How do we know we're maximizing the Lagrangian objective function in PCA?












1












$begingroup$


In Principal Component Analysis, we start with $m$ observations $x_1,dots,x_m$, each of which is an $n$-dimensional vector. Assume we have centered the data; that is, we have subtracted the variable means from each observation. To project each observation into a $1$-dimensional subspace while maximizing sample variance, we want to compute a unit length basis vector, call it $v$. So we get the constrained problem



$$
mathop{arg,max}limits_{v},frac{1}{m}sum_{i=1}^{m}leftlVertlangle x_i,vrangle vrightrVert
$$



$$
text{subject to }v^Tv=1
$$



After some computations, this becomes



$$
mathop{arg,max}limits_{v},frac{1}{m}sum_{i=1}^{m}v^TCv
$$



$$
text{subject to }v^Tv=1
$$



where $C$ is the covariance matrix of the data. Then the Lagrangian objective function is



$$
mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)
$$



If we start taking partial derivatives of $mathcal{L}$, setting these to zero, etc, how do we know that we are maximizing $mathcal{L}$? Do we need to check that the Hessian is negative definite? Even then, I think that only guarantees a local maximum?










share|cite|improve this question









$endgroup$

















    1












    $begingroup$


    In Principal Component Analysis, we start with $m$ observations $x_1,dots,x_m$, each of which is an $n$-dimensional vector. Assume we have centered the data; that is, we have subtracted the variable means from each observation. To project each observation into a $1$-dimensional subspace while maximizing sample variance, we want to compute a unit length basis vector, call it $v$. So we get the constrained problem



    $$
    mathop{arg,max}limits_{v},frac{1}{m}sum_{i=1}^{m}leftlVertlangle x_i,vrangle vrightrVert
    $$



    $$
    text{subject to }v^Tv=1
    $$



    After some computations, this becomes



    $$
    mathop{arg,max}limits_{v},frac{1}{m}sum_{i=1}^{m}v^TCv
    $$



    $$
    text{subject to }v^Tv=1
    $$



    where $C$ is the covariance matrix of the data. Then the Lagrangian objective function is



    $$
    mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)
    $$



    If we start taking partial derivatives of $mathcal{L}$, setting these to zero, etc, how do we know that we are maximizing $mathcal{L}$? Do we need to check that the Hessian is negative definite? Even then, I think that only guarantees a local maximum?










    share|cite|improve this question









    $endgroup$















      1












      1








      1


      1



      $begingroup$


      In Principal Component Analysis, we start with $m$ observations $x_1,dots,x_m$, each of which is an $n$-dimensional vector. Assume we have centered the data; that is, we have subtracted the variable means from each observation. To project each observation into a $1$-dimensional subspace while maximizing sample variance, we want to compute a unit length basis vector, call it $v$. So we get the constrained problem



      $$
      mathop{arg,max}limits_{v},frac{1}{m}sum_{i=1}^{m}leftlVertlangle x_i,vrangle vrightrVert
      $$



      $$
      text{subject to }v^Tv=1
      $$



      After some computations, this becomes



      $$
      mathop{arg,max}limits_{v},frac{1}{m}sum_{i=1}^{m}v^TCv
      $$



      $$
      text{subject to }v^Tv=1
      $$



      where $C$ is the covariance matrix of the data. Then the Lagrangian objective function is



      $$
      mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)
      $$



      If we start taking partial derivatives of $mathcal{L}$, setting these to zero, etc, how do we know that we are maximizing $mathcal{L}$? Do we need to check that the Hessian is negative definite? Even then, I think that only guarantees a local maximum?










      share|cite|improve this question









      $endgroup$




      In Principal Component Analysis, we start with $m$ observations $x_1,dots,x_m$, each of which is an $n$-dimensional vector. Assume we have centered the data; that is, we have subtracted the variable means from each observation. To project each observation into a $1$-dimensional subspace while maximizing sample variance, we want to compute a unit length basis vector, call it $v$. So we get the constrained problem



      $$
      mathop{arg,max}limits_{v},frac{1}{m}sum_{i=1}^{m}leftlVertlangle x_i,vrangle vrightrVert
      $$



      $$
      text{subject to }v^Tv=1
      $$



      After some computations, this becomes



      $$
      mathop{arg,max}limits_{v},frac{1}{m}sum_{i=1}^{m}v^TCv
      $$



      $$
      text{subject to }v^Tv=1
      $$



      where $C$ is the covariance matrix of the data. Then the Lagrangian objective function is



      $$
      mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)
      $$



      If we start taking partial derivatives of $mathcal{L}$, setting these to zero, etc, how do we know that we are maximizing $mathcal{L}$? Do we need to check that the Hessian is negative definite? Even then, I think that only guarantees a local maximum?







