W1 : Mat hidDim inDim) => mse t (mlpVecMat W1 b1 W2 b2 x)) W1)) 1 = outer ((reluDerivCLM (mlpPreActivation W1 b1 x)) ((ContinuousLinearMap.adjoint (matCLM (toMatrix W2))) (mseGrad (mlpOutput W1 b1 W2 b2 x) t))) x

Closed-form gradient of the scalar loss with respect to W1 (under the ReLU “no kinks” hypothesis).

Result has the expected “outer product” form with backpropagated hidden gradient and input x.