      optimization lagrange-multiplier






      share|cite|improve this question













      share|cite|improve this question











      share|cite|improve this question




      share|cite|improve this question










      asked Dec 4 '18 at 17:38









      waynemystirwaynemystir

      475




      475






















          1 Answer
          1






          active

          oldest

          votes


















          0












          $begingroup$

          Define $K:=frac{C+C^T}{2}-lambda I$. Note that $mathcal{L}=v^T Kv+lambda$ is a quadratic, so if $K$ is definite the local turning point, which solves $Kv=0$ from the first derivative of $mathcal{L}$, is also a global extremum. I've written the coefficient as a symmetric real matrix so it's diagonalisable, viz. $mathcal{L}=sum_i K_{ii}v_i^2+lambda$. In our choice of basis $Kv=0$ simplifies to $K_{ii}ne 0implies v_i=0$. This maximises $mathcal{L}$ provided each of the eigenvalues $K_{ii}$ is at most $0$. Indeed, the solution is for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$, with $v$ a unit eigenvector corresponding to $lambda$; then $K$ has $0$ as an eigenvalue, but no positive eigenvalues.






          share|cite|improve this answer











          $endgroup$













          • $begingroup$
            Thank you. Why is the local turning point a global extremum if $K$ is definite? Why does diagonalisability give an obvious result? Which vector space?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:11










          • $begingroup$
            @waynemystir Hopefully my edit makes it clearer.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:28










          • $begingroup$
            I'm confused. What is $V$ or $V_i$? A vector and component? Is $K$ the coefficient? So you diagonalised $K$ and rewrote $mathcal{L}$ relative to this change of basis? Why does the stationary point solve $KV=0$?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:40










          • $begingroup$
            @waynemystir I hope my latest edit is even clearer. $v_i$ is indeed a component of $v$.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:46










          • $begingroup$
            Why is the solution for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$? Taking partials, I see $lambda$ is an eigenvalue. But why does it need to be the largest? The equation $mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)=v^TCv=v^Tlambda v=lambda$ only tells us that $mathcal{L}$ attains the eigenvalue $lambda$. This equation doesn't reveal that it's the largest one.
            $endgroup$
            – waynemystir
            Dec 5 '18 at 0:19













          Your Answer





          StackExchange.ifUsing("editor", function () {
          return StackExchange.using("mathjaxEditing", function () {
          StackExchange.MarkdownEditor.creationCallbacks.add(function (editor, postfix) {
          StackExchange.mathjaxEditing.prepareWmdForMathJax(editor, postfix, [["$", "$"], ["\\(","\\)"]]);
          });
          });
          }, "mathjax-editing");

          StackExchange.ready(function() {
          var channelOptions = {
          tags: "".split(" "),
          id: "69"
          };
          initTagRenderer("".split(" "), "".split(" "), channelOptions);

          StackExchange.using("externalEditor", function() {
          // Have to fire editor after snippets, if snippets enabled
          if (StackExchange.settings.snippets.snippetsEnabled) {
          StackExchange.using("snippets", function() {
          createEditor();
          });
          }
          else {
          createEditor();
          }
          });

          function createEditor() {
          StackExchange.prepareEditor({
          heartbeatType: 'answer',
          autoActivateHeartbeat: false,
          convertImagesToLinks: true,
          noModals: true,
          showLowRepImageUploadWarning: true,
          reputationToPostImages: 10,
          bindNavPrevention: true,
          postfix: "",
          imageUploader: {
          brandingHtml: "Powered by u003ca class="icon-imgur-white" href="https://imgur.com/"u003eu003c/au003e",
          contentPolicyHtml: "User contributions licensed under u003ca href="https://creativecommons.org/licenses/by-sa/3.0/"u003ecc by-sa 3.0 with attribution requiredu003c/au003e u003ca href="https://stackoverflow.com/legal/content-policy"u003e(content policy)u003c/au003e",
          allowUrls: true
          },
          noCode: true, onDemand: true,
          discardSelector: ".discard-answer"
          ,immediatelyShowMarkdownHelp:true
          });


          }
          });














          draft saved

          draft discarded


















          StackExchange.ready(
          function () {
          StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmath.stackexchange.com%2fquestions%2f3025885%2fhow-do-we-know-were-maximizing-the-lagrangian-objective-function-in-pca%23new-answer', 'question_page');
          }
          );

          Post as a guest















          Required, but never shown

























          1 Answer
          1






          active

          oldest

          votes








          1 Answer
          1






          active

          oldest

          votes









          active

          oldest

          votes






          active

          oldest

          votes









          0












          $begingroup$

          Define $K:=frac{C+C^T}{2}-lambda I$. Note that $mathcal{L}=v^T Kv+lambda$ is a quadratic, so if $K$ is definite the local turning point, which solves $Kv=0$ from the first derivative of $mathcal{L}$, is also a global extremum. I've written the coefficient as a symmetric real matrix so it's diagonalisable, viz. $mathcal{L}=sum_i K_{ii}v_i^2+lambda$. In our choice of basis $Kv=0$ simplifies to $K_{ii}ne 0implies v_i=0$. This maximises $mathcal{L}$ provided each of the eigenvalues $K_{ii}$ is at most $0$. Indeed, the solution is for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$, with $v$ a unit eigenvector corresponding to $lambda$; then $K$ has $0$ as an eigenvalue, but no positive eigenvalues.






          share|cite|improve this answer











          $endgroup$













          • $begingroup$
            Thank you. Why is the local turning point a global extremum if $K$ is definite? Why does diagonalisability give an obvious result? Which vector space?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:11










          • $begingroup$
            @waynemystir Hopefully my edit makes it clearer.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:28










          • $begingroup$
            I'm confused. What is $V$ or $V_i$? A vector and component? Is $K$ the coefficient? So you diagonalised $K$ and rewrote $mathcal{L}$ relative to this change of basis? Why does the stationary point solve $KV=0$?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:40










          • $begingroup$
            @waynemystir I hope my latest edit is even clearer. $v_i$ is indeed a component of $v$.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:46










          • $begingroup$
            Why is the solution for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$? Taking partials, I see $lambda$ is an eigenvalue. But why does it need to be the largest? The equation $mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)=v^TCv=v^Tlambda v=lambda$ only tells us that $mathcal{L}$ attains the eigenvalue $lambda$. This equation doesn't reveal that it's the largest one.
            $endgroup$
            – waynemystir
            Dec 5 '18 at 0:19


















          0












          $begingroup$

          Define $K:=frac{C+C^T}{2}-lambda I$. Note that $mathcal{L}=v^T Kv+lambda$ is a quadratic, so if $K$ is definite the local turning point, which solves $Kv=0$ from the first derivative of $mathcal{L}$, is also a global extremum. I've written the coefficient as a symmetric real matrix so it's diagonalisable, viz. $mathcal{L}=sum_i K_{ii}v_i^2+lambda$. In our choice of basis $Kv=0$ simplifies to $K_{ii}ne 0implies v_i=0$. This maximises $mathcal{L}$ provided each of the eigenvalues $K_{ii}$ is at most $0$. Indeed, the solution is for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$, with $v$ a unit eigenvector corresponding to $lambda$; then $K$ has $0$ as an eigenvalue, but no positive eigenvalues.






          share|cite|improve this answer











          $endgroup$













          • $begingroup$
            Thank you. Why is the local turning point a global extremum if $K$ is definite? Why does diagonalisability give an obvious result? Which vector space?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:11










          • $begingroup$
            @waynemystir Hopefully my edit makes it clearer.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:28










          • $begingroup$
            I'm confused. What is $V$ or $V_i$? A vector and component? Is $K$ the coefficient? So you diagonalised $K$ and rewrote $mathcal{L}$ relative to this change of basis? Why does the stationary point solve $KV=0$?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:40










          • $begingroup$
            @waynemystir I hope my latest edit is even clearer. $v_i$ is indeed a component of $v$.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:46










          • $begingroup$
            Why is the solution for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$? Taking partials, I see $lambda$ is an eigenvalue. But why does it need to be the largest? The equation $mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)=v^TCv=v^Tlambda v=lambda$ only tells us that $mathcal{L}$ attains the eigenvalue $lambda$. This equation doesn't reveal that it's the largest one.
            $endgroup$
            – waynemystir
            Dec 5 '18 at 0:19
















          0












          0








          0





          $begingroup$

          Define $K:=frac{C+C^T}{2}-lambda I$. Note that $mathcal{L}=v^T Kv+lambda$ is a quadratic, so if $K$ is definite the local turning point, which solves $Kv=0$ from the first derivative of $mathcal{L}$, is also a global extremum. I've written the coefficient as a symmetric real matrix so it's diagonalisable, viz. $mathcal{L}=sum_i K_{ii}v_i^2+lambda$. In our choice of basis $Kv=0$ simplifies to $K_{ii}ne 0implies v_i=0$. This maximises $mathcal{L}$ provided each of the eigenvalues $K_{ii}$ is at most $0$. Indeed, the solution is for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$, with $v$ a unit eigenvector corresponding to $lambda$; then $K$ has $0$ as an eigenvalue, but no positive eigenvalues.






          share|cite|improve this answer











          $endgroup$



          Define $K:=frac{C+C^T}{2}-lambda I$. Note that $mathcal{L}=v^T Kv+lambda$ is a quadratic, so if $K$ is definite the local turning point, which solves $Kv=0$ from the first derivative of $mathcal{L}$, is also a global extremum. I've written the coefficient as a symmetric real matrix so it's diagonalisable, viz. $mathcal{L}=sum_i K_{ii}v_i^2+lambda$. In our choice of basis $Kv=0$ simplifies to $K_{ii}ne 0implies v_i=0$. This maximises $mathcal{L}$ provided each of the eigenvalues $K_{ii}$ is at most $0$. Indeed, the solution is for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$, with $v$ a unit eigenvector corresponding to $lambda$; then $K$ has $0$ as an eigenvalue, but no positive eigenvalues.







          share|cite|improve this answer














          share|cite|improve this answer



          share|cite|improve this answer








          edited Dec 4 '18 at 20:48









          waynemystir

          475




          475










          answered Dec 4 '18 at 17:51









          J.G.J.G.

          24.6k22539




          24.6k22539












          • $begingroup$
            Thank you. Why is the local turning point a global extremum if $K$ is definite? Why does diagonalisability give an obvious result? Which vector space?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:11










          • $begingroup$
            @waynemystir Hopefully my edit makes it clearer.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:28










          • $begingroup$
            I'm confused. What is $V$ or $V_i$? A vector and component? Is $K$ the coefficient? So you diagonalised $K$ and rewrote $mathcal{L}$ relative to this change of basis? Why does the stationary point solve $KV=0$?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:40










          • $begingroup$
            @waynemystir I hope my latest edit is even clearer. $v_i$ is indeed a component of $v$.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:46










          • $begingroup$
            Why is the solution for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$? Taking partials, I see $lambda$ is an eigenvalue. But why does it need to be the largest? The equation $mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)=v^TCv=v^Tlambda v=lambda$ only tells us that $mathcal{L}$ attains the eigenvalue $lambda$. This equation doesn't reveal that it's the largest one.
            $endgroup$
            – waynemystir
            Dec 5 '18 at 0:19




















          • $begingroup$
            Thank you. Why is the local turning point a global extremum if $K$ is definite? Why does diagonalisability give an obvious result? Which vector space?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:11










          • $begingroup$
            @waynemystir Hopefully my edit makes it clearer.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:28










          • $begingroup$
            I'm confused. What is $V$ or $V_i$? A vector and component? Is $K$ the coefficient? So you diagonalised $K$ and rewrote $mathcal{L}$ relative to this change of basis? Why does the stationary point solve $KV=0$?
            $endgroup$
            – waynemystir
            Dec 4 '18 at 18:40










          • $begingroup$
            @waynemystir I hope my latest edit is even clearer. $v_i$ is indeed a component of $v$.
            $endgroup$
            – J.G.
            Dec 4 '18 at 18:46










          • $begingroup$
            Why is the solution for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$? Taking partials, I see $lambda$ is an eigenvalue. But why does it need to be the largest? The equation $mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)=v^TCv=v^Tlambda v=lambda$ only tells us that $mathcal{L}$ attains the eigenvalue $lambda$. This equation doesn't reveal that it's the largest one.
            $endgroup$
            – waynemystir
            Dec 5 '18 at 0:19


















          $begingroup$
          Thank you. Why is the local turning point a global extremum if $K$ is definite? Why does diagonalisability give an obvious result? Which vector space?
          $endgroup$
          – waynemystir
          Dec 4 '18 at 18:11




          $begingroup$
          Thank you. Why is the local turning point a global extremum if $K$ is definite? Why does diagonalisability give an obvious result? Which vector space?
          $endgroup$
          – waynemystir
          Dec 4 '18 at 18:11












          $begingroup$
          @waynemystir Hopefully my edit makes it clearer.
          $endgroup$
          – J.G.
          Dec 4 '18 at 18:28




          $begingroup$
          @waynemystir Hopefully my edit makes it clearer.
          $endgroup$
          – J.G.
          Dec 4 '18 at 18:28












          $begingroup$
          I'm confused. What is $V$ or $V_i$? A vector and component? Is $K$ the coefficient? So you diagonalised $K$ and rewrote $mathcal{L}$ relative to this change of basis? Why does the stationary point solve $KV=0$?
          $endgroup$
          – waynemystir
          Dec 4 '18 at 18:40




          $begingroup$
          I'm confused. What is $V$ or $V_i$? A vector and component? Is $K$ the coefficient? So you diagonalised $K$ and rewrote $mathcal{L}$ relative to this change of basis? Why does the stationary point solve $KV=0$?
          $endgroup$
          – waynemystir
          Dec 4 '18 at 18:40












          $begingroup$
          @waynemystir I hope my latest edit is even clearer. $v_i$ is indeed a component of $v$.
          $endgroup$
          – J.G.
          Dec 4 '18 at 18:46




          $begingroup$
          @waynemystir I hope my latest edit is even clearer. $v_i$ is indeed a component of $v$.
          $endgroup$
          – J.G.
          Dec 4 '18 at 18:46












          $begingroup$
          Why is the solution for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$? Taking partials, I see $lambda$ is an eigenvalue. But why does it need to be the largest? The equation $mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)=v^TCv=v^Tlambda v=lambda$ only tells us that $mathcal{L}$ attains the eigenvalue $lambda$. This equation doesn't reveal that it's the largest one.
          $endgroup$
          – waynemystir
          Dec 5 '18 at 0:19






          $begingroup$
          Why is the solution for $lambda$ to be the largest eigenvalue of $frac{C+C^T}{2}$? Taking partials, I see $lambda$ is an eigenvalue. But why does it need to be the largest? The equation $mathcal{L}(v,lambda)=v^TCv-lambda(v^Tv-1)=v^TCv=v^Tlambda v=lambda$ only tells us that $mathcal{L}$ attains the eigenvalue $lambda$. This equation doesn't reveal that it's the largest one.
          $endgroup$
          – waynemystir
          Dec 5 '18 at 0:19




















          draft saved

          draft discarded




















































          Thanks for contributing an answer to Mathematics Stack Exchange!


          • Please be sure to answer the question. Provide details and share your research!

          But avoid



          • Asking for help, clarification, or responding to other answers.

          • Making statements based on opinion; back them up with references or personal experience.


          Use MathJax to format equations. MathJax reference.


          To learn more, see our tips on writing great answers.




          draft saved


          draft discarded














          StackExchange.ready(
          function () {
          StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmath.stackexchange.com%2fquestions%2f3025885%2fhow-do-we-know-were-maximizing-the-lagrangian-objective-function-in-pca%23new-answer', 'question_page');
          }
          );

          Post as a guest















          Required, but never shown





















































          Required, but never shown














          Required, but never shown












          Required, but never shown







          Required, but never shown

































          Required, but never shown














          Required, but never shown












          Required, but never shown







          Required, but never shown







          Popular posts from this blog

          Ellipse (mathématiques)

          Quarter-circle Tiles

          Mont Emei